Lecture Notes in Artificial Intelligence

Subseries of Lecture Notes in Computer Science

Edited by J. Siekmann

Lecture Notes in Computer Science

Edited by G. Goos and J. Hartmanis
Editorial

Artificial Intelligence has become a major discipline under the roof of

Computer Science. This is also reflected by a growing number of titles
devoted to this fast developing field to be published in our Lecture
Notes in Computer Science. To make these volumes immediately vis-
ible we have decided to distinguish them by a special cover as Lecture
Notes in Artificial Intelligence, constituting a subseries of the Lecture
Notes in Computer Science. This subseries is edited by an Editorial
Board of experts from all areas of AI, chaired by JSrg Siekmann, who
are looking forward to consider further AI monographs and proceed-
ings of high scientific quality for publication.

We hope that the constitution of this subseries will be well accepted

by the audience of the Lecture Notes in Computer Science, and we
feel confident that the subseries will be recognized as an outstanding
opportunity for publication by authors and editors of the AI community.

Editors and publisher

Lecture Notes in
Artificial Intelligence
Edited by J. Siekmann

Subseries of Lecture Notes in Computer Science

383

K. Furukawa H. Tanaka
T. Fujisaki (Eds.)

Logic Programming '88

Proceedings of the 7th Conference
Tokyo, Japan, April 11-14, 1988

Springer-Verlag
Berlin Heidelberg NewYork London ParisTokyo Hong Kong
Editors
Koichi Furukawa
ICOT Research Center
Mita Kokusai Building 21 F, 4-28 Mita 1-chome
Minato-ku, Tokyo 108, Japan

Hozumi Tanaka
Department of Computer Science, Faculty of Engineering
Tokyo Institute of Technology, 12-10h-okayama 2-chome
Meguro-ku, Tokyo 152, Japan

Tetsunosuke Fujisaki
IBM, Thomas J. Watson Research Center
P.O.Box 704, Yorktown Heights, New York 10598, USA

CR Subject Classification (t987): D.1, D.3.1-4, E4.1-2, 1.2.3-4, 1.2.7

ISBN 3-540-51564-X Springer-Verlag Berlin Heidelberg NewYork

ISBN 0-387-51564-X Springer-Verlag NewYork Berlin Heidelberg

Thiswork is subject to copyright.All rights are reserved,whetherthe whole or part of the material
is concerned, specificallythe rights of translation,reprinting, re-useof illustrations,recitation,
broadcasting, reproductionon microfilmsor in other ways,and storage in data banks. Duplication
of this publicationor parts thereof is only permitted underthe provisionsof the German Copyright
Law of September 9, 1965, in its version of June 24, 1985, and a copyright fee must alwaysbe
paid. Violationsfall under the prosecutionact of the German Copyright Law.
© Springer-VerlagBerlin Heidelberg 1989
Printed in Germany
Printing and binding: Druckhaus Bettz, Hemsbach/Bergstr.
2145/3140-543210- Printed on acid-free paper
Foreword

This volume of the Springer Lecture Notes in Computer Science contains most of
the papers presented at the Seventh Logic P r o g r a m m i n g Conference (April 11-14,
1988 Tokyo). It is the successor to volumes 221,264 and 315.

The c o n t e n t s cover v a r i o u s aspects of logic p r o g r a m m i n g s u c h as

foundations, nonmonotonic reasoning, programming language/systems,
concurrent programming, knowledge base and applications of computer-aided
reasoning and natural language understanding.

The first paper in the "Foundation of Logic Programming" section gives

theoretical results on "narrowing", a natural extension of unification for handling
equalities as well as usual Horn clauses. The second paper gives a proof
method of definite clauses. There is some similarity to inductionless induction
based on the Knuth-Bendix algorithm, b u t its principle is different.

The first and second papers in the "Nonmonotonic Reasoning" section deal
with the problem of belief revision caused by adding new facts. The third paper
gives a new formalism to the problem of nonmonotonic reasoning associated with
inheritance and proposes a parallel algorithm to solve it.

"Logic p r o g r a m m i n g L a n g u a g e / S y s t e m " c o n t a i n s t h r e e p a p e r s on
debugging, explicit control of logic programs and a new programming paradigm
called lattice programming. This last one is particularly interesting both from
theoretical and practical viewpoints. It associates for each computing path of
Prolog programs a value other than true/false, thus increasing its expressive
power.

The "Guarded Horn Clauses" section includes two papers. The first tries to
realize reflection on GHC, and the second paper gives an implementation of full
GHC by communication processes.

The next "Knowledge Base System" section contains two papers. The first
one gives inference methods and semeantics for a knowledge base containing
disjunctive non-Horn clauses. The latter describes a formalism of conventional
network type database access in terms of a refutation procedure.

The "Computer-aided Reasoning" section consists of just one paper. The

paper describes a system to assist h u m a n reasoning, namely a very interesting
and seemingly useful application program written in ESP, a dialect of Prolog that
can be used to write programs in object-oriented style.
 Yl

The last section is "Natural Language Processing", which contains three

papers. The first paper concerns lexical analysis of Japanese sentences. An
efficient algorithm and its implementation in ESP is given. The second paper
presents a system which generates a summary of a given sentence. The system
achieves this difficult goal by restricting the reference domain, thus making it
possible to understand the described situation. The last paper proposes a new
knowledge representation formalism suited for representing dynamic behaviors
by extending the frame system. It also points out the necessity of reasoning in
different situation which depends on multiple world representation.

We thank program committee members, reviewers and authors for their

cooperation and patience. We extend our special thanks to reviewers from IBM
Watson Research Center for their thorough efforts to improve the quality of the
English as well as the contents.

Tokyo, Spring 1989 Koichi Furukawa

Hozumi Tanaka
Tetsunosuke Fujisaki
Table of Contents

Foundation of Logic Programming

A. Yamamoto
Completeness of Extended Unification Based on Basic
Narrowing 1
A. Sakurai and H. Motoda
Proving Definite Clauses without Explicit Use of Inductions 11

Nonmonotonic Reasoning

H. Yuasa and S. Arikawa

Pseudo Extension in Default Reasoning and
Belief Revision by Model Inference 27
T. Shintani
An Approach to Nonmonotonic Inference Mechanism in Production
Production System KOREfIE 38
C. Sakama and A. Okumura
Nonmonotonic Parallel Inheritance Network 53

Logic Programming Language/System

H. Tamura and H. Aiso

Logic Programmig Debugger Using Control Flow Specification 67
M.Hoshida and M. Tokoro
ALEX: The Logic Programming Language with Explicit Control
and without Cut-Operators 82
K. Fukunaga and S. Morishita
Lattice Programming Methodology 96
 Viii

Guarded Horn Clauses

J. Tanaka
A Simple Programming System Written in GHC
and Its Reflective Operations 108

M. Terada
Implementation of Full GHC by Communicating Processes 122

Knowledge Base System

H. Sakai
Inference Methods and Semantics on Or-type Knowledge Bases 136
M. Takizawa, M. Katsumata and S. Nagahora
Access Program to Minimize Redundant Refutations on
the Network Database System 156

Computer - Aided Reasoning

T. Minami, H. Sawamura, K. Satoh and K. Tsuchiya

EUODHILOS: A General-Purpose Reasoning Assistant System
- Concept and Implementation - 172

Natural Language Processing

R. Sugimura, K. Akasaka, Y. Kubo and Y. Matsumoto

Logic Based Lexical Analyser LAX 188
N. Tamura, T. Kamio and M. Kurosaki
Extraction of Characteristic Facts and Abstract Generation 217
S. Kinoshita, H. Sano, T. Ukita, K. Sumita and S. Amano
Knowledge Representation and Reasoning for Discourse
Understanding 238
Executive Committee / Program Committee / Editorial Board

Hozumi Tanaka (Conference Chairman) Tokyo Institute of Technology

Koichi Furukawa (Program Chairman) ICOT
Tetsunosuke Fujisaki IBM
Eiiti Wada The University of Tokyo
Kazuhiro Fuchi ICOT
Yuji Matsumoto Kyoto University
Hitoshi Aida The University of Tokyo
Shigeki Goto NTT
Susumu Kunifuji HAS-SYS Fujitsu Ltd.
Toshiaki Kurokawa IBM Japan
Hiroshi Nakagawa Yokohama National University
Toshihiko Ninomiya Hitachi, Ltd.
Katsumi Nitta ETL
Etsuya Shibayama Tokyo Institute of Technology
Masayuki Takeda Science University of Tokyo
Kazunori Ueda ICOT
Hideki Yasukawa Matsushita Electric Industrial
Co.,Ltd.

Referees

A. Aiba H. Aida K. Appleby-Hougham S. Arikawa

A. Atarashi T. Chikayama H. Fujita K. Furukawa
Y. Futamura K. Futatsugi S. Goto K. Handa
K. Hashida R. Helm K. Hiraki N. Ichiyoshi
K. Inoue H.Ishizaka M. Ishizuka T. Kanamori
H. Koike H. Komatsu A. Konagaya S. Kono
S. Kunifuji T. Kurokawa J-L. Lassez H. Matsubara
A. Matsumoto Y. Matsumoto T. Minami N. Miyazaki
T. Miyazaki R. Mizoguchi Y. Morita K. Mukai
M. Murakami H. Nakagawa K. Nakajima A. Nigam
M. Nilsson K. Nitta N. Noguchi M. Numao
M. Ohki K. Ohno I. Ohsawa A. Ohsuga
H. Ohwada N. Oka Y. Ravin K. Sakai
C. Sakama H. Sato T. Sato J. Sawamoto
H. Sawamura M. Schor H. Seki P. Sheridan
E. Shibayama H. Sugano T. Sugawara M. Takeda
A. Takeuchi H. Taki K. Taki H. Tanaka
M. Terada S. Tomura K. Ueda S. Uzuhara
S. Wada W. Wilson J. Yamaguchi H. Yasukawa
K. Yokota A. Yonezawa K. Yoshida
 Completeness of Extended Unification Based on Basic Narrowing

Akihiro Y A M A M O T O

Department of Information Systems

Interdisciplinary Graduate School of Engineering Sciences
Kyushu University 39
6-1, Kasuga-koen, Kasuga, Fttkuolm, 816 JAPAN

Mailing Address: Research Institute of Fundamental Information Science

Kyushu University 33, Fukuoka, 812 JAPAN

Abstract
In this paper we prove the completeness of unification based on basic narrowing.
First we prove the completeness of unification based on original narrowing under a
weaker condition than the previous proof. Then we discuss the relation between basic
narrowing and innermost reduction as the lifting lemma, and prove the completeness
of unification based on basic narrowing. Moreover, we give the switching lemma to
combine the previous algorithm and our new proof.

1. I n t r o d u c t i o n ,
The main contribution of this paper is a proof that the extended unification procedure based on
basic narrowing is complete under some conditions. The procedure was first given by Hultot [2],
and then improved by Bosco et al. [1]. The improved procedure [1] was obtained by translating
every basic narrowing sequence into an SLD-refutation sequence. It is now very important as a
foundation of logic-functional programming languages, but the proof of its completeness is not
straightforward.
Hullot [2] proved the lifting lemma for basic narrowing as that of SLD-resolution. However,
his proof is incomplete since we found a counter example for Lemma 3 [2]. To prove the lifting
lemma for basic narrowing, we need to construct a basic reduction sequence from every term, but
its existence is not guaranteed in general. By clarifying the function of knnermost reduction, we
overcome this difficulty and show the existence.
We first show, under a weaker condition than [2], the completeness of unification based on
original narrowing. That is, the term rewriting system is simply assumed weakly canonical. Then
we establish the relation between innermost reduction and basic narrowing as the lifting lemma~
and prove the completeness of unification based on basic narrowing. Thus "we suppose the con-
dition that the term rewriting system is both weakly canonical and normalizing with innermost
reductions. Moreover, we show the relation between the completeness theorem and the algorithm
in [1] by clarifying the procedure which searches an innermost redex. Thus the relation is stated
as the switching lemma.
The paper is organized as follows. In the next section, we explain narrowing and basic
narrowing. In Section 3, we prove the completeness of unification based on narrowing under a
weaker condition t h a n that in [2], and discuss the problem in the case of basic narrowing. In
Section 4, we prove the completeness of unification based on basic narrowing under a weaker
condition t h a n t h a t in [2]. In Section 5, we discuss the relation of our proof and the algorithm
given in [1] as a form of the switching lemma.

2. Narrowing and Basic Narrowing.

In the present paper we treat extended unification in a first order equality theory E using a term
rewriting system R. An E-unifier of two terms t and s is a substitution 0 such that E ~ tO = sO,
where ~ represents logical consequence and the predicate symbol = is interpreted as equality.
We consider the case t h a t E is given as a reflexive symmetric transitive closure of the reduction
relation on R, which is defined below.
T h e symbol - denotes syntactical identity. T h e set of all variables occurring in a t e r m t is
denoted by V(t). We use the special function symbol eq only for the description of procedures.
We assume that the t e r m rewriting system R = {ai --~ ~i;i E t } satisfies the following
conditions:
(i) v( d,
(ii) a i is not a variable,
(iii) I is finite.
We define the occurrence of every subterm of a term t as a sequence of positive integers as
follows:
(i) T h e occurrence of t is the empty sequence (denoted by A),
(ii) If the occurrence of the subterm f ( t l , . . . , tn) is u, then the occurrence of the t e r m ti is
u.i.
Let t be a term. T h e n the symbol Oc(t) denotes the set of occurrences of all subterms of t,
and Og(t) denotes the set of occurrences of all non-variable subterms of t. W h e n u e Oc(t), t/u
denotes the subterm of t whose occurrence is u, and t[u ~ s] denotes the t e r m which is obtained
by replacing subterm t / u by a term s. A subterm of t whose occurrence is not A is called a proper
subterm of t.
Let u and v be two occurrences. T h e n we write u ~ v if there is a sequence w such that
?/'W ~--- V .

D e f i n i t i o n 1. Let t and s be terms. T h e n t is reduced to s if there exist u C Oc(t), a --* fl E R,

and a substitution 0 such t h a t f l u =- aO and s =- t[u ~ fl0]. W h e n t is reduced to s, we write
t --* s, or t "-*[-,~-*~,~l s to give fult details. We call t / u a redex of t.

W h e n we regard --* as a binary relation on the set of terms, we call it the reduction relation
on /:~.

A t e r m t is in normM form if there exists no term s such that t --* s. If there exists a finite
reduction sequence
t -~ to --* tl --+ . . . - - * t n =_ s

such that s is in normM form, then s is called a normaI form o f t and is denoted by N(t). Moreover,
let 8 be a substitution {X1 *-- t l , . . . , Xn ~- tn }. If each of t l , . . •, t,~ is in normal form, then 0 is said
to be in normal form. For every substitution 0, the substitution {X1 *-- N ( t l ) , . . . , X , e-- N(t~)},
if it exists, is called a normM form o f O.
D e f i n i t i o n 2. Let t and s be terms. T h e n t is narrowed ~o s, if there exist u ~ Og(t), a variant
--+/3 of a rule in R, and a substitution ~ such that a is an mgu of f l u and ce and s --- (t[u ~/3])~.
W h e n t is narrowed to s, we write t ~ s, or t ~ D , ~ Z , ~ ] s to give full details.

We assume t h a t for every narrowing sequence

t0 ==~[u0,a0--+#0,a0] tl =~Z[u~,al-~fll,at] t2 ==> " " =--~[u.-i,a,-1--/~,-,,~-i] tn,

the variables in every ai -+ /~i are standardized apar~ ([3]). W e also assume that mgu's are
c o m p u t e d by the algorithm in [3, 4]. T h e n the E-uniflcation of two terms t and s based on
narrowing is to generate the set

there exists a narrowing sequence ]

~(t,s) = { o eq(t, s) =-- t o =e~[. . . . o--~o,~o] t:t =:~[. . . . . . ~,a~] t2 =>

L
"'" =~[. . . . . . . . . . ~. . . . . . . t] t~ -- eq(t',s') f
such that cr = ao ... ~n-l#lv(t)uv(8), where # is an mgu of t I and s r

where 8Iv denotes the restriction of a substitution 0 to a set of variables V.

For two substitutions 0 and r, we write 0 =/~ r if E ~ XO = X r for every variable X.

D e f i n i t i o n 3. A set of substitutions E is a compIete set o£ E-uni/iers o£ t and s if, for every

E-unifier 0 of t and s, there exist a E P, and a substitution 7 such that 81y(,)uy(,) = E ~rTIv(0uy(,).

D e f i n i t i o n 4. E-unification based on narrowing is complet;e if, for every two terms t and s,
P,(t,a) is a complete set of E-unifiers of t and s.

Basic narrowing is introduced in order to improve E-unification based on narrowing.

D e f i n i t i o n 5. A narrowing sequence

to ='-'-~[u0,a0--+~0,a0]1~1 =k[ul,~l--.[Sl,a,] t2 =g" "'" =>[u,-t,a.-1-+~,-1,~.-1] tn

is basic if there exists a sequence of sets of occurrences U 0 , . . . , Un which satisfies the following
conditions:
(B-l) U0 = Og(to),
(B-2) ui C Ui C Oc(ti) (0 < i < n - 1) and Un C Oc(tn),
(B-a) U~+I=(U~-{v~Ui;u~_v})U{ui.v;v~Og(~0} (0<i<n-1).
Example 1. Let R = { f ( X , b) -+ g ( X ) , a --~ b } . T h e n a narrowing sequence

f ( a , X ) => f(b, X ) ==~g(b),

is basic because the sequence

U0 = {A,i}, U1 = {A, 1}, U2 = {A)

satisfies the conditions (B-l), (B-2), and (B-a). However,

f ( a , X ) :& g(a) => g(b)

is not basic because U0 and U1 must be {A, 1} and {A} respectively so that Uo satisfies (B-I),
(B-2) and (B-3), but [/1 violates (B-2).

T h e E-unit~cation of two terms t and s based on basic narrowing is to generate the set

there exists a basic narrowing sequence ]

r~B(4,s)={ eq(t, s) =--4o ~[uo,~o--*~o,~0] 41 ~ [ ~ , ~ , ~ ] 42

"'" ~D.-1,,,~-t-*~.-, . . . . 1] tn --- eq(tt, s ')
such that a = a0 " "an-l#lv(ouv(s), where/z is an m g u of 4' and s r

as in the case of narrowing.

Definition 6. E-unification based on basic narrowing is complete if, for every two terms t and
s, ~B(t, s) is a complete set of E-unifiers of t and s.
[1] has improved E-unification based on basic narrowing by using innermost occurrences and
has given an efficient algorithm as explained below.

Definition 7. Let U be a set of occurrences of subterms of t. An occurrence u is innermost in

U if there is no occurrence v in U such that u _ v and v ~ u.

Algorithm 1 (Bosco~ Giovannetti and Moiso [1]).

unify(t, s, S) /* 4 and s are terms and S is a rule to select an innermost occurrence from B */
E := eq(t, s), B := Og(eq(t, s)) - {A}, and ® := e
/ * E is a term, B is a set of occurrences, and O is a substitution * /
LOOP:
i f B = ¢ then
O := O/z and return Olv(t)uv(s) w h e r e / ~ is an mgu of # and a r w h e r e E = eq(#,s ~)
else
select an innermost occurrence u from B according to S,
nondeterministically execute either (1-1) or (1-2), and goto L O O P :
(>1) B := B - (u}
(1-2) select nondeterministically a variant a -~ ~ of a rule in R such ttmt
all the variables in V(a) are new and E / u and c~ are unifiable~
S := (E[u ~ ;31)~r, O := @~, and B := (B - {u}) O { u . v ; v e Og(3)}
w h e r e a is an mgu of E / u and
end i f

T h e search space of basic narrowing is decreased in Algorithm 1 by introducing the compu-

tation rule S (or the selection rule in [1]), which selects one of the innermost occurrences from
the set B. In a basic narrowing sequence, the terms introduced by an mgu may not be selected
afterwards. Algorithm 1 satisfies the conditions of a basic narrowing sequence by selecting an
innermost occurrence from B.
In the following sections, we reconsider the conditions for the completeness of the above three
algorithms.
3. T h e P r o b l e m o n Basic N a r r o w i n g
In this section we point out the problem on the completeness of E-unification based on basic
narrowing. For the purpose of the discussion, we extend the completeness of E-unification based
on naxrowing. It was proved that E-ratification based on narrowing is complete if R is canonical,
that is, confluent and finitely terminating ([2] Theorem 2). The following theorem is the lifting
lemma for narrowing proved in [2].

T h e o r e m 1 ( H u l l o t [2]). Let t and s be terms such that t =- szl for some normal substitution
TI. Then for a reduction sequence

t --= tO "-'~[u0,ao~flo,00]tl ---~[u~,~w-~fll,01]t2 --+ "'" -'+[u,-i,a.-1--fl,-t,a,-i] tn,

there exists a narrowing sequence

and a sequence of normal substitutions ~ o , . . . , zl,~ such that

T0 = ~h r/Iv(s) : a o ' " c r i - l ~ i l v ( s ) (1 < i < n),

ti = s i ~ i (0_<i_<n).

The assumption of Theorem 1 is only that ~ is in normal form. Therefore we can easily
weaken the condition for Hullot's Theorem.

D e f i n i t i o n 8. R is weakly canonical if for every term t there exists only one normal form N ( t )
and for every pair of t and s
E~t=s ~ g(t)=N(s).

The following theorem can be proved in the same way as in [2].

T h e o r e m 2 ( E x t e n s i o n o f H u l l o t ' s T h e o r e m ) . /T R is weakly canonical, then E(t, s) is a

complete set orE-unifiers o f t and s.

P r o o f . Let 0 be an E-unifier of t and s. Because R is weakly canonical, we can let r} be a normal

form of 0. Also we can construct a reduction sequence

eq(t, s)~l =- to -+ tl - + ' " -+ t~ = eq(N(trl) , N(srl)),

such that N(t~l) :- N(s~l). Thus by Theorem 1 there is a narrowing sequence

eq(~, s ) = so ~ sl ~--. ~ s, - eq(t',s').

Then t I and s t are unifiable because there is a substitution r/n such that tl~n -~ Sl?}n. I

To show that the above theorem is a proper extension of HuUot's Theorem, we give an example
which shows the difference between canonicity and weak canonicity.

E x a m p l e 2. Let R = {repeat --, repeat, repeat --~ halt}. Then R is weakly canonical. For every
term t, its unique normal form is a term given by replacing every repeat in t by halt. Moreover,
we can decide E ~ t = s by comparing N(t) and N(s). However, R is not finitely terminating, so
R is not canonical.

Now our problem is to prove the completeness of E-unification based on basic narrowing. A
result similar to Theorem 1 is needed for basic narrowing. Thus we introduce basic reduction.

D e f i n i t i o n 9. Let t and s be terms such that t - s8 for some substitution ~. Then a reduction
sequence

t = to -'*[u0,~0~#0,00] tl -'+[~1,a1-*~1,0~]t2 --~ "'" --+[u,-1,a,-t--*#,-t,e,-1] tn

is based on Og(s) if there exists a sequence of sets of occurrences U 0 , . . . , U,~ which satisfies the
conditions (B-2) and (8-3) in Definition 5 and the following condition (8-1') instead of (8-1):
(B-I') U0 = Og(s).

E x a m p l e 3. Let R = {f(X, b) -~ g(Z), a --* b} . In the same way as in Example 1, we can

show that a reduction sequence
f(a, b) -~ y(b, b) -~ g(b)

is based on Og(t), but

f ( a , b) -~ 9(a) -* g(b)
is not.

To get a theorem corresponding to Theorem 1, it is asserted in [2] (Lemma 3) that

if t and s have the relation t - sT for some nornmlized substitution r/, then
every reduction sequence from s is based on Og(s).
However the second sequence in Example 3 does not satisfy the above assertion. Therefore the
proof of the completeness of E-unification based on basic narrowing does not work. Algorithm 1
is also incomplete because the proof of the completeness of Algorithm 1 in [1] is again due to the
above assertion.
Algorithm 1 is found by using the one-to-one correspondence between SLD-refutation and
basic narrowing sequence in [1]. The efficiency of SLD-resoiution is due to the computation ruie
([3]). Algorithm 1 uses not only the computation rule but also innermost occurrences. In the first
sequence of Example 3, the innermost redexes only are selected. On the other hand, in the second
one, f(a,b) is selected and it is not an innermost redex. This selection makes a counter exasnple
for Lernma 3 [2]. In general, when we construct a basic reduction sequence from a term t to N(t),
the terms introduced by the substitution at each reduction need to be in normal form because
they may not be selected as redexes. From these consideration, we suggest that the innermost
occurrence should essentially owe to the completeness of E-unification based on basic narrowing.
In the next section, we prove the completeness of E-unification based on basic narrowing
along the consideration above.
4. Completeness of Unification Based on Basic Narrowing
In the context of term rewriting systems, the word 'innermost' is not used as the meaning defined
in Definition 7.

Definition 10. Suppose that t --~{u,a-~,e] 8 for some term s. Then t / u is an innermost redex
and the reduction is called an innermost reduction if any proper subterm of t / u is normal.

E x a m p l e 4. Let R = {f(X, b) ~ g(X), a ~ b}. Then the innermost redexes o f f ( f ( a , b), f(b, b))
are a and f(b,b). Thus
f ( f ( a , b), f(b, b)) --~ f ( f ( a , b), g(b))

is an innermost reduction, but

f(f(a,b),f(b,b))-~f(g(a),f(b,b))

is not.

We will clarify the relation of the two definitions of 'innermost', which are in Definition 7 and
in Definition 10, as an algorithm in Section 5.
As suggested by Example 3, innermost reduction furnishes a way to construct the basic
reduction sequence. The following lemma is a key result for our proof of the completeness of
E-unification based on basic narrowing.

L e m m a 1. Let t be a term and 71 be a normal substitution. I r a reduction sequence

tr/~ to --+ "~1 ~ "'" --+ ~'n

consists only of innermost reductions, then the reduction sequence is based on Og(t).

P r o o f . The proof is by induction on the length of the reduction. Suppose first that n = 1, and
consider the reduction sequence
iT/-- ~o --+[uo,~o-+fl,oo]tl.
Then the occurrence of every redex of to is in Og(t) because y is normal. Thus the result holds
by letting
Uo = o g ( t ) ,
u1 = (v0 - {v E u0 ; ~0 ~ v}) u { ~ 0 - v ; ~ e og(~o)).
Now suppose that the result holds for n. Consider the sequence

trl ==-to --~ " " --~ tn-1 "-~[u._~,a._,~fl,_~,0._l] tn --+[u,,a,--.fl.,o.] tn+l.

From the induction hypothesis there is a sequence of sets of occurrences [To,..., Un which satisfies
the conditions (B-I'), (B-2), and (B-3) for the sequence

to --> "" --~ tn-1 --~ tn.

Because t , - 1 / u n - 1 is aza innermost redex, all of its proper subterms are normal, and ~n-1 is also
normal. Thus the occurrence of every redex o f ~ n - l ~ n - 1 is in Og(fln--1). So because un C U,, the
sequence V0,..., Vn+l satis~es the conditions (B-V), (B-2), and (B-3) by lettin,

If R is canonical, then R has the property that there is an innermost reduction sequence from
every term t to N ( t ) because any reduction sequence from t reaches N(t). However the property
is not confirmed in general. Therefore we need to identify the class of term rewriting systems with
this property.

D e f i n i t i o n 11. R is normalizing wi*h innermost reductions if for every term there exists an
innermost reduction sequence to its normal form.

Then the completeness of E-unification based on basic narrowing is proved using the innermost
reduction sequence.

T h e o r e m 3. If R is weakly canonical and normalizing with innermost reductions, then X e ( t, s)

is a complete set orE-unifiers o f t and s.

Proof. Let 8 be an E-unifier of t and s, arid ~7 be its normal form. Because R is normalizing
with innermost reductions, there exists an innermost reduction sequence

eq( t, s)~ = to --~[u0,~0~0,oo] tl -~[. . . . ~-~t,0~] t2 - * . " --+[. . . . . . . . . . p. . . . o._ ~] t , -- eq( N ( t~), N( s~) ).

By Lemma 1 there is a sequence o£ sets of occurrences U 0 , . . . , Un which makes the reduction

sequence basic. By Theorem 1, we can construct a narrowing sequence

Then the narrowing sequence is also basic by using U 0 , . . . , U,, because the two sequences use the
same occurrence at each step. Moreover, t f and s t are unifiable, and there is a substitution 7 such
that
g o ' " e~-l#Tiv(,)uv(~) = rliv(,)uv(~)

where/z is an mgu of t I and #.1

If R is canonical, then R is weakly canonical and normalizing with innermost reductions

because every reduction sequence from a term t reaches N(t). Thus Theorem 3 holds for canonical
term rewriting systems.
5. S w i t c h i n g L e m m a for B a s i c N a r r o w i n g
In this section we discuss the relation between Theorem 3 and Algorithm 1. The most important
point of Algorithm I is the computation rule S. Thus we give the switching lemma that confirms
the introduction of S into basic narrowing as in the theory of SLD-resolution.
At first we consider a procedure for innermost reduction. The following Algorithm 2 is the
representation.

A l g o r i t h m 2.

innermost_reduce(t) /* t is a r~erm */
T := t, B := Oc(t), /* T is a term, and B is a set of occurrences */
LOOP:
ifB = ¢ then
return T
else
nondeterministically select an innermost occurrence u from B,
if there is no rule ~ -~/3 in R such that T / u - aO for some 0 t h e n
(2-1) B := B - {u} and goto LOOP /* T / u is in normal form*/
else
nondeterministically select a --* fl from R such that T / u = aO for some 0
T := T[u ¢= ~0], B := (B - {u}) O {u. v ; v e Og(~)}, and goto LOOP
end if
e n d if

In Algorithm 2 the procedure which searches an innermost redex from a term is clarified by
using the set B and step (2-1). This step shows the difference between Definition 7 and Definition
t0.
Now we apply Algorithm 2 to the construction of the basic reduction sequence in the proof
of Theorem 3. The reduction sequence and the basic narrowing sequence in the proof use the
same sequence U o , . . . , Un. Thus it is clear that we may add a step corresponding to (2-1) to the
procedure for basic narrowing. In Algorithm t the step is (1-1).
Then in order to prove the completeness of Algorithm 1 using Theorem 3, we need to show
that we may restrict the selection of innermost occurrences with a computation rule S. This
restriction is allowable because of the following switching lemma for Algorithm 1. To explain the
lemma, in Algorithm 1, we use symbols E(S, n), B(S, n), O(S, n) to represent the values of E, B,
and O at the beginning of the n-th execution of the LOOP when the computation rule is S.
L e m m a 2. Suppose that there are two innermost occurrences u and v in B ( S , n) and that S
selects u in the n-th execution of the LOOP and v in the (n + 1)-st. Then we m a y use the
computation rule S t which selects v instead of u at ~he n-th execution o f LOOP and u at the
(n + 1)-st. Moreover we can lot E(S, n + 2) and E(S', n + 2) be variants, e ( s , n) and e ( s ' , n) be
~ i ~ t s , and B ( S , ~ + 2) = B ( S ' , n + 2).

It should be noticed that Le~mna 2 is proved by translating every basic narrowing sequence
to an SLD-refutation and applying the switching lemma for SLD-refutation in [1]. Of course, the
false assertion was used in [1]. Thus we omit its proof.
 10

6. Conclusion
We have proved the completeness of E-unification based on narrowing under a weaker restriction
and have proved the completeness of E-unification based on basic narrowing, and have given the
switching lemma. The discussion in [1] is important but was only half way to the completeness
because it showed only the switching lemma.
In the discussions above, we have used the conditions that R is weakly canonical and is
normalizing with innermost reductions. Both [1] and [2] treat the case that R is a canonical
term rewriting system, which is useful for theorem provers using reduction. In the discussions on
narrowing, the restriction for term rewriting systems should be weakened because E-unification
based on narrowing is a refutation procedure. We have shown that it can really be weakened.
We need to find sufficient conditions under which a term rewriting system is weakly canonical
and normalizing with innermost reductions. Especially, we have not clarified any properties of
term rewriting systems which are normalizing with innermost reductions. We conjecture that
weakly canonical term rewriting systems are normalizing with innermost reductions.
We need also to import the property of SLD-resolution other than the switching iemma into
the theory of narrowing. For example, we are considering the complete definitions of functions in
the same way as those of predicates in SLD-resolution.

Acknowledgements
The author would like to thank Prof. Setsuo Arikawa for his constructive comments and en-
couragements. The author also wishes to thank Dr. Makoto Haraguchi for many discussions on
narrowing and SLD-resolution.

References
[1] Bosco, P.G., Giownnetti, E., and Moiso, C. (1987): Refined Strategies for Semantic Unifica-
tion, Proc. TAPSOFT '87 (LNCS 250), 276-290.
[2] Hullot, J.M. (1980): Canonical Forms and Unification, Proc. 5th Conference on Automated
Deduction (LNCS 87), 318-334.
[3] Lloyd, J.W. (1984): Foundations of Logic Programming, Springer-Verlag.
[4] Martelli, A. and Montanari, U. (1982): An Efficient Unification Algorithm, AC1VITOPLAS,
4(2), 258-282.
[5] Yamamoto, A. (1987): A Theoretical Combination of SLD-Resolution and Narrowing, Proc.
4th ICLP, 470-487.
 Proving Definite Clauses without Explicit Use of Inductions*
Akito S A K U R A I Hiroshi MOTODA
Advanced Research Laboratory, Hitachi Ltd.
1-280 Higashi-Koigakubo, Kokubunji, Tokyo 185, JAPAN
A b s t r a c t . We propose an algorithm to prove a definite clause is modeled
by an intended minimal model of a set of definite clauses. In the algorithm,
no induction scheme will be used explicitly, although in effect a mathematical
induction is implied. This algorithm corresponds to an inductionless induc-
tion method using the Knuth-Bendix algorithm in an equation problem. It
resembles the completion algorithm, but in fact it is a completely different
algorithm.

1. I n t r o d u c t i o n
The problem we are concerned with is how to determine if a definite clause is valid under
an intended minimal model of a set of definite clauses. Let us call this an implication
problem, just as the problem to determine the validity of equalities under some theory
is called an equa~ty problem.
The implication problem is the one to be solved by a theorem prover for a first-
order logic. It is, however, well known that if the theory includes inductive definitions
it is hard to determine the validity of clauses. The reason lies in the difficulties of an
inductive proof whose construction is necessary when objects of the theory are defined
inductively.
To implement an induction mechanism in a system, you have to implement heuris-
tics of induction pattern or generalization ([2] etc.). Similar difficulties lie in the equality
problem. Many researchers have tried adoption of the Knuth-Bendix completion algo-
rithm to the equality problem with an idea that if the system augmented with the
new equality to be proved can be completed by the algorithm (and satisfies certain
conditions), then the equality is consistent with the original system([6],[8]).
Can the same approach be used in the implication problem? Unfortunately "No"
is the answer. For example, Hsiang's system [4] seems not to have enough capability to
prove inductive theorems. Although Paul [11] claims his system can conduct inductive
proofs, it usually "proves" a theorem by fMling to halt.
Hsiang [5] proposed an extension of Prolog to incorporate the ability to conduct
inductive proofs. But the ability is quite limited. Firstly, the implication problem
targeted by his system is restricted so that an implication to be proved contains only
equalities in its consequence part. Secondly, the transformations adopted in the system
are so weak that efficient lelnmas are not deducible. For example, the system cannot
express an implication problem with a non-deterministic predicate in its consequence
part, as in the perm-perm problem in Appendix A.3. The rev-rev problem in Appendix
A.2 can be expressed but not proved.
We propose in this paper an algorithm to solve the implication problem, by which
inductive proof is made without using induction explicitly, which is different from Knuth-
Bendix's but to which it bears some resemblance. In Appendix B, we show an example

* This paper is a revised and extended version of [12]

 12

which is easy to understand but which conveys key ideas of the proposed algorithm.
Readers are advised to read it first• In section 2, we introduce very briefly some ter-
minologies, and in section 3, we describe the algorithm. The conditions necessary to
guarantee the validity of a proof or a disproof will be found in section 4. T h e justification
of the conditions is briefly described in section 5.

2. P r e l i m i n a r i e s
In this section we give necessary definitions of terminology to be used•
Syntax
We distinguish between sets of variables, function symbols and predicate symbols. A
constant is a 0-ary function symbol. A term is a variable, a constant, or of the form
f ( t l , . . . , t ~ ) where f is an n-ary function symbol and t l ~ . . . , t n are terms. An atomic
formula (or atom for short) is of the form P ( t l , . . . , tn), where P is an n-ary predicate
symbol and t l , . . . , tn are terms. An atom is a ground atom if it contains no occurrence
of any variables• A definite clause is a clause of the form A +-- B 1 , . . . , Bn n > 0
where A, B 1 , . . . , B~ are atoms, and variables occurring in them are universally quan-
tified. A definite clause is a definite ground clause if it contains no occurrence of any
variables. A substitution ¢r = { X i / t i } is an operation defined on an atom, which replaces
simultaneously all occurrences of given variables Xi with given terms ti. The result is
called an instance, and is denoted by Act for an atom A. A substitution is a ground
substitution if the result is ground. This naturally extends to a set of atoms, clauses
and programs. We use the following terminology in our own meaning.
• We denote an array of variables with a boldface letter.
V(F) and 3 ( F ) are the universal and existential closure of any formula F , respec-
tively.
. A program is a set of definite clauses.
• A conjunct is a conjunction of atoms, and a disjunct is a disjunction of conjuncts.
• An extended definite clause is a formula of the form r ~-- A, where e-- is the
usual logical implication, F and A are respectively a disjunct and a conjunct with
some variables existentially quantified in each conjunct, and with all other variables
universally quantified in the formula. For the sake of clarity and convenience,
existentially quantified variables in different conjuncts are named distinctly. In
other words an extended definite clause is V ( 3 X l r l v . . . v 3XnFn ~'- 3 Y A ) , where
F~ and A are conjuncts.
• If it is not ambiguous, I~1 V .. • V Fn ~-- A or F ~ A is used instead of V(3X1F1 v
• .. 3XnFn ~- 3 Y A ) .
. An internal variable is an existentially quantified variable.
• For a substitution cr = {Xi/ti}~=l, we consider only the substitutions where no Xi
appears in any tj. In our context a substitution so restricted is always available if
we choose appropriate variants of formulae for c~ to operate on.
• Equality (--) as a predicate satisfies equality theory [10], and can be considered
as defined by the following clauses: = ( a , a ) . = ( ~ ( X 1 , . . . ,Xk), ~ ( Y 1 , . . . ,Yk)) +--
=(X1, Y 1 ) , . . . , = ( X k , Yk) where a represents an arbitrary constant symbol, qo an
arbitrary function symbol~ and k its arity.
M ~ G denotes that a formula G is correct under the model M .
MR is a minimal model of a program P. For details, see the following "Semantics".
 13

Note that we do not presuppose any interpreter (i.e., computation rule and search
rule). So, two conjuncts A, B and B, A (or two disjuncts AV B and B v A) are equivalent
in all means.
Semantics
The concept of "correctness" should be made clear before entering into discussing
proofs of "a goal G is correct under a program P". It is natural to consider that a
closed formula VZ(3XF . - 3YA) is correct under a program P, if and only if, for any
ground substitution ~ of the variables Z, " P ~- 3Y~o~ entails P ~- 3XlPo~" where ~- will
be discussed later. Then what should be the meaning of P ~- 3YA~ (or P ~- 3XFo-) ?
It is also natural to consider a completion of P, that is comp(P), as the meaning of
a program P rather than P itself (i.e.,the logical meaning of the program). Especially
when a program P is viewed as defining predicates inductively, employing comp(P)
corresponds to adding to a definition of each predicate the statement '~Only those defined
by the above are ...".
Defining a transformation, Tp from interpretations to interpretations, associated
with a program P under the fixed point semantics, the least fixed point of Tp is con-
sidered to be the meaning of the program P. Note that, since the optimal fixed point
also becomes the minimal fixed point under logic programming, the minimal fixed point
semantics is more appropriate [9l, Adding to this advantage, the fixed point of Tp is
also a model of comp(P) if augmented with appropriate equational theories [10]. We
adopt this minimal fixed point formalism as the semantics.
An Herbrand universe is commonly used as a domain of interpretations. There are
two reasons for this. The first is that the minimal fixed point of Tp on an Herbrand
universe coincides with the minimal Herbrand model/]V/p. The second is that G is a
logical consequence of P if and only if G is valid under/-Mp [10] when negation of the
goal G is of clausal form (general goal clauses of Prolog have this form)
But in our implication problem, naive use of the Herbrand universe is not appro-
priate. The cause is rooted in the existence of ¢r stated above. Terms substituted for
by ~r can not be restricted to the elements of Hfp. In fact if we so restrict, something
unpleasant results:
Set P = {q(a),VXp(X)} and let a goal to be proved a be VX(q(X) ~- p(X)).
Obviously IMp = {q(a),p(a)} implies IMp ~ G. But intuitively we believe that G
cannot hold since q(X) may not so for some X other than a. Moreover, G is not a
logical consequence of P.
The cure is easy. Since the trouble is ascribed to substitutions which may replace
variables with terms consisting of constants other than predicted, we adapt the Herbrand
universe to include an enurnerable (by LSwenheim's theorem) set of constants including
all that we need in future, and call it the extended Herbrand universe Up. We construct
the extended Herbrand base Bp in the same way as we construct the Herbrand base.
Since the difference between//Jp and Up is only the existence of additional constants,
constructions of models, a least model Mp, fixed point and other semantics, etc. are
no different from conventional ones. The implication problem is, therefore, to find if
Me ~ P ,--/X.
 14

3. E q u i f o r m a l i z a t i o n A l g o r i t h m

In this section we describe the proposed algorithm. The algorithm proceeds, trying to
reduce difference of form on both sides of implication while conserving correctness of
implications (the details are not discussed in this paper), to make b o t h formulae be of
the same form. Hence we call it the eqniformalization algorithm.
Following the equiformalization algorithm, we non-determinlstically develop a tree
called a transformation tree. If one non-deterministically constructed transformation
tree is proved to satisfy a proof-condition stated in section 4, then the algorithm termi-
nates with a positive answer. If one tree satisfies a disproof-condition, then the answer
is negative. If all the trees constructed non-deterministically satisfy a failure-condition,
then the total algorithm (including all non-deterministic branchings) terminates with
failure. There also exist some cases when the algorithm never terminates. In this sense
the name 'algorithm' may be inappropriate. Since the name 'Knuth-Bendix algorithm'
is used fairly commonly, we use the word too. At each step of the algorithm, a trans-
formation tree is further transformed by expanding a leaf to a subtree.
Throughout this paper, H, with possible further qualifications, denotes an atom; F,
A, A and II, with possible further qualifications, usually denote conjuncts, but denote
disjuncts in an implied part of ~---; ~, 8 and ~, with possible further qualifications, denote
substitutions.
D e f i n i t i o n 3.1 A transformation tree is a tree, in which
(1) on each node is an extended definite clause (possibiy empty on a Ieaf node),
(2) nodes with an empty clause and nodes marked false have no descendants,
(3) the descendants of a node are constructed by one of the procedures in Transforma-
tions,
(4) each branch is marked either "equivalence-preserved" or "equivaJence-not-
preserved",
(5) each node is marked "false" or not marked.
The equiformalization algorithm proceeds as follows. The algorithm consists of a
big loop and twelve transformations. Here, Co denotes an extended definite clause to
prove under a program P, where all the predicates referred to in Co are defined. And
in the following, the description is for a non-deterrnin~sticalty obtained tree.
A l g o r i t h m 3.2
1. Let To be a tree with only a root node Co. i ~- O.
2. If Ti satisfies a proof-condition, then ~he algorithm terminates: the answer is posi-
tive. If Ti satisfies a dlsproof-condition, then the algorithm terminates: the answer
is negative. If Ti satisfies a failure-condition, then the algorithm fails (only one
non-deterministically selected branch fails).
3. Select non-deterministically a non-empty leaf not marked false C in Ti and also
non-deterministicadly one of the procedures in Transformations and apply k to Ti.
Let the resultant tree be Ti+l.
4. Set i +- i + 1 and go back to step 2.
Transformation procedures are applied to some non-empty leaf of the tree. The
basic idea is to transform a conjunct or a clause on the leaf by "equivalence-preserved"
transformations to obtain its descendants, and to repeat the process until all the leaves
become trivially valid formulae. Of course, the validity of a root node is equivalent to
 15

t h a t of the leaves, a n d t h a t is preserved by transformations. But in fact we use some

"equivalence-not-preserved" transformations, hoping t h a t they will eventually assist us
to find good lemmas. O u r claim is t h a t even in the latter case, if all the leaves become
valid, t h e n under some condition the root node is valid.
Transformations.
T h e t r a n s f o r m a t i o n procedures~ when a tree T and a leaf C : 1~ ~ & are given,
o u t p u t a resultant tree T t. T h e y are one of the following (ones m a r k e d with * are
equivalence-preserved; the others are eqnivalence-not-preserved, a n d a variant is so cho-
sen t h a t its variables do not a p p e a r so far in T):
• (1) If the leaf is f a l s e ~-- true then it is m a r k e d false. If the leaf is F e-- f a l s e or
true *-- A t h e n T ' is T with an e m p t y leaf added as a child of C.
• (2) Simplification: If F is a conjunct and there exists a substitution cr = { X k / t k } where
• Xk is an internal variable of F a n d tk is a t e r m whose variables are of A, and
• for a n y predicate P in F there is a predicate Pv, in A,
then, T ' is T with an e m p t y leaf added as a child of F *- A.
(3) V-deletion: Select non-determlnistically conjuncts in F and delete t h e m f r o m F
forming r l . If all the conjuncts in F are to be deleted, F1 = false. T' is T with
F1 ~-- A added as a child of C.
• (4) Unfold-right: Select non-deterministicaUy a predicate H in A. Assume without
loss of generality t h a t A - A t , H . For every clause in P d o , n being initialized to
0:
if H is unifiable with H ' of a variant H r ~-- II of the clause with m.g.u, g ( t h a t is,
H 0 = HI0) then
set n *- n + 1, set Fn to F0 and set An to A~O, IIO
end-of-do.
If n = 0 t h e n set T ' to T with F ~- f a l s e added as a child of C. Otherwise set T ~
to T with F1 +- A 1 , . . . ,F,, ~ An added as the children of C.
• (5) Unfold-left: Select non-deterministically a conjunct F1 in F and a predicate H in
F1. Assume without loss of generality t h a t Fa -- F~, H , and F - F' V F1.
For every clause in P d o , n being initialized to 0:
if H is unifiable with H ' of a variant H ' *-- II of the clause with m.g.u. 0 ( t h a t is,
HO = H'O) t h e n
(a) set new to = = (consequently newr -
newtS,Ha) where ~ = { X 1 / X ~ , . . . , X d X ~ } , Xj's are the universally
quantified variables in old F1 to be replaced by 0, and X~'s are newly
introduced variables which have never a p p e a r e d in T,
(b) If there is no X j such t h a t it does not a p p e a r in A and is to be replaced
with a non-variable t e r m by 0, then:
set n ~-- n + 1 and set F~,n to F~O',IIO', where 0' is 0 except Xj's a r e
replaced with X ~ s
end-of-do.
If F - - F1 and n = 0 then set T ' to T with f a l s e ~- A added as a child of C.
Otherwise, set T ' to T with F' V F1,1 V . . . V Fl,~ ~-- A added as a child of C.
(6) Cut-right: If there exists an extended definite clause A1 V . . . v An *- II which is
a variant of a clause on a node N other t h a n C in T, and a substitution 6 which
satisfies
 ]6

• A ~- A 1 , A 2 and rio ~ A 1
• for a n y internal variable X in l'i, XO is a variable and does not occur other
t h a n in rio
• ~ replaces different internal variables in II with different internal variables in
A
t h e n let T t a tree obtained f r o m T by adding the children F +- AiO, A2 for all i to
the leaf C. T h e node N is called a reference node, and the node C a cut node•
(7) Cut-left: If there exists an extended definite clause A +-- 1I which is a variant of a
clause on a node N other t h a n C in T, and a substitution 6 which satisfies
• F --__F I , F 2 VF3 where F1 and F2 are conjuncts, and A0 -- F1
• for any internal variable X in A, X 0 is a variable and does not occur other
t h a n in A0
. 0 replaces different internal variables in A with different internal variables in
F1 and F2
t h e n let T ' be a tree obtained f r o m T by adding a child II6,F2 V F~ ~ A to the
leaf C. A is called a cut formula, and 0 is called a cut substitution. Also the node
N is called a reference node, and the node C a cut node.
• (8) Replace-right: H it is already proved t h a t A1 V - . - V An ~ II, II +- At, . . . , and
II ~-- An hold where A1, • " , An, and II are conjuncts a n d the variables in t h e m do
not a p p e a r in T, a n d there exists a substitution 0 which satisfies
A -- At, A2 and II0 _= A1
• for any internal variable X in I I , X 0 is a variable and does not occur other
t h a n in II0
8 replaces different internal variables in ri with different internal variables in
A
then let T ¢ be a tree obtained f r o m T by adding children F +- Ai0, A2 for all i to
the leaf C.
• (9) Replace-left: If it is already proved that A +- ri and II +- A hold where A and II
are conjuncts and the variables in A a n d II do not a p p e a r in T and there exists a
substitution 8 which satisfy
• F =_ F1, F2 V F3 where F1 and r2 are conjuncts, and A0 - F1
• for a n y internal variable X in A, X 0 is a variable and does not occur other
t h a n in AO
- 0 replaces different internal variables in A with different internal variables in
r l a n d F2
then let T ' be a tree obtained f r o m T by adding a child ri0,F2 V F3 +-- A to the
leaf C.
(10) Partial instantiation: If there exists a substltution cr {Xk/tk} where
=

• Xk is an internal variable of I?, and tk is a t e r m whose variables are of A,

then, T ' is T with a leaf ]P~r ~- A added as a child of C.
(11) Delete:
• (11-1) I f A = A I , A 2 and
(a) the variables appearing in A t never a p p e a r other t h a n in Ax, and
( b ) A1 +- t r u e is provable, where all variables are existentially quantified,
then, T ~ is T with a leaf r +- A= added as a child of C.
• (11-2) U A -= A~, Z~ satisfies the above (a) and
 17

(b) false ~ A1 is provable, where all variables are existentially quantified,

then, T' is T with a leaf r ~ false added as a child of C.
*(11-3) If r - r1,P2 v F3 where rl and F2 are conjuncts and
(a) the variables appearing in F1 never appear other t h a n in F1, and
(b) r l e- true is provable, where all variables are existentially quantified,
then, T' is T with a leaf F2 v F3 ~-- A added as a child of C.
*(11-4) i f r = r1,F2 V F3 where r l and F2 are conjuncts and satisfies the above (a),
and
(b') false e-- Ft is provable, where all variables are existentially quantified,
then, T' is T with a leaf F3 +-- A added as a child of C.
*(11-5) If r = F1,F2 V ['3 where F1 and F2 are conjuncts, and
(c) A - A1,A2 and F1 = A1
then, T' is T with a leaf r2 v F3 ~ A added as a child of C.
(11-6) If F -= r x , x , r l , 2 v . . . v Fn,I,Fn,2 where Fi,1 and ri,2 are conjuncts for all i,
A -- A l, A 2, and Fi,~ =- A1 for all i, then T' is T with a leaf F1,2 V . . . V Fn,2
A2 added as a child of C.
(12) Equality-elimination: Suppose an equality X = t is included in some conjuncts in
F or A (where X is a variable and t a term). T' is T with a C on which substitution
i X ~ t } is executed is added as a child of original C, where the conjunct is deleted
from the modified C if X is an internal variable and does not appear in the conjunct
after the substitution.
Note that a cut procedure is quite similar to a fold transformation [13]. The
difference is t h a t the cut replaces an extended definite clause with a prospectively non-
equivalent predicate (or another extended definite clause), whereas the fold replaces it
with its equivalent predicate.
Also note that although from the above description it seems that there is so much
non-determinism t h a t the algorithm generates many lemmas, some restriction of the
application of 'unfold' reduces the lemmas generated to no more t h a n those generated
by the Knuth-Bendix Mgorithm (if the problem is an equality problem). The details are
not discussed here due to the limitation of space.

4. C o n d i t i o n s t o s u p p o r t a p r o o f , a d i s p r o o f or f a i l u r e
In this section, we present conditions for the transformation tree generated by the
equiformalization algorithm to support a proof or a disproof, or to suggest t h a t some
transformation applied is inappropriate. Justification for them is described in the next
section. Some terminology referred in the conditions is defined afterwards.
C o n d i t i o n 4.1 ( c o n d i t i o n t o s u p p o r t a p r o o f )
(1) The leaves are a11 empty and are not marked fa/se.
(2) Any elementary directed closed path in a dependence graph of a transformation
tree is a contraction path.

Condition 4.2 (condition to s u p p o r t a disproof)

(1) There is a node marked false.
(2) A11 the transformations done between the root node and the node marked false are
marked equivalence-preserved.
 18

C o n d i t i o n 4.3 ( c o n d i t i o n t o s u g g e s t a f a i l u r e )
(I) Condition 4.I (1) holds, but not Condition 4.1 (2).
(2) Condition 4.2 (1) holds, but not Condition 4.2 (2).
In Tamaki and Sato's Unfold/fold transformation [13], which, in a sense, is included
in our approach, replacement of equivalent conjuncts is limited, so t h a t the replacer must
be "smaller" t h a n the replacee. In contrast~ in our approach, we do not place such a
condition on replacements, but on cuts (more precisely, when the cut forms a "loop"
with other branches). Clearly our restriction is weaker t h a n theirs. A similar approach
may be applied to fold/unfold transformations.
We prove the correctness of these conditions in the next section. T h e following are
definitions and theorems to be used in the proofs.
D e f i n i t i o n 4.4 If for a substitution O = {Xi/ti}~=l, Xi's do not a p p e a r in any ti's, a
constraint expression w(8) of 8 is: w(8) = X1 = tl, .• • , X n = tn.
D e f i n i t i o n 4.5 A substitution 0 is compatible with a first order formula II if[ M p
3(nO).
D e f i n i t i o n 4.6 A dependence graph of a transformation tree T is a directed graph
whose nodes and branches are those of T with a few more branches added, where,
(1) a direction of a branch originated in T is from a parent ~o a son.
(2) a new brazlch is added from a cut node to its reference node for any cut.
D e f i n i t i o n 4.7 An example o f an extended definite clause F ~- A is a substitution 8
such that
• All the universally quantified variables in the clause are instantiated to ground
terms by O, and AO is ground, and
• M e ~ AO
A counterexample is an example 8 but M e ~ B(F0).

T h e o r e m 4.8 ( p r o p a g a t i o n o f a c o u n t e r e x a m p l e ) / f there is a counterexampte on

a node N in a dependence graph, there is at least one successor node of N on which
there is a counterexample.
(proof) We have cases according to what transformation is applied to N. T h e numbers
used to identify cases in the following correspond to the ones in Transformations in
section 3. Let cr be a counterexample on N and a t be the one on a successor node,
which shall be the required•
(1) In the former case, a son is not generated. In the latter case, we do not have ~r.
(2) We do not have or.
(3),(5),(9),(10),(11),(12) Obviously or' = ~r except that the elements whose replaced
variable does not appear in the resultant node are deleted.
(4) Let us assume that a is compatible with w(0) for some 0 among n O's (for if such
a 0 does not exist, MR ~ Act). Let e ' be a ground substitution whose constraint
expression is equivalent to w(O)cr (the existence of such a o" is easily proved).
(6) Let ~? be a substitution which changes a clause on the reference node to the variant
used in the transformation. We have two cases:
 ]9

(a) MR ~ AiO~r e- IIOe for some i, i.e., MR ~ 3(Ai0cr). Let a" be such that
A i O ~ " is ground and Mp ~ AiOerer". Then let ~r' be a union of ~ and or" but
deleting elements whose replaced variables are not in Ai0, A 2. Since e" does
not replace internal variables in F~ a r is a counterexample.
(b) MR ~= AiOa *-- IIOa for any i, i.e., MR V= (A1 V . . . V An ~- II)0~. In other
words 7/&r is a counterexample on a reference node.
(7) Let r] be a substitution which changes a clause on the reference node to the variant
used in the transformation. We have two cases:
(a) MR ~= (rio, 1"2 v r 3 ~- A)tr. Obviously ~ is a counterexample.
(b) MR ~ (ri0, F2 V I'a ~- A)cr, i.e., MR ~ 3(riO~). It is easily shown that
M p ~: 3(A0~) from MR I= (ri0, F2 V Fs ~- A)~r and conditions on & We have
Mp ~: (A ~ II)&r. In other words, ~0a is a counterexample on a reference
node.
(8) By the equivalence of A1 V...VAn and 1I, for some i MR ~ 3((AiO, A2)tr) Therefore
we can find a ground substitution ~r" such that MR ~ (AI8, A2)er~r" and o-" does
not replace any internal variables in r . Clearly ~r~r" is a counterexample on P ~--
Ai0, A2.
T h e o r e m 4.9 ( p r o p a g a t i o n o f a n e x a m p l e ) If there is an example on a node N
in a dependence graph, there is at least one successor node of N on which there is an
example.
The proof is similar to that of Theorem 4.8, and is easily derived.
D e f i n i t i o n 4.10 A directed closed path in a dependence graph is caI1ed a contraction
path if[:
(1) for any node in the path there exists a well-founded order (~-) among the examples
of the extended de/~nite clauses on the node, and
(2) for any example a on any node (whose clause is r ¢- A ) in the path and any
example or' on the node obtained as a result of propagations of a along the path,
( r ~- A)~ ~ ( r ~- A)~' (strict i , equality)
D e f i n i t i o n 4.11 When a welI-founded order stated in Definition 4.10 is de~ned, we
accordingly define a well-founded order among examples such that:
ig

The well-founded order required by Definition 4.10 m a y be defined differently on

different nodes and on different paths. For example, we can define it for a predicate
p(s,t), or a conjunct A as:
(a) p(sl,tx) ~- p(s2,t2) iff s2 is smaller in structure t h a n Sl
(b) using the weight w defined by Tamaki and Sato [13],
>
The latter is quite orthodox but it may be difficult in a real application, whereas the
former is very easy to use but is limited. But if all the predicates are defined in P in the
way "primitive recursive functions" are defined, that is, the t r u t h value of a predicate
at s(t) is defined by the t r u t h value of the predicate at t, then we can always find an
argument which serves as an argument s in (a), which guarantees equivalence of (a) and
(b). This simplifies a great deal of work to be done to check the proof-condition.
 20

5. J u s t i f i c a t i o n o f t h e A l g o r i t h m a n d t h e C o n d i t i o n s
In this section we show justification of the basic algorithm and the conditions, that is,
if the proof-condition is satisfied then the target clause is modeled by MR, and if the
disproof-condition is satisfied then it is not.
T h e o r e m 5.1 Let N be a node and N 1 , . . . ,N~ be N ' s sons in a transformation tree.
(1) If a transformation applied to N is marked equivalence-preserved, then
M e ~ N iff M e ~ N 1 , . . . , and MR ~ N ,
(2) If it is not marked equivalence-preserved and is neither cu~-right nor cut-left, then
M p [= N if M p ~ N 1 , " ' , and M p ~ Nn
The proof is straightforward.

T h e o r e m 5.2 If a transformation tree satisfies the condition stated in Condition 4.t

(2) and there is a counterexample on a node in a transformation tree, then there is a
counterexampIe on a leaf which is a result of propagations of the former counterexample.
(proof) The proof is by induction on the construction of the tree.
Basis case: When the tree consists of a root node only, the statement is true.
Induction step:
(1) When the last transformation is other than cut, the statement is true by virtue of
Theorem 5.1 and the induction hypothesis.
(2) When the last transformation is either cut-right or cut-left, supposing that there
is a counterexample on the cut-node, we deduce that there is a counterexample on
the resultant leaf, which is a result of propagations. Let T be the transformation
tree. The proof is by reductio ad absurdum.
Let us suppose that we don't have counterexamples on the resultant node, whereas
we have at least one on the cut node. Let cr be the smallest among the counterex-
amples (use the order required by Definition 4.10). By the same argument as in
the proof of Theorem 4.8 (cases (6) and (7)), there should be a counterexample
gr on a reference node of the cut. By virtue of the induction hypothesis, for the
transformation tree T r which is the same as T except for the resultant leaf (i.e. the
tree just before the cut is applied), the statement holds. Therefore ~rt propagates
to some leaf of T ~. Since we are supposing that there is no counterexample on
the leaves of T, the propagated counterexample ~ inevitably comes to the cut
node. Since gtr is a result of propagations along a directed path, by the contraction
property of the path ¢r ~- e ' . This contradicts the minimallty of v'.
T h e o r e m 5.3 Let F +- A be a clause on a root.
(1) If a proof-condition is satisfied, then MR -~- r ~-- •.
(2) If a disproof-condition is satisfied, then M p ~= F +- A.
(3) If a failure-condition is satisfied, then either of the above cases m a y happen.
(proof)
(1) Let us suppose the contrary holds. By the hypothesis, there is a counterexample
on a root node. By virtue of Theorem 5.2, there is a counterexample on a leaf. But
by the proof-condition all the leaves (or more precisely all the nodes one branch up
from the leaves) are obviously true (i.e., modeled by Mp). This is a contradiction.
(2)(3) Obvious by Theorem 5.1.
 2]

6. D i s c u s s i o n
Among interesting topics which we have not discussed in this paper are:
(1) how to reduce the degree of non-determinism occurring in the algorithm
(2) the relation between the Mgorithm we proposed and the inductiouless-induction
method adopting the Knuth-Bendix completion algorithm.
(3) what procedures are to be added to augment capability of the algorithm
Concerning the first item, we are currently investigating two ways. One is to
introduce an order between conjuncts, and allow on a transformation tree only clauses
reduced according to this order. The other is to implement a strategy to find the most
appropriate action to take, which, observing that unless cut or replacement is used non-
trivial simplification or delete can not be applied, incorporates a way to find appropriate
unfold sequences and appropriate lemma.
For the second item, we have three things worth mentioning.
• To simulate the Knuth-Bendix algorithm, we have to:
(i) add some procedures to imdtate the whole capability of superpositions
(ii) modify procedures to use a simplification order
(iii) make some modifications to utilize the functionality of the predicates
• Even without those modifications, the proposed algorithm has capabilities almost
matching the Knuth-Bendix algorithm. Note that "delete" procedure is a powerful
tool, which has no counterpart in the Knuth-Bendix algorithm.
• The proposed method is advantageous on the following points
(i) it proves "implication"
(ii) it knows which clauses must be proved, and which clauses are just prospective
lemma that do not need to be proved.
(iii) it does not rely on a structure-induced order of terms. In the Knuth-Bendix
algorithm it is almost always necessary to make the equality relation directed.
For the last item, we are studying two approaches: procedures to utilize nice prop-
erties of system predicates; and restrictions on transformations (1), (3) and (10), so that
equivalence can be proved by one root node.

7. A c k n o w l e d g e m e n t
We would like to express our thanks to Mr. Yutaka Takuma, general manager of Naka
Works, Hitachi Ltd.; to Dr. Eiichi Maruyama, general manager of the Advanced lZe-
search Laboratory, Hitachi Ltd.; and to Mr. Masahiro Mori, manager of the System and
Software Design Department of Naka Works, Hitachi Ltd. for their giving us the chance
to conduct this research. We also thank Mrs. Mariko Sakurai for her encouragement
and fine typing.

References
[1] Apt, K.R. and van Emden, M.H.: Contribution to the theory of logic programming,
J. A C M 29 (1982) 841-862
[2] Boyer, R.S. and Moore, J.S.: Pro-ring Theorems About LISP Functions, J. ACIVI
22 (1975) 129-144
[3] van Emden, M.tt. and Kowalski, R.A.: The semantics of predicate logic as a pro-
gramming language, J. A C M 23 (1976) 733-742
 22

[4] Hsiang, J.: Refutational Theorem Proving using Term-Rewriting System, Ar~it~ciM
Intelligence 25 (1985) 255-300
[5] Hsiang, J. and Srivas, M.: Automatic Inductive Theorem Proving using Prolog,
Theoretical Computer Science 54 (1987) 3-28
[6] IIuet, G. and IIullot, J-M.: Proofs by Induction in Equational Theories with Con-
structors, Comput. Syst. Sci. 25 (1982) 239-266
[7] Kanamori, T. and Fujita, If.: Formulation of induction formulas in verification of
Prolog programs, ICOT TechnicaI Report TR-094 (1984)
[8] Kapur, D. and Musser, D.R.: Proof by Consistency, Artit~ciM Intelligence 31 (1987)
125-157
[9] Lassez, J.-L. and Maher, M.J.: Closures and fairness in the semantics of program-
ruing logic, Theoretical Computer Science 29 (1984) 167-184
[10] Lloyd, J.W.: Foundations of Logic Programming, Spfinger-Verlag (1984)
[11] Paul, E.: On solving the Equality Problem in Theories Defined by Horn Clauses,
Theoretical Computer Science 44 (1986) 127-153
[12] Sakurai, A. and Motoda, It.: Inductionless Induction Method in Prolog, Proc. 4th
Conference Proceedings of Japan Society of Software Science and Technology (1987)
231-234
[13] Tamald, If. and Sato, T.: Unfold/fold transformation of logic programs, Proc. 2nd
International Logic Programming Conference (1984) 127-138

A p p e n d i x A. E x a m p l e s
In this section some examples of apphcation of the proposed algorithm are presented.
The program under consideration is as follows.
Note that in definite clauses on transformation trees, the variables appearing on
both sides are considered universally quantified~ and those appearing on only one side
are considered existentially quantified.

~ev([],[]). (pl)
rev([AlX], Z) +- rev( X, Y), fins(A, Y, Z). (p2)
Iins(A, [], [A]). (p3)
lins(g, [BIY], [BIZ]) +- lins(A, Y, Z). (p4)
mere(A, [A[X]). (pS)
mere(A, [BtX]) +- mere(A, X ). (p6)
p~m([], []). (pT)
perm([AIX], Z) ~- perm( X, Y), ins(A, V, Z ). (p8)
ins(A, X, [AIX]). (p9)
in4A, [BIY], [BIZ]) ~- ins(A, Y, g). (plO)

A.1 R e v - r e v P r o b l e m
The first example is the "rev-rev" problem [2] stated in Prolog [7]. The problem is to
show the validity of rev(Z, X) +- rev(X, Z). The stem of a complete tree resulted in a
proof tree is created as follows.
 23

,ev(z, x) ~- ,ev(x, z). (rl)

~L unfold-right, cases for (pl) omitted
~ev(z, [AIX]) ~- ~e~(X, Y), tin,(a, Y, Z). (r2)
~L cut-right with (rl).
rev(Z, [AIX]) e- rev(Y, X), lins(A, Y, Z). (r3)
~- unfold-right, cases for (pl) omitted
rev( Z, [AIX]) ~ rev( Y, U), tins(B, U, X), tins(A, [BIY], Z ). (rd)
J~ unfold-right.
rev([B]Z], [AIX]) +-- rev(Y, U), tins(B, U, X), tins(A, ]I, Z). (r5)
unfold-left.
rev(Z, V), tins(B, V, [AIX]) *- rev(Y, U), tins(B, U, X), tins(A, Y, Z). (r6)
,~ unfold-left.
rev(Z, [AIV]) , lins(B, V, X) *-- rev(Y, U), tins(B, U, X), lins(A, Y, Z). (r7)
cut-left with (r3).
rev(Y', V), tins(A, Y', Z), tins(B, V, X)
rev(Y, U), tins(B, U, X), tins(A, Y, Z). (r8)
~ Y'*-Y, V~-U.
[]
Before going to the next example, we show two (stems of) different transformation
trees based on two different formalizations of the "rev-rev" problem. The first one is as
follows:
X = Z ~ rev(X, Y), rev(Y, Z). (sl)
,~- unfold-right, cases for (pl) omitted
[AIX ] = Z *-- rev(X, U), tins(A, U, Y), rev(Y, Z). (s2)
J~ unfold-left.
X = Z', [AIZ' ] = Z ~ ~'ev(X, U), tins(A, U, Y), rev(Y, Z). (s3)
,~ cut-left with (sl)
rev(X, Y'), rev(Y', Z'), [A IZ'] = Z *- rev(X, U), tins(A, U, Y), rev(Y, Z). (s4)
y1 *- U and delete
rev(U, Z'), [A]Z'] = Z ~ tins(A, U, Y), rev(Y, Z). (s5)
unfold-right, cases for (p3) omitted
rev([BlV], Z'), [A[Z'] = Z ,-- lins(A, U, Y), rev([BlY], Z). (s6)
unfold-right.
rev([BIU], Z'), [AtZ' ] = Z ~- tins(A, U, Y), rev(Y, W), tins(B, W, Z). (s7)
J~ unfold-left.
rev(U, V), gins(B, V, Z'), [AIZ'] = Z
,-- tlns(A, U, Y), rev(Y, W), tlns(B, W, Z). (sS)
cut-right with (s5).
~v(v, v), tins(B, V, Z'), iAJZ'] = Z
rev(U, W'), [AfW'] = W, lins(B, W, Z). (s9)
J~ equality elimination.
rev(U, V), Iins(B, V, Z'), [atz'] = z ~ ~ev(u, w'), tlns(B, [AJW'], Z). (sl0)
unfold-right.
,'ev(U,V),tins(B,V,Z'),[AlZ'] = [alz] ~- ~ev(U,W'),Zins(B,W',Z). (s11)
Z j *-- Z, V ~-- W / etc.
[]
 24

We now show the second formalization of the problem. In the following, a predicate
list(X) is defined by list([]), list([AlX]) ~- list(X).
rev(X, Y), rev(Y, X) ~- list(X). (tl)
J~ unfold-right,
rev([AiX], Y), rev(Y, [AIX]) ~ list(X). (t2)
unfold-left.
rev(X, Y'), lins(A, Y', Y), rev(Y, [AIX]) ~- list(X). (t3)
cut-right with (tl)
rev(X,Y'),lins(A,Y',Y),rev(Y,[AIX]) ~-- rev(X,V),rev(V,X). (t4)
Y~ ~ U and delete
Iins(A, U, Y), rev(Y, [AIX]) e-- rev(V, X). (t5)
J~ unfold-right, cases for (pl) omitted
fins(A, [B[U], Y), rev(r, [AIX]) ~- rev(V, Y), Iins(B, V, X). (t6)
J~ unfold-left.
Iins(A, U, Y), rev([BlY], [AIX]) ~ rev(U, V), fins(B, V, X). (iT)
unfold-left.
lins(A,U,Y),rev(Y, W),tins(B, W,[A]X]) e- rev(U, Y),tins(B, V,X). (t8)
unfold-left and v-delete.
fins(A, U, Y), rev(Y, [AIW]),fins(B, W, X) ~- rev(V, Y), fins(B, V, X). (t9)
cut-right with (tS),
Iins( A, U, Y), rev(Y, [AIW]) , fins(B, W, X)
lins(A, U, Y'), rev(Y', [AIV]), fins(B, V, X). (tl0)

[]

A.2 L a s t - i s - a - m e m b e r P r o b l e m
The "last-is-a-member" is an example of a problem where we have to use many heuristics
for without knowing the proposed algorithm [7]. The problem is to show the validity of
VA, X, Z mem(A, X) ~- rev(X, [A]Z]).
One transformation tree is:
mere(A, X) ~- rev( X, [AIZ]). (ml)
J~ unfold-right.
mem(A, [BIX]) e-- rev(X, U), lins(B, U, [AIZ]). (m2)
,~, unfold-right, cases for (p3) omitted
mere(A, [B IX]) e-- rev( X, [AIU]), lins( B , U, Z ). (m3)
J~ unfold-left.
mere(A, X) ~ rev(X, [AIU]), fins(B, U, Z). (m4)
cut-left with (ml).
rev( X, [AIZ']) ~-- rev( X, [AIUI), lin~( B, u, z). (mS)
J~ Z ' ~ U .
[]
 25

A.3 P e r m - p e r m Problem
Show that if a list Z is a permutation of a permutation of a list X, then Z is a per-
mutation of X. The problem seems very trivial if we understand the abstract meaning
of 'permutation', but if we refer only to the formal definition of the predicate "perm",
then it is quite complicated. Readers who have the opposite opinion are challenged to
solve it without using the proposed algorithm.

perm( X, Z) ~ perm( X, Y), perm(Y, Z). (pro1)

unfold-right, cases only ~or (p8)
perm([AIX], Z) ~ perm(X, U), ins(A, U, Y), perm(Y, Z). (pro2)
~- unfold-left.
perm(X, V), ins(A, V, Z) ~ perm(X, V), ins(A, U, Y), perm(Y, Z). (pm3)
cut-left with (pml),
pe m(x, w), perm(W, V), ins(A, V, Z) +
perm(X, U), ins(A, U, Y), perm(Y, Z). (pro4)
"0" W ~- U and delete
perm( U, Y), ins(A, V, Z) ~ ins(A, U,Y ), perm( Y, Z ). (pro5)
~, unfold-right, cases only for (pl0)
perm([B IV], Y), ins(A, t/, Z) *- ins(A, U, Y), perm([B]Y], Z). (pm6)
unfold-right.
perm([B 1U],V), ins(A, V, Z) +- ins(A, U,Y), perm(Y, W), ins(B, W, Z). (pm7)
~- unfold-left
perm(V, V'), ins(B, V', Y), ins(A, V, Z)
ins(A, U, Y), perm(Y, W), ins(B, W, Z). (pmS)
cut-right with (proS)
pe m(v, v'), ins(B, Y', Y), ins(A, V, Z) +
perm(U, Y), ins(A, Y, W), ins(B, W, Z). (pm9)
J~ V ~ *-- Y and delete
ins(B,Y,V),ins(A,V,Z) .- ins(A,Y,W),ins(B,W,Z). (pmlO)
unfold-right, cases only for (pl0)
ins(B, [ClY], V), ins(A, V, Z) ~ ins(A, Y, W), ins(B, [CIW], Z). (pmll)
,~ unfold-right, cases only for (pl0)
ins(B, [CIY], V), ins(A, V, [CIn]) +- ins(A, Y, W), ins(B, W, Z). (pml2)
~- unfold-left and V-delete.
ins(B, Y, V), ins(A, [CtV], [CIZ]) +- ins(A, Y, W), ins(B, W, Z). (pml3)
unfold-left and V-delete.
ins(B,Y,Y),ins(A,V,Z) +-ins(A,Y,W),ins(B,W,Z). (pm14)
~L cut-right with (pmlO)
ins(B, Y, V), ins(A, V, Z) +--ins(B, Y, W), ins(A, W, Z). (pml5)
V+-W
[]
 26

Readers may be interested in what is happening in side branches. One branch stemmed
from ( p m l l ) is as follows.
ins(B, [CIY], V), ins(A, V, Z) ~ ins(A, IT,W), ins(B, [CIW], Z). (pmll)
unfold-right, cases only for (p9)
ins(B, [CIY], V), ins(A, V, [B, ClW]) ~- ins(A, Y, W). (pm12')
unfold-left and V-delete,
ins(B, [ClY], [Big]), ins(A, V, [ClW]) ~- ins(A, Y, W). (pro13')
,~ unfold-left and V-delete.
ins(B, [CtY], [B, CIV]), ins(A, V, W) ~ ins(A, Y, W). (pro14 ~)
~L unfold-left and V-delete.
ins(A, IT,W) +-- ins(A, Y, W). (pro15')

[]

A p p e n d i x B. I n e q u a l i t y
This example is intended to show some key ideas of the proposed algorithm. In this
example, an inequality is proved using an algorithm based on the same principle as ours
for Prolog. Similar applications may be found in other fields.
Suppose you know only three things: arithmetic of non-negative integers, inequality
relations between integers, and the fact that you may deduce an identical term from
both sides of an inequality.
How do you prove (n + 1) 2 > 4n when n > 1?

((n + 1) + 1)2 ~ 4(n + 1) "~\ [(1+ 1)2 ~_ 4 . 1 I

(n+ly +2(n+1)+1 > 4~+4

~ .......
', x
~n + 1)~ + 2(n + 1) + 1 >__(n + 1y + 4
_ _ [ _ _
[2(n+ 1)+ 1 >_ 4J

2((n + 1) + 1) + 1 :> 4] ", [2(1 + 1) + 1 > 4 I

2(n+1
'

1+2>4
"'--"

_
i'.'I

14 + 2 >41
±
 Pseudo Extension in Default Reasoning

and

Belief Revision by Model Inference

Hiroko Yuasa* a n d Setsuo Arikawa

Department of Information Systems,

Kynshu University 39, Kasuga 816, Japan

Mailing address: S. Arikawa, Research Institute of Fundamental Information Science,

Kyushu University 33, Fukuoka 812, Japan

Abstract

This paper is concerned with extension and belief revision in l~eiter's default reasoning.
First we introduce a new notion of pseudo extension which is successively constructed from a
given default theory. We show that the set of pseudo extensions contains the set of extensions,
so that we can get some reasonable candidates of extension in a constructive way. This should
be a first step towards computerization of the default reasoning. We also discuss the belief
revision problem from the viewpoint of Shapiro's inductive inference of models.

1. Introduction

Reasoning we use in our daily problem solving is not always from complete information unlike mathemat-
ical reasoning. One of the main goals of artificiM intelligence research is to computerize our reasoning.
Hence much attention has been paid to such reasoning and formulations have been proposed by many
authors, l{eiter~s default reasoning is one such formulation. He used the notion of default and formulated
incomplete knowledge about a world as a pair : a set of true but incomplete knowledge and a set of
special inference rules called defaults. He enlarged the incomplete knowledge to describe the world as
completely as possible. The enlarged knowledge is ca~ed an extension, which is the set of knowledge
that seems true as a result of reasoning from incomplete knowledge and is a possible world at that time.
Thus the extension is not definite knowledge but just a belief about the world, and hence it needs to be
modified when new facts are found.

In the Reiter's default theory we can reason from incomplete knowledge in a very natural way.
However, the extension can not successively be constructed, because it is defined by using itself in
applying the defaults.

In the present paper, we first try to solve this difficulty. We define a successively constructable set

* Presently at Hitachi Central Research Laboratory, Hitachi Ltd.

 28

of knowledge, which we call a pseudo extension, and discuss the relationship between Reiter's extensions

and ours. We also show that our extension is a natural generalization of l~eiter's.

Unlike the reasoning in classical logic, the reasoning from incomplete knowledge is necessarily non-

monotonic. Hence we sometimes must modify the belief, i.e. the knowledge derived by such reasoning.

In the present p a p e r we take this problem of belief revision as that of extension modification. Then we
can see that the process of belief revision is very similar to Shapiro's model inference. We also discuss

the belief revision from this viewpoint.

2. Reiter's Default Reasoning

We start with recalling the definitions and important results on default reasoning according to [1, 2].

Let L be a first order language. Then a default is a rule of the form

a(x): Mtl(x),... ,Mtm(x)

,4~)
where ,~(~), til(~)," "", tm(X),','(~) are wall-formed formulas (wff, for short) in L, and • denotes a sequence
of free variables. We call a(x) a prerequisite, t l ( x ) , . . . , tim(x) justifications, and ~(x) a consequence.
T h e M is a special symbol to be read as "it is consistent to assume". Hence the default means that

if it is consistent to assume ~(x),... ,tim(X) we may infer a ( z ) ~ w(x). Note that "it is consistent to
assume t i ( x ) " means t h a t there is not sufficient information to conclude -~fli(z). T h e prerequisite a(z)
may possibly be empty, that is,

: Mill(z) .... , Mtm(x)

is also a default to mean that if it is consistent to assume ~ ( x ) , . . . , t , ~ ( x ) then w(x) can be concluded.
If a ( x ) , t l ( X ) , . . . , tim(Z), w(x) have no free variables, the rule is called a closed default and denoted

a : Mfll,o.-,Mflm
O3

A default of the form

a ( x ) : Mti(z)
ti(~)
is called a normal default.
Now let D be a set of defaults, and W be a set of closed wffs. Then A = (D, W ) is called a default
theory. A default theory A is closed if every default in D is closed, and A is normal if every default in

D is normal.
For a set D of default rules, we define Cons(D) by

Thus Cons(D) is the set of all consequences of defaults in D.

 29

For a set S of closed wffs and a closed wffw, we write S ~- ,, to denote that w is derived from S, and

S ~/w to denote that w is not derived from S. We also define a set Th(S) by

Th(S) = {~ I~ is a closed ~ and S ~ ~}.

Let A _-- (D,I4/) be a closed default theory, S C L be a set of closed wffs, and r ( S ) be the smallest
set satisfying the following three properties:
(1) W C r(s).

(2) T h ( r ( s ) ) = r ( s ) .
(3) If (~ : Mill, ... ,Mflm E D, a e F(S), -~j31. . . . . -"tim f[ S then w e r ( s ) .
02
A set E C L satisfying r(E) = E is an extension of A. An intuitive characterization of the extensions is
given by the following theorem :

T h e o r e m 1 (Reiter [1]) Let E C L be a set of closed wits, A = ( D , W ) be a closed default theory,

and E0, E l , . . . be the following sequence of sets of closed wits:
Eo=W

Ei+l = Th(Ei) U ~o(z) Ei F a ~

E ~/-,~1 . . . . , E V"~,~
Then E is an extension of A if and only if

E= UEi.
i=0

The expression of Ei+l in Theorem 1 is slightly different from the original one. Reiter [1] used

conditions a e E l , ~ f l l . . . . ,"tim • E instead of El ~- a , E ~/-~fll . . . . , E ~/ -~fl,~, respectively. These

expressions are essentially the same. In order to check a E E1,-~fll,... ,~flm C E, we need to check
whether they are derivable or not. So we have adopted our expression, which should be more intuitive.

For a closed default theory there may exist two or more extensions or none at all. The existence of

many extensions means the possibility of many worlds at that time. Selecting a correct extension and

constructing an extension are different things. The correctness will gradually become clear as new facts

about the world are known. We say that a dosed default theory has an inconsistent extension if one of
its extensions is the set of all closed wffs of L.

C o r o l l a r y 2 (Reiter [1]) A closed default theory (D, W) has an inconsistent extension if and only
if W is inconsistent.

C o r o l l a r y 3 (Reiter [1]) If a closed default theory A has an inconsistent extension then it is only
the extension of A.

Since a closed default theory with an inconsistent W has only inconsistent extensions, hereafter we
restrict our discussions to default theories with consistent W's, unless otherwise stated. Default theory
may not have extensions, but we know the tbHowing :
 30

T h e o r e m 4 (Reiter [1]) Every closed normal default theory has at least one extension.

3. P s e u d o Extensions and Their Properties

In defining extension E, l{eiter used E itself (Theorem 1). Hence it is a self-recursive definition. We can

use Theorem 1 just to confirm for a given set of closed wffs to be an extension, but can not use it to

construct extensions from a given closed default theory.

This section proposes a new kind of extensions which can successively be constructed from a closed

default theory.

Definition Let & = (D,W) be a closed default theory and let Fi (i >_ 0) be sets of closed wffs
defined as follows:
F0=W

F~+I = Th(Fi) U Cons(DF~) (i > 0),

where DFi is a subset of
I ~ e D,w ~ Cons(U~-=~DFk) ]
= a : M~l,...,Mflm F~ F a
OJ

F~ V-~I~,..., t~ V -~flm
such that DFi • 0 if the above set is not empty. Then we call

F=OFi
i=O

a pseudo extension of the theory A.

The condition Fi V -~/~,... ,Fi V/3~ in applying a default with M ~ , . . . , M f l ~ is diffrent from
tteiter's. Now we can successively construct the pseudo extensions. Moreover in constructing F~+I we

can use an appropriate subset of defaults instead of the whole. A relation between Reiter's extension and

ours is given as follows :

Theorem 5 Let E be the set of all extensions of a closed default theory A, and let ~# be the set of

all pseudo extensions of A. Then

Therefore every extension is also a pseudo extension.

Proof Let E = U~=I El be any extension, and Di be the set of defaults used in defining Ei. Then

we have
El = Th(Ei-1) U Cons(Di).

For a default
= O~ : M ~ I , . l. , M~e-~ E Di,
 31

we easily have
Ei~ a

Z V-~£,-.-,E V ~ m .
By putting
D~ = {~ t~ e Di, Cons(~) ¢ FI-1},

we construct sets of closed wtfs Fz, F2,... as follows :

Po= W

Yl = Th( Fo) U Cons( D~)

Fi = Th(Fi-1) U Con~(D~)

Then, since Ei = Fi for each i (> 0), we have

i=0 i=1

Now let us show that D~ satisfies the condition of DFI in the definition of pseudo extensions. First
assume t h a t for some j and for any k(:> j ) a default satisfies the condition but D~ = 0. T h a t is, assume
that the set
ED, w~t~ stuk= 1 k)
= a : Mfll,...,M!m Fi F- a

lPi V "ill,''" ,Yl V -lira

is not empty but D~ is empty. Then by definition of D~ and D~ - 9, we have

Dj-1 = Dj = Dj+z = ...

Hence we have

E j = Ej+I = Ej+~ . . . .

From the assumption there exists a default

= a : M/3z,..., Mflm
ED,

which satisfies
j--1 Dk)
w ~ Cons(Uk=l !

Fj ¥ -,!1,..., Fi V -,t,~.
By Ej = Fj, obviously
Ej~-a

Ej V-~11,..., Ej V ~im.
 32

Also by
¢~ oo

,~ ¢_co,~(U D¢) c_ Co~((_J D~),

i=1 i=1

there exists at least one k (1 < k < m) such that

Since E = ~ 0 E~, there exists an i such that E~ b ~Sk- This i must be i > j , because E~ ~ ~Sk for

i < j. Since Ei = E j for such i,j, we have

Also by Ej = Fj, we have

which contradicts

Therefore D~ satisfies the condition of D/q, and hence F = (J~0 F~ is a pseudo extension. Thus we have

E = U E~ = Fi=FE/~.
i=O i=O

H e n c e E C _ F . []

Theorem 6 In a closed normal default theory, any pseudo extension which is not an extension is

inconsistent.

Proof Let F be a set of closed wits such that F E f" and F ~ E. Let D ~ be the set of defaults

which is used in constructing F~ of F - (J~=0 Fi. Then by the assumption there exist an i and a default

~= a:M8
8

such that 6 e DF~, F ~- a, F~ ~ -~ fl, F b 8. By 6 E D, F ~- ~, F~ ~ -78, we have

8 E Fi+l C_F.

Hence F ~- "~8 and F [- 8, which means F is inconsistent.[::]

Remark Theorem 6 is not valid if the default theory is not normal. In fact, there exists a default
theory whose pseudo extension is neither an extension nor an inconsistent set. An example follows.
Let A = (D,W) be a default theory with

D={ :MAB ' :M'-Bc ' B^C: }

W=O.
Then F = Th({A, B, C}) is a pseudo extension of A, but neither an extension of A nor an inconsistent
set.
 33

From the discussions above we can conclude the following :

(1) Pseudo extensions can successively be constructed from a given default theory.

(2) Extensions are always pseudo extensions.

(3) In a closed normal default theory, every pseudo extension is an extension if it is not inconsistent.
(4) If W in A = (D, W) is inconsistent then A has a unique pseudo extension, which is inconsistent

and coincides with the extension. This result immediately follows from Corollaries 2, 3 and the definition.

4. E x t e n s i o n R e v i s i o n by M o d e l Inference

Default reasoning is non-monotonic in the sense that an increase in facts does not always mean an increase

in knowledge. Thus the extension is not a definite knowledge, but may need to be revised, because some

defaults used in constructing it become inapplicable when new facts are found. Reiter has defined the

belief as a subset of the extension and has discussed revising it. In this section we define the belief as the

whole set of extensions, and the belief revision as constructing an extension as similar as possible to the

world to be described by the default theory. Thus we identify belief revision with extension revision. Let
us consider the world for the theory to describe as an unknown mode] in a first order language. Then

we can formalize the problem of extension revision in the framework of Shapiro's model inference [3, 4].
Hereafter we assume readers are familiar with the model inference. Our problem can be expressed as
follows:

Given a first order language L, an observational language

Lo = {ground unit clauses of L},

a hypothesis language Lh = L, and an oracle for an h-easy unknown model K which satisfies
the closed default theory A = (D, W), to find a finite Lo-complete axiomatization of K.

We say that the model K satisfies a closed default theory A = (D, W) if it satisfies the following
three conditions:

(1) K k W
which means K satisfies W.
(2) For a default
a : M i l l , . . . , M~.~
~d
ED

if K [=- a and K ~ - ~ f l l , . - - , K ~&-~fim then K ~ w.

(3) Let D' be a subset of D such that I~VU Cons(D ~) is consistent. Then

W W Cons(D') t;t u =:~ M ~ u

for any clause u which is not a ground unit.

 34

The condition above restricts the unknown model K in Shapiro's model inference.
A closed default theory ~ = ( D , W ) with finite D and W is called a finite closed default theory. For

such a theory the following theorem holds :

Theorem 7 Let A = ( D , W ) be a finite closed default theory and E be an extension of A. Then

there exists a finite Lo-complete axiomatization of E.

Proof It suffices to show E ~ T and T ]= L~ for some finite subset T of L. By Theorem 1 any

extension E can be written as E = [.J~0 Ei.

Let Di (i >_ 1) be the set of defaults which are used in defining El. Then Di _ D and

Eo= W

Ei = Th(Ei-1) UCons(Di) (i > 1).

Since D is finite, there exists an N such that D= = Dn-1 for any n > N. Hence we have

Eo C E1 C ... C EN-1 C EN -= EN+I . . . .

and
E = ~J Ei = E N .
i=o
We also have
EIv = Th(Eiv-1)

= Th(Th(EN_2)UCons(DN-1))

= Th(Th(Th(EN-3) o Cons(DN_2)) U Cons(DN_l))

= Th(Th(EN_3) U Cons(Dg_2) U Con~(DN_a))

= Th(Th(Th(Eo) W Cons(D1)) LJ(.JiN=-21Cons(Di))

= T h ( T h ( W ) u U~=I
N-1 Con~(O~))

= T ~ ( w u U ~ ' Con~(DO).
Then
N-1
T = WU U Cons(Di)
i=l
is finite because so are W , D , and obviously E I= T and T t- L E. Therefore there exists a finite Lo-

complete azdomatization of E. D

Theorem 8 Let K be an h-easy unknown model which satisfies a finite closed default theory
A = (D~ W). If Shapiro's enumerative model inference algorithm identifies K in the limit and produces

a guess T as the output, then Th(T) is an extension of a default theory A' = (D, W U W'), where W ~ is

a set of clauses.
 35

Proof Put

W ' = { u I T F u, W u Cons(D') Vu} u {fill < i < k},

where u is a ground unit clause and fi is defined by

(
f ; = ~ a~ ifV~=true
[ -~a{ if V} = false

when a fact < a~,V} >, a pair of ground atom and its t r u t h value, is given as input to the algorithm.
In order to show that Th(T) is an extension of A ' = ( D , W U W'), it suffices to show Th(T) =
F(Th(T)). From the definition of W ' and K ~ W, obviously T F W U W'. Hence we have
W U W ' C_Th(T)

Th(Th(T)) = Th(T).
Since K satisfies A, there exists a default

5= a:Mfll'""Mflm ED

such that K ~- a a n d K ~ - ~ 3 1 , . . . , K ~:-~flbimplyK ~ w . Namely f o r a ~ E D i f T h ( T ) F aand

Th(T) V ~31 . . . . , Th(T) 17/-~3m then w E Th(T). Hence Th(T) satisfies the three conditions on F(S).
Therefore from the minimality of F(S), we have

r(Th(T)) C__Th(T).

We now show the converse. Let v E Th(T). There are four cases to be considered.
Case I. v e W U W ' : v E r(Th(T)) by W u W ' c_ F(Th(T)).
Case 2. A default
5= a:Mfll,...,M/3,~ ED
v

exists and K ~- a, g V= -~fl,..., g ~/3,n : Clearly

Th(T) ~ a

Th( T) ~/ ~31, . . . , Th( T) ~/ ~flm

Th(T) C_F(Th(T)).
Hence F(Th(T)) F a and v E I?(Th(T)).
Case 3. W U W ' U Cons(D") F v for the set D" of all defaults

5- ~ : Mfll''"'M3m .E D
~d

such that K ~ o~,K ~ -~/~I,---,K ~ "3m ~¥e have

v e Th(W u w' u Co~(~"))

c Th(F(Th(T)))

= F(Th(T)).
 36

Case 4. v E Th(T) satisfies none of (1), (2), ( 3 ) : If v is not a ground unit then K ~ v since K
satisfies A. This contradicts that Lo has an evidence to show the falsity of T. If v is a ground unit then
obviously
v e W ' C_ r ( T h ( T ) ) , o

Theorem 7 asserts that an Lo-complete axiomatization exists for any extension of finite closed default
theory. Hence the extension appeaxs in the enumeration Th(T1),Th(T2) . . . . . where T1,T~,... is the
enumeration of at] subsets of Lb. On the other hand Theorem 8 asserts that if we restrict the objects
of the inference to the unknown models which are h-easy and satisfy finite closed default theories, then
the models identified by the enumerative model inference algorithm axe extensions of some closed default

theories. The A t in Theorem 8 is finite~ and hence it has at most finite number of extensions. Therefore
if for a finite closed default theory there exist two or more extensions in which the unknown model K
is included, the algorithm will find it in the search space of all extensions. Moreover if all the pseudo
extensions of a finite closed default theory can be enumerated, the algorithm will find an extension for
the unknown model by deleting guesses which are pseudo extensions but not extensions or contradict

newly given facts.

5. Discussions

There are some problems in Reiter's default theories. The extensions can not be constructed successively

and even their existence is not guaranteed with exception of the closed normal cases. Many efforts to
attack the problems have been made by many researchers, l~eiter and Criscuolo [5] have introduced
seminormal defaults. The seminormal default theories can describe some natural examples which the
normal default theories can not, but still the existence of extensions is not guaranteed. Lukaszewicz
[6] has studied a default theory of propositional logic and introduced another definition of extensions,
for which the existence is guaranteed. Murakami et al. [7] have constructed a set of knowledge called
a modified extensions, which is, in a sense, a generalization of IZeiter's extension, and guaranteed the
existence. [6, 7] have thus guaranteed the existence for any default theory, but still their extensions can
not be constructed successively from W. Murakami et al. [8] have dealt with this problem independently.
The)" have defined sets of knowledge called a nonrecursive extensions for closed normal default theories.

The nonrecursive extension can be uniquely determined.

In case the closed normal default theory A = ( D , W ) has an inconsistent extension, that is, W is
inconsistent, the nonrecursive extension is also inconsistent and equal to the extension of A. However,
the inconsistency of the nonrecursive extension does not mean that of W. In other words, if A has two
or more extensions~ the nonrecursive extension is inconsistent or equal to some of th9 extensions. If A
has just one extension, the nonrecursive extension coincides with it.
On the other hand, our pseudo extensions may include non-extensions but include all the extensions.
 37

The extensions should be candidates for possible worlds. In this regard, our approach is better than the
others.

We have introduced the notion of pseudo extensions which can be constructed successively and made
a step towards the computerization of default reasoning. However, it is still not computable, because
the condition Ei ~/ "~1,..- ,El ~ ~m for applying defaults is not even semi-computable. This aspect is
common to all other approaches. To solve the problem we need some concept like the h-easiness.

References

[1] Reiter, R. : A Logic for Default Reaso~dng, Artificial Intelligence, 13, 41-72 (1980).

[2] Shinmura, T. : On the Inconsistency in Non-monotonic Logic, Master's Thesis, Dept. of Information
Systems, Kyushu Univ. (1986) (in Japanese).

[3] Shapiro, E.Y. : Inductive Inference of Theories From Facts, TR-192, Dept. Compt. Sci., Yale Univ.
(1981).

[4] Ishizaka, tI. : Model Inference Incorporating Least Generalization, Master's Thesis, Dept. Informa-
tion Systems, Kyushu Univ. (1986) (in Japanese). To appear in J. In/. Processing Soc. Japan.

[5] Reiter, R., Criscuolo, G. : On Interanting Defaults, Proc. 8th IJCAI~ 270-276 (1983).

[6] Lukaszewicz, W. : Considerations on Default Logic, Proc. AAAI Non-monotonic Reasoning Work-
shop, 165-193 (1984).

[7] Murakami, K., Aibara, T., Shitanda, H. : Modified Extension and Its Properties in Default l~eason-
ing, Trans. Inf. Processing Soc. Japan, 12, 1280-1287 (1987) (in Japanese).

[8] Murakami, K. Aibara, T., Shitanda, it, : Nonrecursive Extension and Its Properties in Default
Reasoning, J. Japanese Soc. Artificial. Intelligence, 3~ 359-367 (1988) (in Japanese).

[9] Yuasa~ It.: Pseudo Extensions in Default l~easoning and Belief t~evision Using Model Inference,
Master's Thesis, Dept. Information Systems, Kyushu Univ. (1988) (in Japanese).
 An Approach to Nonmonotonic Inference Mechanism in
Production System KORE/IE

Toramatsu Shintani
IIAS-SIS FUJITSU LIMITED,
140 Miyamoto, Numazu-shi, Shizuoka 410-03, Japan
E-mail: tora%[email protected]

Abstract

In this paper, we propose a method for constructing a mechanism for

nonmonotonic inference in KORE/IE which is a Prolog based forward-
chaining production system. In the implementation, we utilize the form of
default reasoning proposed by Reiter, and introduce the framework of
TMS(Truth Maintenance System) for resolving contradictions. Relational
information such as data dependencies is independently managed by
KORE/EDEN, a network-based knowledge m a n a g e m e n t system, for
improving dependency-directed b a c k t r a c k i n g in TMS. Resolving
contradictions is achieved by undoing the inference. The mechanism for
resolving a contradiction can be realized by extending the inference steppers
in KORE/IE along the framework of TMS. As an example of the mechanism,
we show a maintenance mechanism for pairwise comparisons. Consistency of
the comparisons is effectively maintained by using the mechanism.

1. Introduction

KORE/IE(Shintani 1988) is a fast production system on Prolog. Its

inference mechanism implements a recognize-act cycle in a manner similar to
OPS5(Forgy 1981) whose type of reasoning is forward chaining.
Generally, inferences of production systems are monotonic. Namely,
consistency among facts in WM (working memory) is ignored, since WM is
only used as a global data base for rules. However, if we want to treat inexact
knowledge effectively, we need to create a mechanism for nonmonotonic
reasoning (Matsumoto 1987) that can maintain consistency. This mechanism
will be useful for improving the quality of knowledge representation for
heuristics (or expertise) and constructing rule-based applications.
This paper describes a method that supports nonmonotonic inferences in
KORE/IE. In the method, data dependencies are independently managed to
improve efficiency of the inferences. The data dependencies are supervised by
 39

KORE/EDEN (Shintani 1986,1986b) which provides functions for presenting

and managing many relationships between knowledge in a knowledge base.
This paper consists of five sections. In Section 2, we present a basic
architecture in KORE/IE for nonmonotonic inferences. In Section 3, details of
the mechanism for the inferences are presented. In Section 4, an application
implemented by utilizing the mechanism is described. In Section 5, some
concluding remarks are presented.

2. The basic architecture

Inexact knowledge is a useful part of flexible problem solving. It can be

classified into the following two types; (1)knowledge for choosing alternatives,
generally needed for problem solving such as diagnosing; (2)knowledge for
common sense information, needed for problem solving such as scheduling.
The knowledge of type (1) usually has numerical values which indicate
uncertainty. Consistency within the knowledge does not have to be
maintained. Knowledge of type (2) requires consistency maintenance. In
order to maintain the consistency, a problem solver needs to revise a part of
past problem solving. In KORE/IE, we realize a nonmonotonic inference
mechanism for managing the knowledge of the latter type. Its inference is
based on default reasoning(Reiter 1980) which provides frameworks for
managing the knowledge. We use the framework of TMS(Doyle 79) which is
used as a mechanism for dissolving contradictions between knowledge. In
order to effectively construct the mechanism of TMS, we utilize functions of
inference steppers of KORE/IE, and make a new mechanism for managing
default knowledge.

2.1 Utilizing inference steppers

In KORE/IE, when an inference is performed by executing a recognize-act

cycle, cancellation of the inference is achieved by putting back the cycle. The
recognize-act cycle can be illustrated as shown in Fig.l, which is a basic
control mechanism for an inference engine. The cycle is a sequence of
operations, which are (1)matching, (2)conflict resolution, and (3)an act
process. In order to put back the cycle effectively, we record changes to
WM(working memory) and the conflict set instead of recording every WM and
conflict set. The elements are called WM elements and instantiations,
respectively. Recording the changes is necessary for efficient memory use.
An instantiation is a consistent set of bindings for a rule. Putting back the
cycle by one step corresponds to changing the current state to a previous state.
The state is determined by specifying WM and the conflict set.
KORE/IE provides inference steppers for controlling the recognize-act
cycle, which are the "run" and "back" commands. The "run" command is used
 40
for performing inferences in forward chaining. In contrast, the "back
command" is used for undoing effects of the inferences (or rule firings) by
using the record of changes made while tuning to undo the changes, restoring
the WM and conflict set to a previous state. Each of the commands takes a
positive number argument which represents number of repetitions of the
cycle.

~ L~Confllctresolution,1

" "1 "1

Fig.1.Therecognize-actcycle
In order to record the changes, KORE/IE asserts a Prolog fact (or
assertion) for every forward-running cycle. We call the assertion back data.
The back data is expressed as follows;
back_data(Step, Flag, Instantiation, WM list)
where the first argument "Step" represents a step number of a recognize-act
cycle which corresponds to an inference step. The larger the step number, the
more recently the back data was asserted. The second argument "Flag"
represents a status which is used for nonmonotonic inferences. Details will be
discussed in Section 3.2. The third argument "Instantiation" represents an
instantiation which is selected in a conflict resolution process. The fourth
argument "WM_list" represents a list which keeps input-output information
of WM elements. By using a Prolog term, the input information which
indicates WM elements added to WM is described in the list as follows;
in_wm(WM_ELEMENTS)
where "WM_ELEMENTS" represents the WM elements. The output
information which indicates WM elements removed from WM can also be
described in the list as follows;
out wm(WM_ELEMENTS).
Fig.2 is given to clarify the description, which illustrates the internal
mechanisms of the run command and the back commands. In Fig.2, the
"run(5)" is an example of the run command which runs 5 steps of the
recognize-act cycle forward. The "back(2)" is an example of the back
command which undoes the effects of the 2 previous recognize-act cycles. By
using information of the back data in reverse order and restoring current WM
to a previous state, the inferences are undone. In the example, for undoing
the inferences, the back data for Step 5 is applied first, and then the back data
for Step 4 is applied. Keeping the back data is started by the command
 41

"ie_debug". The end of keeping the back data is specified by the command
"ie no debug". These commands are used for avoiding needless recording of
back data and for improving the efficiency of the inferences. The back data
are only needed for debugging a rule base and e n a b l i n g n o n m o n o t o n i c
inference.

ie debug

Inference step.~
Step 3 l---: - -~ back(2) t

Step 4
......~..~... • Undoing
Step 5 m
...... ,'"'L'""
" ie_no_debug

Fig.2. An internal mechanism of the run and back commands

2.2. R e p r e s e n t a t i o n for d e f a u l t k n o w l e d g e and c o n t r a d i c t i o n

Default reasoning is reasoning based on inexact knowledge. "Default

knowledge" is used as long as no contradictory evidence is present. In a
default rule(Reiter 1980), the knowledge can be represented by attaching a
modal operator M as follows;
~(x): M 1~(x)
7"(x)
where 13(x) corresponds to the default knowledge. This default rule is
interpreted as saying "For all individuals x, if a(x) is valid and if ~(x) is
consistent with our beliefs, then y(x) may be believed". If a contradiction is
detected by adding new knowledge to a knowledge base, an inference engine
for default reasoning needs to correct the default knowledge which causes the
contradiction and undo results of inferences which depend on the default
knowledge. Namely, a contradiction is due to the default knowledge. The
correction of the default knowledge is achieved by asserting its negation,
because the validity of the knowledge is assumed as long as the negation of the
knowledge is not proved.
In KORE/IE, a mechanism for resolving the contradiction is implemented
using a method for maintaining consistency adopted in TMS(Doyle 1979).
When a contradiction is detected, TMS evokes its own reasoning mechanism,
 42

dependency-directed backtracking, to resolve the contradiction by altering a

m i n i m a l set of b e l i e f s . In TMS, t h e b a c k t r a c k i n g u s e s d a t a
dependencies(McDermott 1983) kept in each TMS node, and is inefficient. In
KORE/IE, the dependencies are recorded independently for improving the
efficiency of the backtracking (for further details, see Section 3.1).
We use two flags in WM elements for representing default knowledge.
An internal expression for a WM element can be shown as follows;
Class_name(Time_tag,Default,Not_Contradicted,Yaluel,...Yaluen)
where the first argument " T i m e t a g " , the second argument "Default", and the
third argument "NotContradicted" are built-in arguments. The second and
third arguments are used as flags for representing the default knowledge: If
the value of the second argument is true, the WM element represents default
knowledge; otherwise it is not default knowledge. If the value of the third
argument is "nil", it represents a negative element; otherwise it represents an
affirmative element. If values for the second and third arguments are
unknown, the values automatically take "nil" and "true", respectively. The
first a r g u m e n t is a time tag which is a u n i q u e n u m b e r d e t e r m i n e d
automatically by the system. It is used by an inference engine for conflict
resolution. The larger the time tag, the more recently the element was
entered or modified. The rest of the arguments "Val u e 1, ---,Valu en" are values
defined by a user. The functor name "Class_name" is a class name for a WM
element defined by the user.
A WM element can be matched by a user by using a pattern description.
The pattern description consists of (1)a class name and (2)some number of slot
descriptions which are attribute-value pairs. In KORE/IE, the pattern
description is expressed as follows;
Class_name(Slot 1 =Value1 ..... Sloth = Yaluen)
where Slotn and Valuen are a slot name and its value, respectively. There are
t h r e e p r e r e s e r v e d s l ot s w h i c h a r e " t i m e _ t a g ' , " d e f a u l t " , a n d
"not_contradicted". These slots are used for representing the "Time_tag",
"Default", and "Not_Contradicted" arguments in the internal expression
mentioned above, respectively.
By using "make" command a default knowledge can be added to WM as
follows;
make(man(defauJt = true,name =taro,weight = 80))
where the argument is a pattern description which represents the knowledge.
The "man" is a class name. The "default" slot is the prereserved slot. The
"name" and "weight" are slot names defined by a user. For user's
convenience, the "make_assumption" command can be used for making the
same default knowledge as follows:
make_assumption(man(name = taro,weight = 80)).
The order in which slot descriptions are specified is not significant. Thus the
knowledge can also be made as follows;
make_assumption(man(weight = 80,name = taro)).
 43

The pattern description compiled into the internal expression as follows;

man(12,true,true,taro,80)
where the first and third arguments are a time tag and the value of the
"not_contradicted" slot taken by the system, respectively. The second, fourth
and fifth arguments are the value of the default slot, the name slot and the
weight slot, respectively. The order of the arguments except the built-in
arguments are determined by "literalize" command as follows;
literalize(man,[name,weight])
where the first argument defines the class name of the pattern. The second
argument is a list in which the order mentioned above is defined by arranging
the slot names.

3. The m e c h a n i s m for a n o n m o n o t o n i c inference

In KORE/IE, TMS is used as a mechanism for maintaining consistency

among WM elements. In order to implement the mechanism, we need the
following functions: (1)recording inference chains, (2)backtracking based on
the chains, and (3)resolving a contradiction for checking and denying default
knowledge which causes the contradiction.

3.1. Recording inference chains

If a rule is fired, an inference chain is recorded by k e e p i n g the

relationship between WM elements which matched LHS patterns in its
LHS(left-hand side) and new WM elements made by its RHS(right-hand side)
actions. In order to record the relationship, we use time tags for those WM
elements. The records for the relationships are managed by KORE/EDEN as
shown in Fig.3.
For example, consider following rules;
e x l : if a(slot = 1) & b(slot = I) t h e n modify(1,slot = 2).
ex2: if a(slot = 2) then make(b(slot = 2)).
where a rule consists of (1)the name of the rule (such as "exl" and "ex2"),
(2)the symbol ":", (3)the symbol "if", (4)the LHS which has some LHS patterns
(such as "a(slot= 1)", "b(slot= 1)", and "a(slot=2)"), (5)the symbol "then",
(6)the RHS which has some RHS actions (such as "modify" and "make"), and
(7)the symbol ".". The symbol "&" is a delimiter. The "make" command is
used to create a new WM element. The "modify" command is used to modify
one or more slot values of a WM element, in which the first argument
indicates which WM element to modify. After the following commands (that
is, Prolog goals);
?- make_assumption(a(slot = 1)),
rnake(b(slot = 1)).
 44

~ rl: if a & b then make(c). I

step 1

(~ ~step

~step 3

Inference chainq"-'~?step n
Fig.3. Recording inference chains by KORE/EDEN

then, the following WM elements are generated;

2: b(slot= 1)
1: a(slot = 1)*
where the numbers before the symbol ":" represent time tags. The symbol "*"
indicates that the WM element to which the symbol is attached is default
knowledge. The representation is used for displaying WM in KORE/IE. If
the WM elements are created, the rules mentioned above can be fired
successively and generate the inference chains as shown in Fig.4.

1
Firing the rule exl

~ rule ex2

Fig.4. An
Inference chains
example of inference chains

The inference chains can be kept by executing the following "add_to"

commands of KORE/EDEN;
add_to((3,a,fact),[(1 ,a,assurnption),(2,b,fact)],exa mple).
add_to((4,b,fact),[(3,a,fact)],example).
where the first, the second, and the third argument represent a node, a ]ist of
parent nodes of the node, and a name of the chains, respectively. In Fig.4, the
 45

node is illustrated by the WM element encircled by an ellipse. The

representation for nodes can be shown as follows;
(Time_tag, Class_name, Type)
where the first, second, and third argument represent a time tag, a class
name, and a type of the node. If the type is "assumption", then the node is
default knowledge; otherwise the node is not default knowledge.
KORE/EDEN provides basic functions (such as searching, deleting, appending
relationships, and etc.) to manage the relationships. For example, in Fig.4,
we can retrieve the default knowledge on which the node for the time tag 4
depends by achieving next Prolog goals;
?- su pers((4,Class,Type),Result,example),
bagof(Tag, C^member((Tag,C,assumption),Result),Answer).
where the "supers" is a KORE/EDEN command for searching reachable nodes
in the direction of parents from a node which is given the first argument. The
result returns to the second argument "Result". The third argument is the
name of the chains mentioned above. In the second goat which is a built-in
Prolog predicate, the default knowledge is bound to the third argument
"Answer". For the convenience and improving the efficiency, the
"assumption_super" command can be used, which is exactly the same as the
"super" command except that the result will contain only nodes whose type are
assumptions.

3.2. Controlling the TMS b a c k t r a c k i n g

In KORE/IE, the backtracking mechanism for TMS uses the inference

stepper mentioned in Section 2.1. The back data includes information for
finding and revising default knowledge which caused a contradiction. A flag
kept in each back data is used to prevent loops while attempting to resolve
contradiction. The flag takes on values of "real", "unreal", or "contradiction".
The "real" flag is used when the back data is kept during usual recognize-act
cycles. The "unreal" flag is used when the back data is kept during recognize-
act cycles under applying a procedure for resolving a contradiction. The
"contradiction" flag is used for marking back data which become the objects of
backtracking. The flags are needed for resolving a given contradiction by
revising one of current default knowIedge WM elements. While resolving the
contradiction, the flags are used to avoid entering the same procedure
recursively when a new contradiction is found. If a new contradiction is
found, the procedure needs to check the rest of the default knowledge instead
of trying to resolve the new contradiction. Fig.5 is given to clarify the
description.
Fig.5 illustrates the mechanism for using the flags when a contradiction
is detected, in which the default A and B represent default knowledge. It
shows the current inference steps by asserting negation of the default A in
order to resolve the contradiction. When the system backtracks to the step
 46

default B

Inferenc s t e p s ontradicti-~n ~ x ~ _ j /
Fig.5. Using the flags in back data

which includes the default A, the flags of the back data which depend on the
default are changed to "contradiction". The backtracking is realized by using
the back command in KORE/IE. These back data are used for undoing the
backtracking when the contradiction is not resolved by the assertion. By
asserting the negation, when new inference steps are done, the flags of the
newly recorded back data are set to "unreal". The flag "unreal" used for
indicating temporary back data. If the negation causes a new contradiction,
the system checks the rest of the default knowledge (such as the default B) by
undoing the inference steps which depend on the negation. The back data
with the flag "unreal" are used for undoing the inference steps. If the
negation does not cause a new contradiction, resolving the contradiction is
achieved by removing the back data with the flag "contradiction" and
changing the flag "unreal" of the back data to the flag "real".

3.3. The procedure for resolving contradictions

In order to maintain consistency among WM elements, a procedure is

used in KORE/IE. It is called the "contradiction", and provides a mechanism
which corresponds to that of TMS. In KORE/IE, for example, the procedure is
used as an RHS action in RHS as follows;
cry: if cry(name = X, f o r = Toy) then contradiction.
where the "contradiction" in the RHS is the procedure. This rule means that
it is inconsistent if there exists a WM element which matches the LHS
pattern. If the rule is fired, the procedure is applied and resolves the
contradiction by finding and revising a default knowledge which causes the
contradiction. The procedure is achieved by performing following steps:
Step 1: Get a list L of default knowledge on which a
contradiction depends. GotoStep 2.
 47

The elements of the list L are arranged in the order of decreasing time. The
time can be determined by using the time tags. The list L consists of nodes
which can be obtained by using the "assumption_supers" command mentioned
in Section 3.1.
Step 2: If the tist L ~sempty then stop the procedure and
the procedure fails; otherwise go to Step 3.
The procedure fails in the same manner as TMS when a contradiction needs to
be resolved by negating two or more assumptions together, since it only looks
for one at a time. If the procedure fails, the system helps us to resolve the
contradiction by showing the all default knowledge and inference chains
checked by the system. Then, for resolving the contradiction, we need to undo
and redo some inference steps by revising some default knowledge or some
rules.
Step 3: Get a defauit knowledge from the list L, which is
the first element (that is, Lhead) of the list L (that is,
L = [LheadlLtail] ), and backtrack to the inference step which
includes the default knowledge by using the back
command. Then, assert a negation of the default
knowledge. Go to Step 4.
When default knowledge was used multiple times, the procedure backtrack to
the step which includes the earliest default knowledge. As mentioned in
Section 2.2, the negation can be generated by modifying the value of the
"not_contradicted" slot to "nil" in the pattern (i.e. WM element) which
represents the default knowledge. This modification can be achieved by using
the "modify" command in KORE/tE.
Step 4: If a new contradiction is detected, then undo the
result of Step 3, let L = Ltail , and go to Step 2; otherwise
go to Step 5.
In order to undo the result, the procedure removes the back data which have
the "unreal" flags, and return the current state to the previous sate in which
Step 3 is not yet applied by using the back data which have the
"contradiction" flags.
Step 5: Change the "unreal" flags to the "real" flags, stop
the procedure, and the procedure succeeds.
In the procedure, removing the back data and changing the flag of the
back data are simply realized by using built-in Prolog predicates such as
"assert" and "retract" in which backtracking mechanism of Prolog can be
utilized effectively.

3.4. Other d e t a i l s

In order to resolve a contradiction, TMS changes statuses (i.e.IN or OUT)

of all the nodes which depend on the assumption node causing the
 48

contradiction. The process is performed in a bottom up manner by using

dependency-directed backtracking. The statuses need to be changed for
maintaining consistency of belief for a data base (or a knowledge base). In
this approach, for representing and maintaining the belief, it needs to record
data dependencies in each node. The data dependencies are information for
dependency-directed backtracking. In our approach, the procedure
"contradiction" resolves the contradiction in a top down manner. Namely, in
the process, default knowledge which correspond to the assumption nodes are
found first by KORE/EDEN. Then, the default knowledge are successively
tested and revised for resolving the contradiction. Since the belief is managed
dynamically by performing inferences, it need not record the data
dependencies in WM elements, which correspond to nodes of TMS. The
information for backtracking, which corresponds to dependency-directed
backtracking in TMS, is kept independently by using KORE/EDEN as
mentioned in Section 3.1.
An assumption node in TMS must have a IN list which is an empty list.
In KORE/IE, a default knowledge does not have such a constraint. Namely,
the default knowledge can depend on knowledge which is not default
knowledge. For example, a rule can be described as follows;
ex3: if a(slot = 3) & b(slot = t) t h e n m a k e _ a s s u m p t i o n ( c ( s t o t = 1)).
where a default knowledge is generated by achieving the RHS action. The
rule can be fired if the following WM elements are made:
2: a ( s l o t = 3 )
1: b(slot = 1)
By firing the rule, a new WM element (that is, a default knowledge) is added to
the WM as follows;
3: c(slot = 1)*
2: a(slot = 3)
1: b(slot = 1)
where the default knowledge which has the time tag 3 depends on the two WM
elements which have time tag 1 and 2. In order to resolve the contradiction
which depends on the default knowledge, the procedure "contradiction" needs
to modify not only the default knowledge but also the rule. In our approach,
the RHS of the rule is modified as follows;
ex3: if a(slot = 3) & b(slot = 1) t h e n m a k e _ a s s u r n p t i o n ( \ c ( s l o t = 1)).
where the pattern used in the make_assumption command is modified. The
pattern is obtained by changing the value of the not_contradicted slot to "nil".
In KORE/IE, the pattern for representing such negation can be simply
expressed by adding the symbol "V' to the pattern.
In KOREKE, there are no cyclic inference chains because the relationship
between knowledge is generated by using time tags. Namely, patterns are
classified and kept if they have different time tags. An aim of TMS is to
realize logical deduction for belief revision. The aim of our approach is to
realize a mechanism for nonmonotonic inference on a forward-chaining
 49

production system such as KORE/IE,OPS5, etc. The main difference between

TMS and our approach is that the purpose of TMS is to maintain the validity
status of nodes, and that of our approach is to maintain the validity of
inference steps.

4. An application

We have implemented an application to see how flexibly and effectively

the nonmonotonic inference mechanism can be used. The application is
called KORE/CDSS(Choice Decision Support System). KORE/CDSS helps
decision makers to make a reasonable choice from alternatives. Two
processes are used. One is a process which clarifies attributes used for
estimating the alternatives. The other is a process which determines the
priorities of the alternatives by estimating the attributes. In this section, we
sketch the mechanism for the estimation processes by using the nonmonotonic
inference in KORE/IE.
In the estimation, a pairwise comparison method based on AHP(Analytic
Hierarchy Process; Saaty 1980) is used. AHP is one of the weighting methods,
and has an advantage of being able to treat subjective judgments of a decision
maker. However, in the method, it is difficult to maintain consistency of the
judgments. In AHP, the consistency can be checked by using the consistency
index (C.I.) which can be represented as follows;
(~max"n ) l ( n - I)
where -~max and n are the maximum eigenvalue and the dimension of a
pairwise comparison matrix, respectively. The matrix is a square matrix
whose rows (or columns) consist of alternatives used for the comparisons.
Elements of the matrix are weighting numbers for the comparisons, where if
the value V for element aij is determined then the reciprocal value 1/V is
automatically entered in the aji. The value V takes a positive integer no
greater than 9. The larger the number, the more importantly the judgment is
agreed upon. If C.I. is tess than 0.1, the consistency of a matrix is considered
acceptable; otherwise it is inconsistent. If the inconsistency has occurred, we
need to revise a few pairwise comparisons for maintaining consistency by
finding inadequate comparisons which are the comparisons causing the
inconsistency. However, it is difficult to find them. In the worst case, it
forces us to revise all the comparisons.
KORE/CDSS provides a mechanism for maintaining the consistency of
the pairwise comparisons automatically. It uses the nonmonotonic inference
mechanism in KORE/IE. In the mechanism, we can group weighting
numbers into two classes such as (1)temporary numbers and (2)reliable
numbers, which can be generated by using the "make_assumption" and the
"make" command, respectively. If inconsistency has occurred, the temporary
numbers are checked and revised first for maintaining consistency. Fig.6
illustrates the process. In Fig.6, the symbol in a square represents the name
 50

run
!comparison1 I
Undoingin£erences
comparison2J

comparison1
[comparison2 change_weight]
t cOmparisOn-checkl

consis---teney contradiction

Inference chains ~ b lck

Fig.6. Maintainingconsistency of pairwise comparisons

of a rule. The rule "start" is used for initializing pairwise comparisons and
does the "ie_debug" command described in Section 2.1. The rules
"comparison1" and "comparison2" are used for obtaining a pairwise
comparison matrix. By the rule "comparisonl", weighting numbers are
entered and grouped into the matrix. The rule "comparison2" determines and
enters their reciprocal values into the matrix. The rule "comparison_check"
checks whether the matrix is filled or not. The rule "check" is used for
computing the C.I. of the matrix. The rule "consistency" is used for achieving
the procedure "contradiction" if the C.I. is not less than 0.i. The rule can be
described as follows:
consistency: if matrix(ci > 0.1)then contradiction.
The rule "change_weight" is used for revising the inadequate pairwise
comparison, which can be described as follows;
change_weight:
if \matrix_element(name = X,column = C,row = R,weight = W)
then
remove(I) &
new_weight(W,W2) &
make(matrixelement(name = X,column = C,row = R,weight = W2).
where the LHS pattern corresponds to the negation of the inadequate
comparison which needs to be revised. The negation is generated by the
procedure "contradiction" if the procedure finds the inadequate comparison as
a default knowledge. By firing the rule "change_weight", the inadequate
 5J

comparison is revised. In order to revise the comparison, the RHS action

"new_weight" is used. The action changes the old weighting value "W" to the
new weighting value "W2" which decreases the C.I. of the matrix. The range
of the change is +_2 because we assume that the old value is not so inadequate
for maintaining the consistency. If the C.I. becomes less than 0.1, the
procedure "contradiction" is stopped and the consistency is automatically
maintained. If the system can not resolve the contradiction by using the
procedure "contradiction", the system try to change several temporary
numbers at once. The temporary numbers to change are determined
systematically in which the numbers reduce Xma×effectively.

5. Conclusions

In KORE/IE, in order to realize a nonmonotonic inference, we introduce

the framework of the mechanism in TMS. Belief is managed dynamically by
performing inferences in the system. The mechanism for resolving
contradictions is simply implemented by utilizing and extending inference
steppers such as the run and back commands. The main features of our
approach can be described as follows: (1)The mechanism for the nonmonotonic
inference can be effectively realized without drastically changing the
architecture of KORE/IE. The method for implementing the inference is
generally applicable to forward-chaining production systems; (2)In order to
improve the dependency-directed backtracking used in TMS, the information
for relationship (such as data dependencies) is independently managed by
using KORE/EDEN; (3)The mechanism for checking the contradiction can be
implemented without using meta programming which is usually used as an
easy method for the mechanism in Prolog. The mechanism used in KORE/IE
contributes to speeding up the nonmonotonic inference in ProIog.
In Section 4, we have shown an example for using the mechanism in
KORE/IE. In the example, in order to maintain the consistency of pairwise
comparisons, we utilize the mechanism. By using the mechanism, the
consistency is effectively maintained with simple rule programming in
KORE~E.

Acknowledgment

The author would like to acknowledge the continuing guidance and

encouragement of Dr. Tosio Kitagawa, the president of IIAS- SIS, and Dr.
Hajime Enomoto, the director of IIAS-SIS. The author is deeply grateful to
Dr. Mitsuhiko Toda, IIAS-SIS, for reading the draft of this paper and giving
him many valuable comments. Also many thanks to Mr. Yoshinori
Katayama and Mr.Kunihiko Hiraishi, IIAS-SIS, for their useful comments.
 52

This research has been carried out as a p a r t of Fifth Generation C o m p u t e r

Project.

References

[Doyle 79] J.Doyle: Truth Maintenance System, Artificial Intelligence Vol.12,

pp.231-272(1979)
[Forgy 81] C.L.Forgy: OPS5 User's Manual, CMU-CS-81-135,July(1981)
[Matsumoto 87]Y.Matsumoto: Nonmonotonic Reasoning and Its Application,
The Journal of Institute of Electronics, Information and
Communication Engineers, pp.804-807(19887) (in Japanese)
[McDermott 83]D.McDermott: Contexts and Data Dependencies: A Synthesis,
IEEE Trans., VoI.PAMI-5, NO.3, pp237-246(1983)
[Reiter 80]R.Reiter: A Logic for Default Reasoning, Artificial Intelligence
Vol.13, pp.81-132(1980)
[Saaty 80]T.L.Saaty: The Analytic Hierarchy Process, McGraw Hill(1980)
[Shintani 86]T.Shintani, Y.Katayama, K.Hiraishi, and M.Toda: KORE:A
Hybrid Knowledge Programming Environment for Decision Support
based on a Logic Programming Language, Lecture Notes in
Computer Science 264, Logic Programming '86, pp.22-33(1986)
[Shintani 86b]T.Shintani: Knowledge Table: An Approach to Knowledge
Utilization Mechanism for Knowledge base, IIAS-SIS FUJITSU
LIMITED Research Report No.70, p38(1986)
[Shintani 88]T.Shintani: A Fast Prolog-based Production System KORE/IE,
Logic Programming: Proc. of the Fifth International Conference and
Symposium (edited by R.A.Kowalski and K.A.Bowen), MIT
Press(1988),pp.26-41(1988)
 Nonmonotonic Parallel Inheritance Network

Chiaki S a k a m a a n d A k / r a Okumura

Institute for New Generation Computer Technology

Mita Kokusai Bldg. 21F, 1-4-28, Mita, Minato-ku
Tokyo 108, Japan

Abstract

This paper discusses a theory of nonmonotonic inheritance reasoning in a semantic

network and presents a parallel inheritazace algorithm based on this approach.

1 Background

First, consider such an inheritance hierarchy in a semantic network.

Elephants are gray.
African elephants are elephants.
Clyde is an African elephant.
This hierarchy Js represented by a set of first order formulae as follows.

W = {WElephant(x) D Cra~,(x),
YzA frieanEIephant(x) D Elephant(x),
A f ricanEl ephant( clyde) }

In this case, Gray(clyde) is deducible from W. That is, inheritance is realized by the repeated
application of modus ponens.

However, when there are exceptions in the hierarchy, the case becomes somewhat more
complicated. Consider the following hierarchy.
Elephants are normally gray.
Royal elephants are elephants, but are not gray.
Clyde is a royal elephant.
This hierarchy can be represented by a set of first order formulae as follows.

W = {VxRoyalElephant(x) D Elephant(x),
 54

VxRoyaI Elephant( x ) D "~Gray(x ),

VxEIephant(x) h -~RoyaIElephant(x) D Gray(x),
RoyaIElephant( clyde) }

Suppose we add the fact, Elephant(taro), to W. When taro is not a royal elephant, the fact,
-~RoyaIElephant(taro), also has to be represented explicitly in W to deduce Gray(taro). Thus,
to represent such an inheritance hierarchy in first order formulae, all the negative information
has to be represented explicitly.
[ER83,Eth87a] represented such an inheritance hierarchy by l~eiter's default logic [l~eiS0].
For example, the above hierarchy is represented as:

W = {Vxl~oyalElephant(x) D Elephant(x),
RoyaI E l ephant( elyde ) }

D = {Elephant(x): Gray(x)A -~RoyalElephant,x,}(~-

Gray(x)

D is a set of defaults and read as: "when Elephant(z) holds, and Gray(x)A -~RoyaIElephant(x)
is consistent with this, then infer Gray(x)." As a result, Elephant(clyde) is deduced from W,
but RoyaIElephant(clyde) in W blocks the derivation of Gray(clyde) from the default D.
Besides, when Elephant(taro) is added to W and taro is not a royal elephant, Gray(taro) can
be deduced from D without -~RoyalEIephant(taro). Such an inheritance hierarchy is called a
nonmonotonic inheritance hierarchy, and it enables us to treat exceptions implicitly.
This formulation, however, seems to have some drawbacks since it requires as many in-
heritance rules as exceptions in a hierarchy. Moreover, update of such a hierarchy requires
modification of all the affected default rules as we~ as the corresponding first order state-
ments. It will become increasingly complex as the network grows, and does not make the most
of default reasoning:
The following sections present the formalization of nonmonotonic inheritance reasoning by
default logic in a different manner from Etherington, and also give a parallel algorithm based
on this approach.

2 T h e o r y of N o n m o n o t o n i c I n h e r i t a n c e R e a s o n i n g

First, two predicates are introduced.

 55

1. IS._A(x, y): an acyclic relation between an individual x and a class y, or a subclass x and
a supercla~s y; that is, IS_A(x, y) i f f x E y, or x C_y.

2. Property(x, y) (resp. -~Property(x, y)): a class or an individual x has (resp. has not) a
property y.

Now a nonmonotonic inheritance network is defined.

D e f i n i t i o n 2.1 A nonmonotonic inheritance network A = (W~ D) is defined as follows.

W : a consistent set of ground instances of either IS_A(x,y),Property(z,w),

or -~Property(u, v).

D
= {IS_A(x,y)^ Property(y, ): Property( ,z)
Property(x, z)
I S-A(x, y) A -~Property(y, z ) : ~Property(x, z) } []
-~Property( x, z)

E x a m p l e 2.1 Suppose the followingclassical hierarchy.

Molluscs are norma//y sheflbearers.
Cephalopods are molluscs but are not normally shellbearers.
Nautili are cephalopods but are shellbearers.
Fred is a nautilus.

In this hierarchy, cephalopods become an exception to molluscs with respect to shellbearers,

and nautili also become aa exception to cephalopods with respect to shellbearers.
Such a hierarchy is represented by A = (W, D), where

W = {IS.A(cephalopod, mollusc),
I S-A( nautil us, cephalopod),
IS_A(fred, nautilus),
Property(mollusc, has_shell),
-~P r opert y( cephaI opod, has_shell),
Property(nautilus, has_shell)}

As a result, the extension of A becomes:

E = W U {Property(fred, has_shell)}.
(InformaLly, an extension denotes a set of logical consequences of a default theory.) []
 56

The same hierarchy is represented by [ER83,Eth87a] as: 1

W = {VxCephalopod(x) D Mollusc(x),
VxNautilus(x) D Cephalopod(x),
Vxgautilus(x) D Shellbearer(x),
Nautilus(fred))

D = {M°llusc(x): SheUbearer(x) A ~Cephalopod(x)

Shellbearer(x)
Cephalopod(x) : -,S hellbearer( x ) h -~g autilus( x ) )
-~Shellbearer( x )
Compairing these two formulations, our approach describes the data in the network, W,
apart from the inheritance rules, D. This provides a simple description of inheritance hier-
arclfies, since default rules are independent of the network. It also facilitates updation of the
network, since it requires changing only the corresponding data in W and does not need to

modify defaults.
Such a system, which consists of a set of classes and a set of inheritable properties associated
with each class, is called a class/property inheritance system [Tou86]. Our approach defines
exceptions as nonmonotonic properties of classes, while IS_A hierarchy defines a monotonic
relation between classes. In the IS_A hierarchy, transitivity is not assumed since it generates a
redundant link. For example, IS_A(fred, cephalopod) is deduced using transitivity in Example
2.1, then it leads to an extension which contains -~Property(fred, has_shell), which is an un-
desired result. To derive a transitive class-subclass relation, it is necesarry to add, for example,
Property(nautilus, upper(cephalopod)) to W, to derive Property(fred, upper(cephalopod)).
A is called a normal default theory and has at least one consistent extension for every
consistent W [Rei80]. Nonmonotonic inheritance networks are classified by the number of

their extensions.

Definition 2.2 A nonmonotonic inheritance network A is definite iff it has only one extension.
[]

Example 2.1 is a definite case. However, there is a network with multiple extensions which
are inconsistent with each other. Consider the notorious example of Nixon diamond. The
problem is: Nixon is both a Quaker and a Republican, and Quakers are typically pacifists,

1[Eth87b] employs a different manner of representation, based on Touretzky's approach.

 57

while Republicans are typically not. Then, whether Ninon is a pacifist or not?
This hierarchy is represented in A with W:

W = {IS_A(nixon, quaker),
IS_A(nixon, republican),
Property(quaker, pacifist),
-~P r operty( r epubI ican, pacifist)}.

In this case, there exists the following two extensions which are inconsistent with each
other.
E1 = W u {Property( ixon,paci/ist)}
E2 = W U {-~Property(nixon, paci fist) )

Such a network which has more than one extension is called indefinite, and there axe two
attitudes for treating such an indefinite network [THT87]. A skeptical reasoner draws no
conclusion from ambiguous information, and hence offers no conclusion as to whether Nixon is
a pacifist or not. A credulous reasoner, on the other hand, tries to draw as many conclusions
as possible, and hence offers two alternatives: Nixon is a pacifist in one case, and is not in the
other.
From an algorithmic point of view, a skeptical reasoner always generates a unique extension,
then its algorithm seems to be simpler and more efficient than that of the credulous one, which
generates multiple possible extensions that grow exponentially as ambiguity increases. The
credulous attitude, on the other hand, seems to be more expressive than the skeptical attitude,
since the explicit representation of ambiguities suggests that there is some disagreement in the
network structure.

To take advantage of each attitude, an algorithm which can treat ambiguities but does
not generate multiple extensions is considered. The next section discusses such an inheritance
algorithm and its parallel execution.

3 Parallel Inheritance Algorithm

Inheritance algorithms combined with parallelism have been studied over the past few
years. NETL [Fah79] is a pioneering semantic network system. In NETL, inheritance is
performed by parallel marker propagation over nodes in a network. However, as is pointed out
by [ER83,Eth87a], NETL does not treat nonmonotonic cases correctly.
 58

[Tou86] has proposed some inheritance algorithms for a nonmonotonic inheritance system.
Those algorithms are based on the choice of inference paths in multiple inheritance and limited
parallelism is achieved. They offer a credulous inference system and also a skeptical version is
discussed in [HTT87]. These methods, however, require each derived path to contain its entire
derivation history and seem to become overloaded as the size of the network increases.
[EP~83,Eth87a] have shown a parallel algorithm based on their formalization and proved its
correctness, that is, all inferences lie within a single extension. However, this algorithm is not
complete in general; there might be some extensions which do not come out of the algorithm.
l~elated to this work, [Cot85] has shown a parallel connectionist axchitecture, but there is no
assurance of correctness.
The remainder of this section shows a r (parallel inheritance} algorithm for the nonmono-
tonic inheritance network presented in the previous section. First, notation in the algorithm
corresponding to A is given:

(a) property(class, O) iff Vcprop, Property(class, cprop) ¢ W and -~Property(class, cprop) ¢

W.
Otherwise, property(class, CProps) if] Vcprop E CProps, Property(class, cprop) e W,
and Vnot(cprop) e CProps, -~Property(cIass, cprop) E W.

(b) is_a(class, O) iff Vupper, IS_A(class, upper) ¢ W.

Otherwise, is_a(class, Uppers) iff Vupper e Uppers, IS-4(class, upper) e W.

Here, 0 denotes an empty set and notation which begins with a capital letter denotes a set.
Now, the r algorithm is presented below.

procedure ~r(input : class, output : Props);

begin
call property(class, C Props );
call is_a(class, Uppers);
Tamp ¢- 0;
while Uppers ~ O do
begin
select upper from Uppers;
cal~ ~r(upper, UProps);
Tamp ~-- Tamp U UProps;
 59

Uppers ~- Uppers - {upper}

end
call reverse( C Props, RevC Pr ops) ;
Props ¢-- CProps U (Temp - RevCProps)
end

procedure reverse(input : CProps, output : RevCProps);

begin
RevCProps = 0;
while CProps ~ ~ do
begin
select eprop from CProps;
if cprop = not(prop) then
RevCProps ~ RevCProps U {prop}
else RevCProps ~ RevCProps U (not(cprop)};
CProps ~ CProps - {cprop}
end
end

E x a m p l e 3.1 In Example 2.1, ~r(fred, {has.shell}) where

property(mollusc, {has_shell}), property(cephalopod, {not(has_shell)}),
property(nautitus, {has_shetl} ), property(fred, O), is_a(moIlusc, O),
is_a(cephalopod, {mollusc}), is.a(nautilus, {cephalopod)), and is_a(fred, {nautilus}).

In Nixon Diamond, r(nixon, {pacifist, not(pacifist)}) where

property(quaker, {pacifist}), property(republican, { not(paci fist ) } )r property( nixon, ¢),
is_a(quaker, 0), is-a(republican, 0), and is_a(nixon, {quaker, republican}). D

The procedure r produces a set of inheritable properties for an input class downwards
from upper classes to their subclasses. Nonmonotonic inheritance is achieved through over-
riding higher properties by conflicting lower ones. When there is more than one upper class
at is.a(class, Uppers), each upper class calls the recursive rr procedure independently. This
process will be executed in paxat]eI on a massively parallel architecture, where each processor
will have cost proportional to the length of the inheritance path. The 7r algorithm is imple-
mented in the parallel logic programming language GHC (Guarded Horn Clauses) [Ued86],
 60

and is shown in Appendix A.

The next proposition assures the soundness and completeness of the ~r procedure with
respect to a nonmonotonlc inheritance network A.

P r o p o s i t i o n 3.1 Suppose a nonmonotonic inheritance network A then

Vclass, ~r(class, Props) iff
Prop = {prop ] 3E , r 'opertV(d ,prov) e
U{, ot(prop) l e EA,
where E~ and Ej are extensions of A.

Proof See Appendix B. D

4 Discussion

The previous sections presented a formalization of the nonmonotonic inheritance network

and its parallel execution algorithm. Our approach enables us to define inheritance rules apart
from data in a network, and simplifies description and maintenance of the network. The
problem is, as is mentioned in an earlier section, a redundant IS-A link often causes some
ambiguities in the network.
The ~r algorithm produces a set of inheritable properties for an input class, and parallel
execution is achieved in multiple inheritance. When a network is definite, the properties
generated by the algorithm are included in an extension, while in the case of an indefinite
network, it collects ambiguous information from multiple extensions. Note that the output of
the algorithm is not an extension itself, thus there is no problem of logical inconsistency in the
indefinite case. It may seem that such ambiguities give us no information, but they suggest
the problem in the network structure and help us to reconstruct it.
In general, however~ it is not straightforward to decide whether an inheritance network is
definite or not. Let us consider an example from [San86] (Figure (a)).

[San86] has defined some basic structure types for inheritance networks and given sound
inference rules for these structures. According to his inheritance system, Clyde is not gray in
the above network. Whereas Touretzky's system [Tou86] yields two alternatives; Clyde is gray
in one case and is not in the other, and our system also returns gray and not(gray) for an
input class Clyde. This shows that the above network is interpreted as definite in Sandewall's
system, and indefinite in Touretzky's and ours. In this example, it seems to be more intuitive
 61

(a) gray (b) beer drinker

elephant ~ ~ ' ~ l man

royal chaplain ~ /
african marine

Clyde George

to infer that Clyde is not gray. However, as in shown in [THT87], there is a counter-example
which has the the same structure with (a), but the Sandewall's inference seems to be doubtful
(Figure (b)). In this case, whether George drinks beer or not is not dear and it depends on
the rate of beer drinking among marines°
These examples suggest that it is difficult to infer the intended meaning of the network
from its structure alone.

Acknowledgment

We would like to thank David W. Etherington for his comments for improving earlier drafts
of this paper. We are also grateful to the colleagues at ICOT for useful discussions and
comments.

References

[Cot85] Cottrell,G.W.: "Parallelism in Inheritance Hierarchies with Exception", IJCAI'85,

pp.194-202, 1985.

[ER83] Etherington,D.W. and P~eiter,R.: "On Inheritance Hierarchies with Exceptions", AAAF83,
pp.104-108, 1983.

[Eth87a] Etherington,D.W.: "Formalizing Nonmonotonic l~easoning Systems", Artificial In-

telligence 31, pp.41-85, 1987.
[Eth87b] Etherington,D.W.: "More on Inheritance Hierarchies with Exceptions", AAAI'87,
pp.352-357, 1987.
 62

[Fah79] Fahlman,S.E.: "NETL: A System for Representing and Using Real-World Knowledge",
MIT Press, Cambridge, MA, 1979.

[ttTT87] Horty,J.F., Thomason,R.H. and Touretzky,D.S.: "A Skeptical Theory of Inheri-

tance", AAAI'87, pp.358-363, 1987.

[Rei80] Keiter,R.: "A Logic for Default Reasoning", Artificial Intelligence 13, pp.81-132, 1980.

[San86] Sandewall,E.: "Nonmonotonic Inference Rules for Multiple Inheritance with Excep-
tions", Proc. of IEEE, vol.74, pp.1345-1353, 1986.

[Tou86] Touretzky,D.S.: "The Mathematics of Inheritance Systems", Research Notes in Arti-

ficial Intelligence, Pitman, London, 1986.
[THT87] Touretzky,D.S., Horty,J.F. and Thomason,R.tI.: "A Clash of Intuitions", IJCAI'87,
pp.476-482~ 1987.

[Ued86] Ueda,K.: "Guarded Horn Clauses", Lecture Notes in Computer Sciences 221, Springer-
Verlag, Berlin, 1986.

Appendix A
Here, we show an implementation of the 7r algorithm in GHC. GHC is the parallel logic
programming language developed as the kernel language of fifth generation project at ICOT.
The syntax of a clause in GHC is in the following form:
H :-G1,G2,...,Gm I B1,B2, ...,B~.
where the part preceding '[' is called a guard, and the part succeeding it is called a body.
A clause with an empty guard is a goal clause. The declarative meaning of the clause is the
same as Prolog.
The execution of a GHC program proceeds by reducing a given goal clause to the empty
clause as follows.

(a) The guard of a clause cannot export any bindings to the caller of that clause.

(b) The body of a clause cannot export any bindings to the guard of that clause before
commitment.

(c) When there are several candidate clauses for invocation, the clause whose guard first
succeeds is selected for commitment.
 63

Under these conditions, the execution of goal reduction is done in parallel. Now the proce-
dure is shown with an example of shellbearers.

/*** Nonmonotonic Parallel Inheritance Network in GHC ***/

~ inheritance procedure ~
pi(Class,Props,Tail):- true
property(Class,Props,Temp),
is_a(Class,Uppers),
has_propertyCUppers,UProps,Res),
filter(Props,UProps,Res,Temp,Tail).
has_proper~y([UClassIResZ],UProps,Tail):- true [
pi(UClass,UProps,Temp),
has.property(Rest,Temp,Tail).
has_property([], UProps,Tail):- true I UProps=Tail.
filter([CProplRest], In,Tail1~Out,Tail2):- CProp\=not(~ I
filter2(not(CProp),In,Taill,Temp,Tail3),
filter(Rest,Temp,Tail3,0ut,Tail2).
filter([not(CProp)~Rest],In,Taill,Ou~,Tail2):- true
filter2(CProp,In,Taill,Temp,Tail3),
filter(Rest,Temp,Tail3,0ut,Tail2).
filter(Dut, In,Taill,Out,Tail2):- true I
In=Dut,Tail1=Tail2.
filZer2(CProp,[PlIP2],Taill,Temp,Tail2):- Pl\=CProp I
Temp=[PllRest],
filter2(CProp,P2,Tail1,Rest,Tail2).
filter2(CProp,[PlIP2],Taill,Temp,Tail2):- Pl=CProp
filter2(CProp,P2,Taill,Temp,Tail2).
filter2(CProp,Taill, Taill,Temp,Tail2):- true I
Temp=Tail2.

~Z~ data ~ZZ

is_a(mollusc, Uppers):- true I Uppers=[].
is_a(aquatic, Uppers):- true I Uppers=[].
is_a(cephalopod, Uppers):- true I Uppers=[mollusc,aquatic].
 64

is_a(nautilus, Uppers):- true [ Uppers=[cephalopod].

is_a(fred, Uppers):- true [ Uppers=[nau%ilus].
property(mollusc, CProps,Tail):- true I CProps=[soft_body,has_shellITail].
property(aquatic, CProps,Tail):- true I CProps=[swimminglTail].
property(cephalopod,CProps,Tail):- true I CProps=[noZ(hasmhell),haslegsITail3
property(nautilus, CProps,Tail):- true I CProps=[has_shsll,not(swimming) ITail].
property(fred, CProps,Tail):- true I CProps=[american~Tail].

~Z~ execution results ~

?- ghc pi(fred,Props, O).
21 msec.
Props = [american,has~hell,not(swimming),haslegs,soft_body]
yes

This GHC program is easily translated into a Prolog program, which performs sequential
inheritance in a network.

Appendix B
First, some notation used in the proof is given. For an IS_A(x, y), the closure of x is defined
as a set, Upperk(x) (k >__0), as follows:

1. x 6 Uppero(x)

2. z e Upperk+l(x) i f f 3y, y e Upperk(x) and IS_A(y, z).

Note that the IS_A hierarchy is acyclic, then x ~/[Jk>l Upperk(x). Now we show the proof

of the proposition.

P r o p o s i t i o n 3.1 Suppose a nonmonotonic inheritance network A then

Vcl ass, ~r(class, Props) iff
Props = {prop l]Ei, Property(cla$s,prop) 6 Ei}
u{~ot(prop) 13Ej, ~Propert~(~1~ss,prop) e Ej}.
where E~' and Ej are the extensions of A.

Proof Suppose first that 3n, n + 1, Uppern(class) ~ O, Uppern+l(class) = O, then

vc~ e Vppe~(czass) , is_a(c~, 9) holds.
 65

Assume r(c,~, Props,~) and property(c,~, CProps~) then

Propsn = C Propsn U (Tempn - RevC Props,~)
= CPropsn U (9 - l:tevCPropsn)
= CPropsn
= {eproplProperty(cn, cprop) 6 W} U {not(eprop) I -~Property(cn, cprop) 6 W}

Next assume Vea e Uppera(class) (0 < k <_n), r(ca, Propsa) where

Propsk = {prop I 3E{, Property(ca,prop) e Ei}
U{not(prop) 13Ej, "~Property(ck, prop) e Ej} holds.

Let Vc~_~ E Upperk_l(elass), ~r(ck-1, Propsa-1), then

Propsa_l = CPropsk_l U (Tempa_l - RevCPropsk_l)
where property(ek-1, CPropsa_l).
(a) If uprop e CPropsa_l or not(uprop) 6 CPropsk-1, then
Property(ck_l, uprop) 6 W or -~Property(ck-1, uprop) E W.
(b) Otherwise, uprop 6 Tempa_l - RevCPropsk_l
or not(uprop) e Tempa-1 - RevCPropsa_l, then
uprop E Tempa-1 or not(uprop) E Tempk-1 where Tempk-1 = Propsk.
By the assumption, Vea 6 Uppera(cIass), then
3Ei, Properly(ca, uprop) e Ei or 3Ei,-~Property(ca, uprop) E Ej.
tn case uprop 6 Tempk_l, clearly uprop • RevCPropsa_l, then
"~Property(ca_l, uprop) ~. W and 3El, Property(ca_l, uprop) e E;.
In case not(uprop) E Tempa-1, dearly not(uprop) • RevCPropsa_l, then
Property@k-i, uprop) ~. W and 3Ej, "~Property(ca_t, uprop) e E i.
Therefore,
Props~_l C_{prop [ 3E{, Property(ea_l,prop) EEi}
U{.o (prop) I 3Ei,- Proper y(ca_ ,prop) Ej) (,).

While, let Vca-1 E Upperk_l(class), 3El, Property(ca_l, uprop) EEi,

or 3zs, Property(ea_ , prop) Zi.
(a) If Property(ck_l, uprop) E W or -~Property(ck_l, uprop) e W, then
uprop E CPropsk-1 or not(uprop) E CPropsa-1.
(b) Otherwise, 3Ei, Property(ca, uprop) EEi, or 3Ej,-~Property(ca, uprop) E Ej.
By the assumption, Vca 6 Uppera(class), uprop E Propsk,
or not(uprop) 6 Propsa, where ~r(ck,Propsk).
 66

In case 3Ei, Property(ck-1, uprop) E Ei, -~Property(ck_l, uprop) ~_W holds, then
uprop ¢ RevCPropsk_l and uprop E Propsk - RevCProps~_l.
So, uprop e CPropsk-1 U (Tempk_l - RevCPropsk-1).
In ease 3Ej, "~Property(ek_l, uprop) e Ej, Property(ca_l, uprop) • W holds, then
not(uprop) ~. RevCPropsk-1 and not(uprop) e Propsk -- RevC Propsk_l.
So, not(uprop) E CPropsk-1 U (Tempk-1 - RevCProp$k-1).
Hence uprop 6 Propsk-1, or not(uprop) 6 Propsk-1 holds, where r(ek-1, Propsk-1).
Therefore,
Propsk-1 ~_ {prop i 3El, Property(c~_l,prop) 6 Ei}
U{not(prop) ] 3Ej,-~Property(ek_l,prop) e Ei} (t).

Together from (,) and (4),

Propsk-~ = {prop l 3Ei, Property(ck_l,prop) EEi}
u{ ot(p op) 13E , P ope ty(ek_,,p op) e
By induction, we have the desired result.
 L o g i c Programming Debugger
Using C o n t r o l Flow S p e c i f i c a t i o n

Hideo Tamura Hideo Aiso

Keio University
Japan

In this paper , a control specification for logic

programming and the Deterministic Trimming Method for bug
detection are Proposed.
Using information from the control specification, the
query complexity of our Deterministic Trimming Method is
less than that of Shapiro's debugging system.

§1 Introduction
Prolog is a popular language for programming prototypes
of parallel language interpreters and natural language
processors, and for knowledge representation. But it is
said that it is difficult to show Prolog's execution flow
explicitly, so one cannot understand another's programs. To
reduce the programmer's load, many Prolog debugging systems
have been proposed and used. They are classified into two
groups.
Programmer oriented
Using debugger's facilities to detect bugs by programmer.
[Byrd78][Numao85]
System oriented
Debugging system asks programmer about a test program,
analyses that information and detects bugs.
[Shapiro83][Pereira,L~M 86][Lloyd87][Maechi87]
[Takahasi86]
If there were a program-specification that did not put
additional load on programmer and did not let down the
Prolog's abilities, it would be very useful and make Prolog
more powerful. In this paper, a control specification is
proposed, that is based on the Control Flow Model [Byrd78],
very familier to Prolog users as the Box Model. This
control specification can represent the programmer's
intention of how the program is to perform.
We p r o p o s e the Deterministic Trimming Method, which uses
control specification information based on Shapiro's
Algorithmic Program Debugging [Shapiro83].

§2 Control Specification
If there are specifications of each goal's mode
declarations, argument types and determinaney, it is easy
to debug and maintain a program. Even a beginner will
understand the program's meaning easily.
There is not a general Prolog program specification
notation, even if there were a new one proposed, it
probably would be difficult to get aquainted with, and
 68

programmer's load would increase linearly with the number of

program steps.
Therefore a control specification based on the poppular
control flow model is adopted and applied to this debugging
system.

§2.1 Control flow specification

Whereas the control flow model represents the dynamic run
time behavior of a program, control specification represents
it a statically. Although control specification cannot
represents a program's entire behavior, it can represent the
programmer' practical intention.
In the following, control specification is described, and
we i n d i c a t e how the programmer's intention is represented
and its relationship with the debugger.

The control specification mainly represents relations

between goals of program text. In figure 2.1, its difference
from the control flow model is an additional "FAIL' line.
This additional 'FAIL' line is used for 'cut fail', which
means backtracking invoked after a 'cut' operation dose not
affect goals before the 'cut' in the same clause.
F A I L ~
CALL~ pEXIT
FAIL ~
~---REDO
Fig 2.1
There are 5 patterns of control specification, which the
programmer can use to express control flow.

Fig 2.2
O Goal @ succeeds deterministically does not allow back-
tracking (deep backtrack).
O Goal ~ finitely fails.
O Goal @ always succeeds. If there is backtracking after
the goal's success, there must be alternative solutions
of the goal. There is never the ease of no solution of the
goal.
O Goal @ permits both success of failure. If the goal
succeeds the goal does not allow backtracking (deep
backtracking).
Examples of control specifications for some built-in
predicates is shown in figure 2.3.
F A I L ~
CALE--t~XIT CALKS---" 1 CALL i 'r "i--~EXIT
i~ ~EDO FAIL ~--i, REDO

CUT FAIL REPEAT

Fig 2.3
 69

§2.2 The relation of goals

Control specification notation represents control flow
between goals, though it does not regulate the behavior
within the subgoals. It regulates control flow on the
clause level of the program text. Each subgoal's control
flow is regulated on the level of it's own clauses. A goal's
determinancy is not specified, therefore the goal may suc-
ceed even though it causes shallow backtracking. A control
specification box represents a goal's final behavior of
success or failure or backtracking.
Example 2.4
Goal (~(~(~) m e a n s it does not fail. Backtracking might
occur inside of the goal, but it must eventually succeed.
Goal (~) have the posibility of backtracking, and eventually
it gets an alternative solution or fails. Cut operator
inhibits backtrack to th goal ~ from FAIL ~, and switches
the flow of control f r o m REDO l i n e to upper FAIL line.

NN-
Fig 2.4
Example of actual simple control specifications are fig
2.5 and fig 2.6. There are four types of 'append' usages,
type I being most common. The control specification in fig
2.5 allows type I, H and m but not type IV. I n f i g 2.6, all
types are allowed. In this way, control specification sets
constraints on goals, therefore reducing bugs. If a
programmer wants 'append' to be used as type I and IV,
control specification must use type IV.
append( [ HIT ],L,[ H]Z ] ):- append( T,L,Z ).
append( [],L,L ).

Fig 2.5 Fig 2.6

I append( +,+,- )
ex ?- append( [a,b,c] , [d,e],Ans )
Ans = [a,b,c,d,e]
II append( +,-,+ )
ex ?- append( [a,b],X, [a,b,c] )
x = [c]
]1I a p p e n d ( -,-,+ )
ex ?- append( X, [b,c], [a,b,c] )
X = [a]
IV a p p e n d ( -,-,+ )
ex ?- append( H,T, [a,b.c,d] )
H = [a,b,c,d]
T = [];
H = [a,b,c]
W = [d] ;
H = [a,b]
T = [c, d] ;
 70

§2.3 Clarify backtrack area by control specification

One of the reasons why another programmer's Prolog
program is difficalt to understand is that there is no sign
of where the backtrack area starts and ends in the program.
How t h e b a c k t r a c k area is represented in control specifi-
cation is discussed as following.
The program in Fig 2.7 searches for multiples of three
from numbers in the data base. If it finds one , it is saved
in the database. The reagion of backtrack is between the
'generate' and "test' goals, so there is a link between them
w i t h a F A I L a n d REDO l i n e . The backtrack invocation goal is
the last goal within a backtrack region, namely 'is'
tripple:- generate(A),test(A),assert(tripple(A)).
generate(A):- num(A).
test(A):- 0 i s A mod 3 .
hum(l).
num(2).
num(3).
num(4).
num(5).

Itripp]ei:--->lgenerate"/~[!~L~
[generatel:-'->l n ~ ....i ~
i I:' 1 'i ->
I hum I
Fig 2.7
Example of quick sort control specification
sort( X,Y ):- qsort( X,Y-[] ).
qsort( [],R-R ):- !.
qsort( [KeylOther],B1-B2 ):-
partition( Key,Other,Small,Large ),
qsort( Small,Bl-[KeylTmp] ),
qsort( Large,Tmp-B2 ).
partition( __,[],[],[] ):- !.
partition( Key,[ ToplOther ],[ ToplSmall ],Large ):-
T o p =< K e y , ! ,
partition( Key,Other,Small,Large ).
partition( Key,[ ToplOther ],Small,[ ToplLarge ] ):-
Key < Top , !.
partition( Key,Other,Small,Large ).
List 2.2 quick sort

Ipartit~on~ IToP =( ~ y ~ - ~ - - ' ~ l ~ i t i o n [ - - ~

[partitio~[:---~ [Key < TopI ~--~--~ [p~rtitio.[ "~

F i g 2.8
 7l

§2.4 The relation of Control Specification and debugger

When a user begins to debug, the debugging system will

show a default control specification for the user's program
in order to reduce programmers load to construct all control
specification. There may be many symptoms of bugs that
affect default control specification, so the programmer may
debug at this first stage.
At the next stage, the programmer must correct the
erroneous control specification. This step proceeds with a
control specification editor.
The debugging system needs the information below to.
analyse a bug.
OA goal's relations that can be known from control
specification.
( Existence of backtrack region )
OThe goals that violate the control specification.
( There is the case that the goal which should not fail
actually does fail. )

§3 Deterministic Trimming Method

The Deterministic Trimming Method is a bug detection
algorithm that is based on Shapiro's deride-and-query
algorithm, using the control specification. In processing
this algorithm, the information below is required.
Whether backtracking affects a particular branch; that
is, which nodes fail or succeed.
Whether a goal succeeds with a value that the programmer
does not expect. Identification of a node includes that
value.
Backtracking information ~ is obtained from the control
specification notation. For example, figure 3.1 shows that
a subtree rooted at the goal 'a' has a possibility of chang-
ing its computation tree due to the result of a subtree
rooted at the goal "b'. However, a subtree rooted at the
goal 'b' is not affected by a subtree rooted at the goal 'c'.

Fig 3.1
The information ~ about unexpected values is obtained
from the user. The debugging system has facilities for
managing information regarding binding values and nodes, and
the system finds the position where an indicated argument
value was produced.
For example, it will query the user about a goal that has
an erroneous argument.
When 'qsort( [4;1,2,5,3],[3,1,2,4,5] )' succeeds, the system
asks the programmer:
Query
qsort( [4,1,2,5,3],[3,1,2,4,5] ) (yes/no) ?n
The programmer responds "no".
The second argument is indicated and we call it 'false
instance production node'.
 72

Deterministic Trimming Method

step 1 Call the root node "R'.

step 2 Calculate t h e w e i g h t *~ o f t h e n o d e 'R', if it is 1
then report that there is a bug in the clause that
matched with this node.
step 3 Calculate the middle node of root "R', call it 'Q'.
Query the programmer at the node 'Q' is true or not.
step 4 If the node 'Q' is true, then trim the subtree rooted
at the node "Q', go to step 2.
step 5 If the node 'Q" is flase then indicate erroneous
value in the argument. ( The programmer may indicate
values in the argument ). The system will find nodes
that produce only indicated values, we call that set
of nodes as 'B'.
step 6 If there is at least one node of set 'B' outside of
the subtree rooted node "Q', trim the branches that
have no possibility of backtracking on the path from
the node "Q' to the node "R'. Trim the subtree rooted
at the node 'Q' and go to step 2.
step 7 If all nodes of set 'B' are in the subtree rooted at
the node 'Q', trim the branches that have no
possibility of backtracking on the path from the old-
est node within the set 'B" to the node 'Q'. We
rename the node 'Q' as 'R' and go to step 2.
step 8 If there is no indication of a false term at step 5,
rename the node "Q" a s 'R" and go to step 2.

note *-1
To calculate weight, assume that
- a node of a built in predicate = 0
- a leaf node ( except for built in predicate ) = 1
- other node = Z(weight of child node ) + 1

Example figure 3.2 represents step 7, Figure 3.3

represents step 7, These examples are assumed to be
non-backtracking case. In Shapiro's algorithm, region "A"
can be trimmed from the search area at one query. In Determi
nistic trimming Method, region 'A' and 'B' can be trimmed
from the search area in one query.

i n~tar~e
product ion node
p r ~ ion noae

Fig 3.2 step 6 Fig 3.3 step 7

 73

More detail of example step 7 and step 8

The case that there is a false instance node in a subtree
rooted at the queried node ( figure 3.4, step 7 ).
We c a l l the queried node 'Q',the false instance node 'B ~ .
In this case, we find a false instance node 'Q' whose false
value was actually produced by the node 'B', so there is a
high probability that there is a bug on the path from "B' to"
Q' ( There is aIso a possibility that the true bug is in
another path ). Therefore the path from 'B" to 'Q' and the
paths that affect it cannot be trimmed. Not only the "B'--'Q"
path but also the nodes that precede that path are bug
affection nodes. The noticeable nodes are after the 'B'-'Q'
path ( after the node 'B" ). If these nodes are reduced
deterministically, there is no problem, but if backtrack
occurs, the previously successful nodes are re-invoked in
order to search for another solution. Therefore we cannot
trim the nodes that invoked backtracking to the "B'~'Q" path.
In the case of figure 3,4 , we cannot trim the node "D"
that invokes backtracking to the node 'C' ( this information
is given by the control specification notation ). It does
not affect the 'B'~'Q" path that the node 'F' invokes
backtracking to the node 'E' and 'E' itself, so they are to
be trimmed. Even if the node 'D' possibly invokes backtrack-
ing, the child nodes of 'D" are to be trimmed because of
their determinancy.
Figure 3.5 and step 6 show tha case that a false instance
production node is in the outside of a subtree rooted at the
query node.
We a s s u m e queried node 'Q', a false instance production
node "B', and root node "R'. Similar to the case of figure
3.4 , considering the nodes that may affect "B'~'Q' path,
the nodes in the 'Q'~'R' path are doubtful. We c a n n o t trim
the node "D" that may invoke backtrack to the node 'C' on
the "Q'~'R" path, but any other node that is after "Q'~'R'
path except node 'D' can be trimmed.

§4 Bug Detection Strategy

The features of this program diagnosis is follows.
The diagnosis of a successful procedure with bug uses
the deterministic trimming method.
The step of diagnosing a failure procedure with bug is
same as that of Shapiro's trace query algorism, but we
improve the querying step by providing instances of
unbound variables in order to make it easier step for the
programmer. In a case that a goal has matched many
clauses ( causing shallow backtracking ), programmer can
select which clause must be matched, as this debugger uses
a history of test program's execution.
If there is a goal for which execution is inhibited
by the control specification, remove the execution history
after that goal. Because this removed area is generated
after production of a bug. If this is done before ~ and ~,
redundant query times can be reduced. For example, in
figure 4.1 it is the case that the node "J" behaves
 74

O~(--bug instance

/~B E F
bug instance product i on node

~ node t h a t _pr~__ibly a f f e c t s t h e above p a t h

O node t h a t does n o t a f f e c t
. . . . _pr~__ibly buggy p a t h
t h e above p a t h

Fig 3.4

bug i n s t a n c e
p r o d u c t i on ~) YJ~

bug i n s t a n c e g//~
d i scovery nodef/~

~ node t h a t possibly affects t h e above p a t h

O node that
possibly
does n o t a f f e c t
buggy p a t h
t h e above p a t h

Fig 3.5
 75

illegaly with respect to the control specification, the

subtrees of node 'C' and node 'D" are very big. In the
dividing query algorism, when a query node is selected in
order to devide a search area in half, then the selection-
is based to a big subtree, so the bug detection step
becomes roundabout. Given the control specification, we
can know that some bugs are in the nodes ('A','B','E','I',
"J') so irrelevant parts can be trimmed.

Fig 4.1
Combination of Deterministic Trimming Method and
trace query algorism

Practical diagnosis of a Prolog program has a case that

even if the program suceeds with bugs, it may include
subprograms that fail with bugs. For example, there are
clauses 'c1'c2' that may match with goal 'a'.
clause cl a:- gl,g2,g3,- ..... .
clause c2 a:- kl,k2,k3,. ..... .
First of all, match goal 'a' to clause 'el', but clause
"cl'is assumed to have bugs in its body and invoked back-
track.
Next, match goal 'a' to clause 'c2', goal 'kl','k2' are
executed successively. This program succeeds with bugs so
deterministic trimming method is adopted, but it cannot find
bugs in the successful path. Deterministic Trimming Method
only finds out that goal 'a' must not match with clause 'c2'
In this case, trace query algorism must be applied to the
failed path in which goal 'a' is matched with clause 'cl'
Therefore the system has to decide which clause to choose
and switch to the trace query algorism. If the system
applies all steps of the trace query algorism, a bug will be
found hut many queries are necessary. In order to solve this
problem, we combine both algorisms.

Fig 4.2

Figure 4.2 shows the execution history of a test program

using an execution tree representation ( hierarchical part
means 'OR' subtree of 'AND' tree ). Each triangle also
represents an execution tree. Oblique parts are successful
trees. If the bug search area is in a successful tree, the
system applies the Deterministic Trimming Method, otherwise
it applies the trace query algorism.
 76

The bug search step in figure 4.2 is shown below.

First of all, the system applies the Deterministic Trimm-
ing Method because execution tree rooted at the node 'l'.
Since node '3' has four alternatives, the system urges the
programmer to select a valid one. If the lowest layer of
node '3" is selected, continue to apply the Deterministic
Trimming Method, otherwise apply the trace query algorism.
If the programmer indicates that node '3' must not match
with that fourth layer, the system will trim the subtree
under node '3' when using Deterministic Trimming Method. It
will report that the node ~3' is a bug when using the trace
query algorism.

Bug search strategy related to an assert predicate

Features of an "assert' predicate are
-Assert predicate always succeeds.
• Variables are not bound by an 'assert" predicate.
Therefore execution flow is not affected directly by an
'assert' predicate. A bug symptom will appear since a goal
refers to a false instance produced by 'assert' predicate.
We w i l l show an example of using 'assert' predicate. At
position node @ 'assert( animal(cat) )" has done, then node
picks up an instanced 'eat' to bind variable 'A' of
'animal( A )'. We a s s u m e that this step caused binding
'A' with "eat'. All previous algorism diagnose this test
program as that the node @ is bug, and it cannot diagnose
before the node @.
If a bug node that was obtained by the all previous
algorism refered to the asserted term, then the system
searches this "assert ~ node , and calls this node 'Z'. The
system switches the bug search space to the subtree rooted
at the node that most recently queried an ancestor node of
'Z" ( if there is no such node, it witl be a present root
node ). Using all previous algorism, if the queried node is
true, then trim its subtree. However, if there is node 'Z'
in the trimmed area, the system begins to search for a bug
in the subtree rooted at that query node.

§5 Debugging Sequence

Debugging sequence of this system is shown below.

Execute test program.

OEnvironment is almost like C-Prolog
Modify control specification
OSystem shows test program's execution flow on a display
with control specification. Some times a programmer
notices a bug at this step.
~Modify control specification with a special editor.
Bug search step with query and respond.
OThe system will indicate a clause that has bug.
We w i l l show our debugger the diagnosis test program below.
Underlined statements are inputed by a programmer and a pair
of curly brackets means comments.
 77

cat isort
isort( [ XlXs ],Ys ):- isort( Xs,Zs ),insert( X,Zs,Ys ).
isort( [],[] ).
insert( X , [ Y I Y s ] , [ Y J Z s ] ) : - Y > X, i n s e r t ( X,Ys,Zs ).
insert( X , [ Y I Y s ] , [ X , Y I Y s ] ) : - X ~ Y.
insert( X,[], [ X ] ).
pgdb ( invoke the debugging system }
?- [+isort].
?- pgdb( isort( [2,3,1],hns ) ).
Ans = [2,3,1]
yes
9- halt.
Iogin
lisort/Z #- 4 1 i s o r t / Z I* ~[i.~t/3r ~ $,o
[ isort/2 I
I insert/3~- ~_[ ) / z ]~._ ~_[ i n s e r t / 3 ] - *
[ insert/3~ ~1 = ( / 2 ]~
letPt(h) r i g h t ( I ) up(k) ¢k0wl(j) ctl (spc) q u i t ( q )
Fig 5.1 Modify mode of control specification
( In Figure 5.1,you can move cursor of curved box by using
"h','j','k','l' keys( this is like 'vi' mode ). Each time
you push the space bar, the cursor's box control specifi-
cation will change. '/2' and '/3' represent the number of
arguments in a goal. }
Iogin
lisortl2 ~-- -*[isort/2 ]-~ ~ [ i n s e r t / 3 f ~ ~;
I isort/2 ]
[insert/3~- ~--I )/z I~ ~-1 i n s e r t / 3 1 ~
[ insert/3~ ~[ = ( / 2 ]~
l e f t ( h ) r i g h t ( I ) up(k) doNn(j) Ctl (spc) ¢lUit(q)
Fig 5.2 Modify mode of control specification
( In Figure 5.2, there must not be 'REDO' control after the
'>' built-in predicate. The programmer can modify it with
the box cursor. }
Iogin
[isort/Z ~-- - ~ l i s o r t / Z I -~ -~1 i n s e r t / 3 1 - *
[isort/Z J
[ insertl3~--
[insert/3~-
~1 ) I Z
~1 = ( / 2

Fig 5.3 Modify mode of control

{ In Figure 5.3, predicate "insert'
!~
I -~
~1 i n s e r t / 3 1 ~

l e f t ( h ) r i g h t ( I ) up(k) down(j) ctl (spc) q u i t ( q )

specification
should not fail,
k
but
actually it did fail. So t h e p r o g r a m m e r moves the cursor to
it and modifies its control specification. To e n d t h e
modification step, push "q' }
 78

( Bug search mode with query and reply }

Input y(es) or n(o)

Query isort( [3],[3] ) success ?

Input y(es) or n(o)

Query insert( 1,[3],[3,1] ) success ? n

select bug instance

space .. change scope, b.. bug,
q..end selection

insert( 1,[3],[3,1] )
* ...? space key (space bar was pushed}
insert( 1,[3],[3,1] )
*** ...? space key
insert( 1,[3],[3,1] )
...? space key
insert( 1,[31,[3,11 )
***** ...? b
insert( 1, [ 3 ] , [ 3 , 1 ] )
...? b
nsert( 1,[3],[3,1] )
...? b
insert( 1,[3],[3,1] )
• --? It

Bug clause is
insert( 21,[181 19],[ 181 20] ):-
18 > 21,insert( 21, 19, 20 ).
called by insert( 1,[3],[3,1] )
( Bug clause was found. }

§6 Evaluation
§6.1 Shapiro's deride and query algorism
The weight of a computation tree is reduced by one query
as described below,
N N W ~ (1- IN (6.1)
"N" i s l e n g t h of computation tree
'b' is factor of branches
Figure 6.1 and 6.2 represent the example where the reduced
weight i s n o t I/2N. T h e c a s e w h e r e t h e b r a n c h i n g factor
equals 2 is shown in figure 6.1. Within the weight below
the middle node ( weight lO ) o f r o o t n o d e ( w e i g h t 20 ) ,
the heaviest node weight is only 5 not 10. If this node is
true, and adopt this middle node, trim a subtree rooted
at this node. The total weight of this computation tree is
1
{ 1--2V }N
If the middle node is false, trim outside of the subtree
rooted at that node. the total weight of this computation
tree is
1N
2b
 79

In the divide and query algorism, the upper bound and

lower bound of total weight depend on the branching factor
'b', shown in formula ( 6.1 ). Figure 6,2 shows the case
where branching factor equals 3.
2O 3O

IG

5 S 5
Fig 6.1 Fig 6.2
The lower bound of query times, QL, is represented by
formula ( 6.2 ).
-1
QL = logz(1/2b) x 10g~N ( 6.2 )
The upper bound of query times, Qa, i s represented by
formula ( 6.3 ).
-I × log2N ( 6.3 )
Qa . . . . [ o g z ( l - I / 2 b )

Proposal ( Deterministic trimming method )

The weight of computation tree is reduced by one query.
--2-~--N =< W =< ( 1--21b~} aN ( 6.4 )
The lower bound of query times, QL, i s represented by
formula ( 6.5 )
-1
QL = l o g z ( a / 2 b ) x log2(N-B) ( 6.5 )
The upper bound of query times Qa i s represented by
formula ( 6.6 )
QH = - 1
[ogz{(l-1/2b) a} × logz(N-B) ( 6.6 )
Where a is the trimming factor , 0 < a _-< 1, a n d
B is the number of node that was trimmed due to
illegal control flow.

Comparing upper b o u n d Q~
Comparing Shapiro's algorism with Deterministic Trimming
Method is accomplished by comparing the coefficients of
formula ( 6.6 ) for different values of a. This comparison
is given in figure 6.3.
Figure 6.3 shows that the upper bound of query times in
Shapiro's algorism depends on the branching factor linearly,
but the Deterministic Trimming Method shows the increase in
query times by a percentage of the trimming factor.

Comparing lower b o u n d Q~.

As for the upper bound, comparing the lower b o u n d QL i n
both algorisms is the same as comparing coefficients in
formula ( 6.8 ) for different values of a. This comparison
is shown in figure 6.4.
Figure 6.4 shows that Deterministic Trimming Method is not
as powerful, but is lower, than Shapiro's method.
 80

14
--~ 13 trimming
12 f a c t o r c~ ~ j ~
"" II rl 1.00
-0
..~ I0 0 0,98, ~ - r,

v 8
• o,ss
7

0
6
S
4
3
2
1
0
I l 1 I l ) ( 1 i )
1 2 3 4 5 5 ? 8 9 10
branching f a c t o r b
Fi~ 6.3 Upper bound of query times,

0
J~

0
111
0 6-
trimminc
factor
0
0
+
A
1.00
0,80
0,60
0.40
m 0.20
0 5- • 0.I0

0 4-

0 3-

0 2-

0 I , J,,

2 3 4 5 6 ? 8 9 10
branching f a c t o r b
Fig 6.4 Lower bound of query times.

System configuration
This debugging system consists of
-Execution block for test Prolog program,
( This block is based on C-Prolog source code )
-Editor for control specification notation,
.Bug search block.
They are invoked in sequentially, communicating by
intermediate files.
 81

Conclusion
The size of the intermediate files produced by the
debugging system i s 164 K B y t e s w h e n t h e n o n - b a c k t r a c k test
program fills the local stack ( 128K B y t e s ), so it has no
practical problem.
We h a v e implemented a system that support meta-predicates
( "call', "not' ) . We a r e c o n s i d e r i n g how to cope the
"assert" predicate. There are two types of bug the "retract"
predicate can cause:
Q The case where variables are bound to the arguments of
"retract',
The case where a term is deleted by the "retract"
predicate.
This system can cope with the case Q, but not ~.

Acknowledgement
This research, and the development of the debugging system,
was supported by Also Laboratory at Keio University. The
authors would like to thank Dr. Yang Rong, Mr. Show Ei Ken,
and members of the Logic Programming Group.

References
[Lloyd87]Lloyd,J.W.
Declarative Error Diagnosis,
New G e n e r a t i o n Computing,5 1987 Springer-Verlag
[Maechi87]Machi MAEJI and Tadashi KANAMORI
Top-down Zooming Diagnosis of Logic Programs,
35th Conference of Information Processing Society of Japan,
1987
shinichi MORISHITA and Masayuki NUMAO
BPM-Visual prolog computation model and debugger PROEDIT2,
Proc. of The Logic Programming Conference "86 , 1986
[Numao85]Masayuki NUMAO
PROEDIT Screen oriented Prolog programming Environment,
Proe. of the Logic Programming Conference "85, 1985
[Pereira,L.M 86] Pereira,Luis Moniz
Rational debugging in logic programming,
Lecture Note in Comp. Science 225, t986 Springer-Verlag
[Shapiro83]Shapiro,E.
Algorismic Debugging, M.I.T. Press, 1983
[Takahashi86]Hidehisa TAKAt~SHI, Etsuya SHIBAYAMA
PRESET - A Debugging Environment for Prolog,
Lecture Note in Comp. Science 1986, Springer-Verlag
Naohisa TAKAHASHI, S a t o s h i ONO a n d M a k o t o AMAMIYA
Parallel-processing-oriented Algorismie Bug-detection Method
for Functional Programming Languages,
Study of Basic Software,1984
 ALEX • The Logic Programming Language
with Explicit Control and without Cut-Operators

Masaki Hoshida* and Mario Tokoro

Department of Electrical Engineering

Keio University
Yokohama 223, Japan

We propose a Prolog-like new programming language "ALEX". It is an extension of

Prolog from the view point of procedural l)rogramming, keeping the merit of declarative
feature of Prolog. In ALEX, the backtrack mechanism is divided into two categories :
"Recall" and "Retry". "Recall" corresponds to "essential trim and error" and it appears
rarely in the program, while "Retry" corresponds to "branch on condition" and it appears
frequently. In ALEX, these actions are controlled by different methods, and "Recall" is
restricted in default state. Control in ALEX is described as prefix of predicate. As a
result, cut-operators are excluded, and we achieve easy debugging and higher readability.

1 INTRODUCTION
Prolog has a backtrack mechanism and enables declarative programming. Its drawbacks
are that over-strong backtrack and cut-operators make the program difficult to write, read
and debug. Prolog's mechanism has the following problems.

1) The way of controlling the backtrack mechanism seems unnatural to users.

We waaat to use the backtrack mechanism for a limited part of the Protog
program. At the moment, cut-operators are used to control backtrack. They
are inserted at the parts that the user wants to execute deterministically.
However, since the deterministic part in a program is generally larger than
the nondeterministic part, this way of controlling backtrack seems unnatural
for users. It is also the major cause of bugs.

*Currently with ICOT on leave from Matsushita Electric Industrial Co.,Ltd.

 83

2) Backtrack control cannot be described on the source code level on which that control
is achieved.

If we read just one clause of' a Prolog program, we cannot tell which predicate
may fail and which predicate may be reactivated by backtracking. To find out
these things, we must read at a deeper level of the Prolog program. In other
words, in the Prolog program, the source code level on which a backtrack
is achieved cannot contain a description of which predicate is reactivated by
backtracking. The description must be made at a deeper level to insert the
cut-operators. This makes the program difficult to code, debug, and read.
In any programming language, the control of the program should be described
at the source code level on which the control is achieved. It is unnatura~ that
the callee program describes the control of the caller program.
In the hierarchy of the program, if the deeper level is related to the shallow
level, and the modification or the bugs in the deeper level affect the shallow
level, there is no point in making the program hierarchical.

3) Execution errors are not categorized

In Prolog, "program execution error" and "failure in logical meaning" cannot

be distinguished. Only "fail" appears and causes the backtrack. This makes
it difficult to detect bugs; we have to scrutinize the program execution trace.

We think that the most important feature of a programming language is that the user
can express what he wants to do easily and exactly. Even a programming language that
has excellent theoretical features is impractical if it is difficult to write, read and debug.
In order to make Prolog more practical and more useful, cut-operator should be scrapped.
Many Prolog debuggers have been tried, but we think that the control specification of
Prolog itself should be re-examined. In this paper, we propose "ALEX" (Advanced pro-
Log with Explicit expression), a programming language designed to improve the control
specification of Prolog.
In ALEX, the problems we have just described are solved as follows.

Problem 1)

In ALEX, the backtrack mechanism is divided into two categories. One of

them corresponds to "branch on condition", and the other corresponds to
"essential trial and error". And the latter is restricted.

Problem 2)

In ALEX~ control of backtrack can be described on the source code level on

which that control is achieved. And that control is described explicitly.
Problem 3)

In ALEX, the concepts "fail-possible-predicate" and "fail-impossible-predicate"

are introduced.
 84

2 BASIC CONCEPT OF ALEX

2.1 Recall and Retry
In ALEX, in order to restrict the backtrack, the backtrack is divided into two categories;
Recall and Retry.

Recall:

Recall is backtrack that the clause that finished execution is abandoned, and
its alternative is re-selected.

7-1 $-1
.... p,Q . . . . . . . . p,Q ....

P:-A,B.--(1) P:-A,B,C. --(3)

P:-C,D. --(2)
B:-X. --(4)
E B:-Y.-(s)
(a) (b)
Fig 1 Example of Recall

(a) shows a case where predicate Q fails and predicate P is reactivated by backtracking,
that is, the cIause for the call P is re-selected. Suppose that when P was called, clause
(1) was selected and the calls A,B were executed deterministically.
In this case, for the call P, clause (1) was selected and executed successfully. We
named this kind of backtrack "Recall".
(b) shows a case where predicate Q fails and predicate B (called from P) is reactivated
by backtracking. Suppose that when B was called, clause (4) was selected and the call C
and the call X were executed deterministically. In this case, as in Fig 1, the clause that
finished execution is abandoned, so this backtrack is categorized to "Recall".
Recall corresponds to "essential trial and error". Because the clause that finished
execution is abandoned and its alternative is re-selected.

Retry:
Retry is backtrack that, during the execution of a clause, the whole clause
fails and its alternative is re-selected.
 85

..... e . . . .

P:-A,B,C. --(1)

~ P:-D,E.--(2)
Fig 2 Example of Retry

Fig 2 shows a case where P was called, clause (1) was selected, A finished execution
deterministically, the next B failed, and clause (2) was re-selected.
In this case, clause (1) has not finished execution. In other words, the clause for the
call P has not been decided yet. In this sense, Retry corresponds to "branch on condition"
like "if then else" or "case".

2.2 Basic Concept

As we said earlier, Recall corresponds to "essential trial and error", and Retry corresponds
to "branch on condition". Generally, "branch on condition (Retry)" appears frequently
in the program, while "essential trial and error (Recall)" appears rarely.
tt is better to mark on the part which appears rarely. Because, if we mark on the part
that appears frequently, it is easier for us to forget to put on marks.
In Prolog, we have to mark on the parts that appear frequently, deterministic part,
by cut-operators. It seems unnatural. Further more, in Prolog, if we omit a cut-operator,
it is hard to find out what we did wrong.
The reason why Prolog has such problems is that both Recall and Retry are controlled
only by cut-operators. Cut-operators restrict both Recall and Retry at once.
In ALEX, the two categories "Recall and Retry" are controlled in different ways.

• Recall cannot be performed in default state; it can be performed only when it is

explicitly permitted.

* Retry can be performed in default state; it can always be performed unless it is

explicitly prohibited.

The prohibition of Recall in default state reduces the number of bugs caused by omis-
sion of cut-operators.

3 SYNTAX AND SEMANTICS OF A L E X

3.1 Abstract of Syntax and Semantics
Syntax:

An ALEX procedure is a set of horn clauses like Prolog. The pattern of horn
clauses are same as Prolog, except for the following Dature.
 8{5

• A predicate in the head or the body of a clause can take a prefix such as "?", ":",
"??', "::', " + ' . The prefix is used as control information.

• A clause can take control of a structure named "block". "Block" is a line of predi-
cates which is clipped between "<" and ">".

* The head and body of a clause are divided by "<-".

Semantics:

An ALEX clause is executed "depth first, left to right" like Prolog. The
unification mechanism and the pattern of atomic formulas are the same as in
Prolog except for the backtrack mechanism.

3.2 Control on Recall (clause level)

In ALEX, when a clause finishes execution successfully in default state, its alternatives
are prohibited from executing in the case of backtrack. In order to let its alternative
execute in the event of backtrack, a prefix "+" must be added to the head of the clause.

P<®A,B,C.--(1)
+P<-D,E. --(2)
P<-F,G. --(3)

Fig 3 Control on Recall No.1

In Fig 3, if clause (1) finishes execution successfully, (2) and (3) (alternatives Of (1))
cannot be executed in the event of a backtrack that reactivates the predicate P. If clause
(2) finishes execution successfully, (2)'s alternatives (3) can be executed in the event of
backtrack.
Retry can be performed in default state (in spite of the prefix " + ' ) . In Fig 3, if
predicate A fails, clauses (2) and (3) can be tried.

3.3 Control on Recall (predicate level)

In ALEX, the predicates in the body of a clause are mainly categorized into three classes
as follows.

1) Predicate without Prefix

• Never fail.
• If it fails, ALEX interpreter halts with a error message.
• If it finishes execution successfully, it cannot accept "Recall".
- Error message shows the functor of the failed predicate, the contents of
the variable when the predicate was called, and the call number.
 87

- After the indication of the error message, the ALEX interpreter asks the
user whether they want to continue the execution and regard the error as
a failure or to make an abort.

2) Predicate with prefix "?"

• It can fail.
• ff it fails, it causes Recall or Retry.
• If it finishes execution successfully, it cannot accept Recall.

3) Predicate with prefix ":"

• It can fail.
• If it fails, it causes Recall and Retry.
• If it finishes execution successfully, it can accept Recall.
Note that the prefix "+" is necessary on the clause level to accept
Recall, in addition to the prefix ":".

Table 1 Prefix and Control

no prefix "?" prefix %" prefix

possibility of failure x o o
possibility of Recall X X O

Table 1 shows the essence of the control on Recall. In this table, "possibility of failure"
cannot be described in Prolog. "Possibility of Recall" is expressed by using cut-operators
in Prolog, but it is implicit from a view point of well-understandable programming. Be-
cause when we want to know the determinicity of a predicate, we have to read the deeper
level of the program.
In ALEX, user can know "possibility of failure" and "possibility of Recall" of the
predicate only to see the prefix of the predicate. So ALEX is easier to read.
Predicate that cannot fail or accept Recall are most often used in the program. There-
fore in ALEX this kind of predicate is expressed as the predicate without prefix.
In this way, ALEX is designed to reduce the amount of the prefix for control in the
source code so far as the necessary information is available. With ALEX, it is easy to
detect bugs caused by the user omitting a prefix at a place where it was necessary.

P<- :A,B,:C,?D,?E,F.

Fig 4 Control on Recall No.2

In Fig 4, when ?D or ?E fails, only :B (inside :B) can accept Recall. C and ?D cannot
accept Recall. If :B fails, P also fails as a result. In this case, if P has alternatives, the
alternatives are tried. "Inside :B" means the predicates that are directly or indirectly
called by :B.
 88

.... :P,Q,?R . . . .

P<-A,:B,C.
P<-D.

A.
+B.
B<-:B.
C.
D.

Fig 5 Control on Recall No.3

Fig 5 shows the case in which :P and Q are executed successfully, and ?R fails. Then,
as a result, :B is Recalled. In this case, the prefix ":" is necessary on both of P and B.
And P is not reactivated by a failure of ?R. Because P does not have a prefix "+" at the
head of the clause.

.... P,Q,?R . . . .

P<-A,:B,?C.

A~
+B.
B<-:B.
C.

Fig 6 Control on Recall No.4

Fig 6 shows the modularity of the ALEX program. Suppose that the user wants to
permit :B to accept Recall only from ?C. When programming in Prolog, in a case like
this, the user has to insert the cut-operator behind ?C to protect :B from the backtrack
from the upper level, (?R) in this example.
In ALEX, in this case, P has no prefix, so inside P is protected from Recall that is
caused by a failure that occurs outside P.
If the user writes B for :B by mistake, when ?C fails, P fails and the error occurs as
the result. Therefore the error is shut in a local part.
"Protection from Recall" and "Confinement of error" can achieve the easy debugging
and the higher confidence of program. And the explicit expression of the backtrack control
by the prefix of the predicate can achieve higher readability.
In Prolog, it is often difficult to see the reason why a cut-operator was inserted in a
place. Unnecessary cut-operators are often inserted, too.
To make a failure unconditionally, an operator "?" (not a prefix) can be used. When
this operator is executed, it makes a failure.
 89

3.4 Control on Retry

Retry can be achieved in default state; and it is caused by the failure of the predicate
with the prefixes "?" and ":". In order to prohibit Retry, the prefix "??" and "::" are
used for "?" and :q".
The definition is shown as follows in addition to the section.

4) Predicate with prefix "??"

• It can fail.
• If it fails, its parent-predicate fails immediately.
• If it finishes execution successfully, it cannot accept Recall.

5) Predicate with prefix "::"

• It can fail.
• If it fails, its parent-predicate fails immediately.
• If it finishes execution successfully, it can accept Recall.

"P's paxent-predicate" means the predicate that calls P directly.

.... O,?P,Q . . . .

P<-?A,??B,C. --(1)
P<- D . - - ( 2 )
A.
B.
C.
D.

Fig 7 Control on Retry

In Fig 7, if ??B fails, ?P fails immediately. (2) is not tried. In ALEX, Retry can be
prohibited in such a way. Note that, Retry is prohibited very rarely, so these two prefixes
are not often used.
To make the parent-predicate fail unconditionally, the operator "??" (not a prefix)
can be used. When this operator is executed, the parent-predicate of this operator fails
immediately.

3.5 Block
To restrict the area where Recall can be performed, the control structure "block" is
introduced. "block" can protect an area from Recall. "block" is the sequence of the
predicates clipped between "<" and ">".
"block" has the following features.

• A failure that occurs outside "block" does not reactivate inside "block".
 90

• If a whole "block" fails, the "block" makes an error, and the system halts with an
error message.

• To give the possibility of failure to "block", it is necessary to put a prefix "?"

or "??" at the head of the "block". "?" means mere fail, and "??" means the
parent-predicate fails.

• "block" can be nested.

P<- A, ?<:B,C,?D>, ?E.

Fig 8 Block

In Fig 8, Recall can be performed from ?D to :B. Once ?D succeeds, :B cannot be

reactivated any more.

4 PROGRAMMING IN ALEX
In this section, we show an example of ALEX program.

gps(S,[l,S,[l).
gps(S,G,S,~)<-?subset(G,S).
gps(S,G,Sa,ANS)<-
sorting(G,GN),
difference(S,GN,DIFLIST), --(1]
:member(DIF,DIFLIST), --(2)
operator(DIF,OPLIST), --(3)
:member(OP, OPLIST), --(4)
precondition(OP,PRECOND), --(5)
?gps(S,PRECOND,S1,ANS1), --(6)
apply(OP,S 1,$2), --(7)
?gps(S2,G,S3,ANS2), --(8)
cat(ANS1,OP,ANS1,ANS). --(9)
difference(_,[],[]).
difference(S,[XIY],R)<-
?member(X,S), --(10)
differen-ce(S,Y,R).
difference(S,[X IY],[X t R] )<-
difference(S,Y,R).
+member(X,[X I -]).
member(X,[_ t Y])<- :member(X,Y).

Fig 9 A Part of the GPS Program in ALEX

 91

Fig 9 is a part of the GPS (Genera1 Problem Solver) program to resolve the monkey
and banana problem.
The calling pattern at top-level is

?-gps(<initial state : input>,

<goal state : input>,
<final state : output>,
<operator sequence : output>).

to be concrete

?-gps([at(monkey,banana,),at(box,b),empty],
[hold(banana)],
FS,
ANS).

GPS calculates the difference between current state and goal state, -(1), and select the
most important difference, -(2). Next, GPS lists all operators to reduce the difference,
-(3), and select the most effective operator, -(4). Then, GPS applies the GPS algorithm
itself recursively regarding the precondition of this operator as the new goal, -(5)(6). If
this application succeeds, GPS applies the operator, -(7). Next, GPS regards that result
as the new current state, and GPS applies the GPS algorithm recursively further, -(8).
At the end, the sequences of the operators are concatenated, -(9).

In this ALEX program, we can easily know the control from the prefixes as follows.
In the 3rd clause of gps/4, when ?gps/4 at (6) ol- (8) fails, :member/2 at (4) is reacti-
vated by Recall, and if (4) has no alternative, :member/2 at (2) is reactivated next. All
predicates in the 3rd clause of gps/4 except (2) and (4) cannot be reactivated by Recall.
Thus, in ALEX program, the control of a program can be expressed on the source-code
level at which the control is achieved.
When we translate this program into a Prolog program, we have to insert the cut-
operators inside sorting/2, different/3, apply/3. It is not natural.
When we read the Prolog program that corresponds to this program, we cannot know
the control of the program only from the 3rd clause of gps/4. We have to read the deeper
levd.
Also, :member/2 at (2) and (4) has a different function from ?member/2 at (10) in
spite of calling the same substance. :member/2 at (2) and (4) are used to generate some
data repeatedly by Recall, so it is necessary to be nondeterministic. While ?member/2
at (10) is used only to test the data and it should not be reactivated by Recall, so it is
necessary to be deterministic.
In ALEX, the use of the predicate is expressed in the caller side, so only one substance
is necessary, and the use of the predicate can be clearly understood.
 92

5 PROLOG AND ALEX

In this section, we try to categorize the use of cut-operator of Prolog from the point of
view of how the programmer uses it. And we show how these categories are expressed in
ALEX about each case.

- Here, the word "freeze" is used to mean fixing the choice of clause.
- The word "procedure P" means the clauses that have the P at their heads.

1) The case that a clause freezes itself

For using a procedure deterministically, prohibiting the reactivation of procedure
can be achieved to insert cut-operators into clauses which construct the procedure.
An example of a Prolog program and an ALEX program that have the same func-
tion is shown below.

Prolog

member(X,[X 1_]):-!.
member(X,[_ I Y]) :-member (X,Y).

ALEX

member(X,[X I -]).
member(X,[_t Y])<-?member(X,Y).

In this example, member/2 in Prolog freezes itself. This kind of use of cut-operator
is usually called "Confirming the Choice of a Rule". Most cut-operators are used
like this.
In ALEX, this determinicity is realized in default state, which is without prefix "'"
and "+" .
That is, Recall can be achieved only when the following two conditions are satisfied
at the same time.

- The prefix " + " is at the head of a clause, on the clause level.
- The prefix ":" is at the predicate, on the predicate level.

It means that ALEX has a twofold protection mechanism, so it is easy to find bugs.
But in ALEX, it is necessary to put the prefix "?" at member/2 in the body. Be-
cause it may fail. If the programmer forget to put this "?" and member/2 fails,
it becomes a bug. But this bug comes out as an executior~ error, so we can find it
out easily. This is the essential difference between the hug caused by the tack of the
prefix "?" and the bug caused by the lack of the cut-operator. In the latter case,
we have to spend a lot of effort looking for bugs.
 93

• 2) The case that a procedure, though it can freeze itself, is frozen by the other one

Prolog

member(X,[X [ _]).
member(X,[_[ Y]):-member(X,Y).

subset ([X 1Y],Z):-member(X,Z),!,subset(Y,Z).

subset([],_).
ALEX

member(X,[X i_]).
member(X,[_ I Y])(-?member(X,Y).

subset ([X [Y],Z)(-?member(X,Z),??subset(Y,Z).

subset(I],_).

In case 1), member/2 freezes itself. In case 2), subset/2 freezes member/2, even
though member/2 can freeze itself. In actual programming in Prolog, cases 1) and
2) are often used redundantly. It means, for instance, cut-operators are inserted
into both member/2 and subset/2 in case 2). The larger a program becomes, the
less we can rely on determinicity of the predicate on the lower level. So we insert
more redundant cut-operators. We can also insert cut-operators to achieve higher
readability, although we know they are redundant. In ALEX, by a twofold protection
mechanism on the clause level and the predicate level, such determinicity is achieved
in default state, cases 1) and 2) are the most often used cases of cut-operators.
according to our experience, more than 80% of cut-operators are used in this way.

• 3) The case that a procedure cannot freeze itself, and is frozen by the other one
(Whether a procedure is frozen or not depends on the condition of its surroundings)

This is the case that a predicate is frozen when the test by the other predicate suc-
ceeds. It is often called "termination of generation and test". This case is essentially
different from cases (1) and (2) because a predicate cannot freeze itself.

Prolog

member(X,[X I _]):-!.
member(X,[_l Y]):-member(X,Y).

moreThanFive(X,LIST):-
member(X,LIST),X:,5,!.

?-moreThanFive(X,[2,3,5,7,t 1]). (call)

 94

ALEX

member(X,[X I _]).
member(X,[_ I Y])<- ?member (X,Y).

moreThanFive(X,LIST)<-
?< :member(X,LISW), ?(X>5) >.

?-moreThanFive( X, [2,3,5,7,11]). (call)

In this example, moreThanFive/2 is defined as the predicate that returns a number,

which is more than five and found first, member/2 generates the element, and "XL5"
tests it. member/2 cannot determine when it should be frozen. So member/2 cannot
freeze itself. In ALEX, in such case, "block" is usually used. By using "block", when
"?(X>5)" succeeds, :member/2 cannot be reactivated by Recall any more. And by
using "block", it becomes clear that "generation and test" is achieved only between
:member/2 and "?(X>5)". In this example, if the number more than five cannot be
found, moreThanFive/2 fails. It is expressed by "?" at the head of the block.
If moreThanFive/2 is always called without the prefix ":", it is not necessary even
to use "block".

• 4) The case that the special predicate "fail" forces "the parent goal" to fail
This case is often called "cut-and-fail" or "negation". In ALEX, it corresponds to
the operator "??". That is, "!,fail" has same function as "??".

Here, we tried to categorize the use of the cut-operator in Prolog. It is often ambiguous,
for what kind of purpose a cut-operator is used. In ALEX, a prefix expresses the control
explicit, so it is easier to read. The control of Recall can be expressed on the level on
which it is achieved, so we can do natural programming.

6 CONCLUSION
We believe that the most important feature of a programming language is that it should
be easy to write, debug, and read. Prolog has high potential, but it is inconvenient in
above feature. Many extensions of Prolog have been developed. But if it takes more time
to debug, the extension is not practical.
In ALEX, we suggests a way to improve the execution control of Prolog.
Backtrack mechanism is divided into two categories. One of them corresponds to
"essential trial and error" and the other corresponds to "branch on condition". We named
the former "Recall" and the latter "Retry". Generally, "branch on condition (Retry)"
appears frequently in the program, while "essential trial and error (Recall)" appears
rarely. So Recall cannot be achieved in default state in ALEX. This enables "Protection
from Recall" and "Confinement of error", and it reduces the number of bugs by omission of
cut-operators. As a result, easy debugging and higher confidence of program are achieved.
 95

Control in ALEX is described with the prefix at the head of predicate. The predicates
in the body are categorized into five categories. And a prefix shows the possibility-of-
failure or possibility-of-backtrack of a predicate. By the prefix, the user can know the
control of program on the source code level on which the control is achieved.

References
[1] Clocksin,Merish "Programming in Prolog", Springer-Verlag(1981)
[2] Naish,L "PROLOG CONTROL RULES", IJCAI86(1986)
[3] Sawamura,H "The Determinacy of Prolog Predicate(Call)",
Computer Software Vol.4 No.4,(1985) in Japanese
[4] Chikayama,T "Unique Features of ESP',FGCS84(1984)
[5] Chikayama,T "ESP Reference Manual", ICOT TR-044,(1984)
[6] Nakashima,H "Prolog/KR User's Manual" METR 82-4,
Dept. of Mathematical Engineering, University of Tokyo(1983)
 LATTICE PROGRAMMING METHODOLOGY
K o u l c h l ~hlkunaga Shlnichi M o r i s h l t a
IBM Research, Tokyo 1~esearch Laboratory
5-19 Sanban-ch% Chiyoda-ku, Tokyo 102 JAPAN

1 B a s i c idea

Logic programming gives us a way to look at conventional algorithms f~om a different viewpoint
and to reorganize or classify them. The separation of the logical aspect of the problem f~om
the procedural control for searching for a solution makes the comparison of algorithms easier.
As a result, in some cases algorithm derivations become easier in logic programming.

For example, many sorting algorithms can be regarded as the specializations of the following
naive algorithm.

sor~(Lis~, SortedLis~) ~- perm~ta~ion( Lis~, Sor~edLis~) & ovdered( SortedLis~).

Since the logic and control of this program are separated, it is easy to understand what this
program does. If we assume the depth-first search as the control strategy, it does generate-
and-test by backtracking between permutation and ordered. In this case, its inefficiency comes
t o m the excessive backtracking and the specializations are concerned with reducing unnecessary
backtracking by incorporating the test function of ordered into permutation. This clarity of
algorithm derivation is one of the advantages of logic programming.

However~ logic programming has still disadvantages. Since it only provides syntactic unification,
programmers have to invent the way to represent their data structures by using this unification.
These problem specific inventions sometimes obscure the declarative clarity of logic programs.
In these cases, we would rather prefer to directly represent objects and their properties of the
problem domain.

We found a way to clearly represent some algorithms whose goal is to produce minimal (or
maximal) objects under the given ordering. Examples of the category include sorting, minimum
spanning tree, and shortest path finding, which we will explain later in this paper.

We conclude that all the advantages of logic programming come f~om one theorem : its declar-
ative semantics coincides with its procedural semantics. If we look carefully at the proof of this
theorem, we find it depends on a technique which is not specific to logic but more widely appli-
cable. The technique is largely based on lattice theory. This observation suggests an extension
of logic programming as follows.

* {true, false} ~ c o m p l e t e lattices

 97

• implication --* lattice order

• and -~ meet ~ functions over complete lattices

• or ~ join

As we will briefly explain in the later section~ the above theorem still holds under some conditions
on functions over complete lattices. Since lattice is the simplest algebraic structure which many
algebras of d a t a structures satisfy. Thus, this extension has the potential for ~epresenting some
d a t a structure producing algorithms in a natural way. We call this exiension lattice programming
and show its apphcations in the following sections.

2 Lattice valued programs

In lattice programming, we use lattice valued programs in which each atom has a complete
lattice as a range and each clause is associated with a function over complete lattices.

A complete lattice Z is a partial ordered set which has least upper bound (US) and greatest
lower b o u n d ( ~ S ) for every subset S of L. The least upper bound and the g~eatest lower bound
of L itself are denoted by TL and -kL respectively.

In a lattice valued p r o g r a m P , atoms which have the same predicate symbol and the same arity
are associated with the same complete lattice. Let LA denote the complete lattice foz A. A
lattice valued program P is a finite set of clauses each of which is associated with a function over
complete lattices. Suppose that a clause A ~- B 1 , ' " , B,~ (n > 0) is given, then it is associated
with a function f : LB~ x . . - x LB~ ~ LA. If a clause has the fo~m A *--, then it is associated with
a constant function, t h a t is, an element of LA. We will denote the p a i r constitutes of a clause
A 4- B 1 , - ' - , Bn and its function f by (A ~ B I , . . . , B,~, f ) which we call a clau,e pazr. We also
associate a goal e-- A 1 , - . . , A,~ with a function F fzom LA~ x . . . x LA~ to any complete lattice
L. We will denote the pair constituting of a goal ~-- A 1 , " ' , A n and F by (~- A 1 , - - - , A n , F )
and call it a goal pair.

We assume that every function F : L1 × --- × Ln --+ L in any program satisfies the following two
properties:

* Elementwise least upper bound preserving :

F(al~-.., a~_l,uS, ai+~,"-, an) = U { F ( ~ , . . . , a~_~, ~, a,+~,..-, aN) I z e s}
holds for any subset S C L~(i = 1 , . . . , n ) and for Vax C L x , . . . , V a ~ - t E Li-l,a~+x C
L~+i,---,Va,~ E Ln

• Elementwise strict :
F(ai,-", an) = -LL whenever at least one al = I-L i
 98

We wilt refer this requirement on a program by (a). Note that an elementwise least upper bound
preserving function is monotonic increasing.

Informally, the declarative meaning of a clause (A *- B I & - . - &B,~, f) is that , for any ground
substitution p, the value of Ap is greate~ titan or equal to f(bi,'", b,~) (bi is the value of Bip).
Let denote this fact by Ap ~ f(Bip,"',Bnp). We also describe the value of a goal pair
(~---Aip,'",A,~p, F) by F(Aip,'", A,,p) for any ground substitution p.

Suppose that a goal pair (~- A i , ' " , A n , F) is given and ~ Ai,...,A,~ is ground. We can
restrict the value of F ( A i , " ' , An) as follows ; if Ai is selected from A i , - - - , A , , and Ai and A
in (A ~ B i , - " , B m , f} are unifiable by an mgu 0. Since A ~_ f ( B i , ' " , B m ) holds and F is
monotonic increasing~ we have

F(Ai,...,AI,...,An) D .F(AiO,'",f(BiO,'",BmO),'",A,~O).

Suppose that we have the following sequence, after deriving goal pairs repeatedly.

F(A1,'",AI,"',An) ~ F(AIO,...,f(BtO,...,BmO),...,A~O) ~ ...... ~ F'

Then this process is called a derivation and an equality F ( A i , - . . , An) -7 F ' is obtained.

By the way, in each step of deriving goals there are two choices, that is a choice of an atom to
be selected and a choice of a clause. In our framework, we can fix one rule for selecting an atom
(see section 5), however, there still exist a choice of a clause. Due to this choice in each step, we
may have many derivations for the same goal pair.

These derivations can be made up into a tree structure like a SLD-tree in usual logic program-
ming. See Fig 1. Such a tree is called a derivation tree. Irt a derivation tree, each branch from the
root to a leaf corresponds to a derivation. Since we can regard a derivation as an inequality which
gives a lower bound of the given goal pair, each leaf node gives a lower bound for the initial goal.
For example, in Fig.l, the leaf F ' is a lowez bound for F ( A i , ' " , A,~) (F(Ai,-..,A~) ~_ F').
In this way, from a derivation tree, many lower bounds for the initial goal may be obtained.

F(A~,...,A.---,A.)

A n f(B1,---,Bm),0

f(AiO,...,f "" O),...,A,O)

Fn
Fig.1 A Derivation Tree
 99

In section 5~ we show that the least upper bound of these lower bounds is equal to the solution.
In this section, we discussed the declarative and p~ocedural semantics informally. Its ~igorous
treatment is found in section 5.

3 Programming methodology

The steps of lattice programming proceeds as follows.

• Formulate the complete lattices to represent all of the potential data structures (including
the p~oblem solution) which may be handled by the program.

• Show the solution be restricted by a set of inequalities and represent them as clauses

• Show the least upper bound of ~he lower bounds derived from the inequalities is the
solution.

The following example illustrates this process. Suppose we are given a problem of producing a
sorted sequence out of a sequence of natural numbers (with +co). The first step is to construct
a lattice of sequences of natural numbers Z,~q. A sequence is denoted by ( z r , . - - , z,~) (n _> 0).
Let Ln,,,q be the set of sequences of the same length n. In L,,,s~q, elements are ordered in
lexicographic order _.

v(~,,~,.-.,~,,),(~,~,~,~,...,~,,,) e L .... q ( ~ , ~ , - ' - , ~ , , ) E (.~,~,---,~,,,)

iJr~, > ~,~ or (~, : ~ and ( ~ , . . - , ~ , , ) _E (~,~,...,~,,))

Note that Ln,,~q is totally ordered. The sequences of different lengths are unordered. The whole
lattice L,~q is obtained by taking the direct sum of {L. . . . q I n is a natural number } whose top
(T,¢q) and bottom (±,¢q) are newly introduced.

Next, consider a binary function % over L,~q. % concatenate two sequences. For T,cq and ±~¢a,
the meaning of % is defined as follows:

T,¢¢ % s = T,~q, s % T,~q = T,~q, ±~¢q % s = i,~¢, s % i,~q = i , e q for any s in L,~q.

It can be easily shown that % satisfies requirement (a).

Let sequence(S) be an atom with which L,¢q is associated. Let d_union(S1, $2, S) be an atom
expressing that S is a disjoint union of $1 and $2. d_union(S1, $2, S) can be defined by a
conventional logic program, .and it is associated with {true, false). The following discussion
shows sequence(S) defined below calculates the sorted sequence of S.
 100

(sequence({A}) ~-, (A)) for each element A

f(X1,X2, X3) a2-fif Xl=true then X2%X3 else 2,eq

Since % satisfies requirement (a), it is obvious that f also satisfies (cO. To clarify the declarative
meaning of this p~ogram, we will replace ~ and & with .~ and % respectively. Furthermore
we will use A- as a syntax sugar for d_union. Then we get the following inequalities which
sequence(S) satisfies.

sequence({A}) ]_ (A) for each element A

sequence(S1 + $2) _ sequence(St) % sequence(S2) (i)

Now, let us execute a sample goal pair, for example, (4- sequence({4, 3, 1, 2}),I>, where I is the
identity function over £,eq. This goal pair is intended to calculate the value of sequence({4, 3, 1, 2})
itself. After searching exhaustively (see l~ig.2), we get the set of inequalities from the derivation
tree.

( ,~q~ce({4, 3, 2, z, }) ~ x I x is a permutatio, of (~, 3, 1, ~) } (2)

~equencK{4, 3, 1, 2})

4",
~quenc~({~, 3}) S sequence({1, 2}1
d',
(~eq~enc~({4}) % ,eq~en~e({3})) % (~eq~nee({i}) % sequen~({2}))
d',
((4) % (3)) % ((~) % (2))
( = <~, 3) s (1, 2) = (¢, 3,1, 2) )
Fig.2 (Case %) h Derivation Tree for (~---sequence({4, 3, 1,2}),I)

Since the least upper bound of {X [ X is a permutation } is (1, 2, 3, 4), the above set of inequal-
ities is reduced to sequence({4, 3,1, 2}) _~ (1, 2, 3, 4). Hence, (1, 2, 3, 4) becomes the solution of
the given goal.

There is one problem in this program. The set (2) is quite large. This is caused by the naivety
of concatenation. ]?or example, during computation, we may be able to conclude that :

sequence({4, 3}) ~ (4, 3) s~q~e=ce({1,2}) ]_ (1, Z)

 101

Then, the above program restricts seq~ence({4, 3, 1, 2}) by using the simple concatenation as
follows.

,eq~e~e({4,3,1,2)) _~ ,eq~ce({4,3)) % ,*q~e~e(b,2}) _~ <4,3) % 0,2) = (4,3,1,2>

If we use the merge operation $ defined below instead of %, we can restrict sequence((4, 3, 2, 1})
more strongly. Because (4, 3)$(1, 2} = (1,2,4, 3) is much greater than (4,3, 1,2). Moreover,
we do not lose the solution (1, 2,3,4) by this change. This observation suggests the following
program which realizes the so-called merge ,ort.

sequence({A)) ~_ (A) for each element A

seq~e~ce(S~ + S2) 3 seq~e~ce(S~) $ ,eq~e~ce(S2)
T,~q$s = T,~q, sST,~q = T,~q, ±,~q$s = ±,~q, s$±,~q = ±,~q
(,~, a ~ , - . . , a,,)$(b~, b2, . - . , b,,) ~
if ,~ _< b~ then (~,(~2,.-.,~,~)*<b~,b=,---,b,,))
else (b~,(~,~,...,~,,)S(b~,-..,b,,)> (3)

'$' meets requirement (ix). It is easily shown that, given two sorted sequences, merge produces
the sorted sequence of them. Hence, for the goal (*- seq~tence({4,3,1,2}),I), we obtain the
same lower bound (I, 2, 3, 4) f~om any derivation of this goal, however S is decomposed into $1
and $2 (see Fig.3).

sequence({4. 3.1.2})
d',
sequence({4.3}) * ~eq,,~ce({1.2})

4",
((4), (3>), ((1) * <2))
( = (3.4>, 0.2) = (1.2.3.4))
Fig.3 (Case $) A Derivation T~ee for (e- sequence({4,3, 1,2}),I)

This gives the chance to reduce the complexity by fixing the method to decompose S. For
example, if we decide $2 to be a singleton, we get the insertion aort as follows.

~eq~e~ce({A}) ~_ (A) for each eloment A

sequence(S1 + {A}) ~ sequence(S1) * sequence({A})
 102

In this way, lattice programming give the following stepwise refinement method from naive
algorithms to more efficient ones.

* Reduce the number of derived lower bounds for each goal by introducing new operations
while retaining the solution.

• Incorporate an efficient search strategy into the program which utilizes the properties of
the resulting set of lower bounds.

4 A further example

The above argument which clarifies the relation between algorithms by showing their algorithm
derivations also applies to other problems. Due to the space limitation, we only show one example
program which solves the single source shortest paths problem and its algorithm derivation from
more naive versions is left to the reader.

Let G = (V, E ) be a directed graph which consists of a vertex set V and a edge set E each element
of which has non-negatlve length. The length of an edge e(v, ~ ) from the vertex v to the vertex ze
is denoted by length(e(v, w)). A path from vl to vn is denoted by e(vl, v2).e(v2, v 3 ) - ' - e(v,~-l, v,,).
The length of a path p denoted by length(p) is the sum of the lengths of the edges on p. The
problem is :

Given a source s, find a shortest path from a source s to every vertex in V

(For simplicity, we assume that there exists an edge from a source s to itself whose length is
zero). First of all, we construct a complete lattice £ , ~ h , whose elements are sets of paths from
s. Every element z in L,~ath, is an array whose length is the number of vertices V. Each cell is
indexed by a vertex v E V, and its content z[v] is a path p from s to v or nil which means there
is no path. We define length(nil)=+co. A order E in L,~o, th, is defined as follows.

Vz,y e L,a,~,h, : x E Y <~ Vx E V~ length(z[v])_ tength(y[v])

Secondly we define another complete lattice L¢~g¢, of edges E. Leage, is a fiat lattice in which
all edges in E are unordeIed and its top(Teag¢,) and bottom(-Le~g~,) are newly introduced. It
can be proved that/~,~ath, and L,doe, are complete lattices.

Now we will give a function f : L~g~, x L , ~ h , ~ L,~h, which incrementally calculates path
sets from s by adding the effects of an edge. (By the way, to meet elementwise strictness, we
will add a-,~a~h, to L,~a~h,, where -L,~th, is less than every element in L , ~ t h , . )
 103

V~ E ~,~#af, hs -- {-Ls,paf.h, } , Ve(1/i,~2) E ~edges -- {T,dg¢,,-L~dg¢,} f(e(~Jl,V2), ~) = lJ st.

if (~[~,] ¢ nil a~d ~[~.] = nU)
then ( ~ [ ~ ] = , [ ~ , ] . e ( ~ , , ~ ) and ~[~] = =[~] for a , y ~ C V - { ~ } ) else ~ = ±,~o,~,

It is easily verified f meets requiremen~ (a). Then the following program calculates the shortest
paths from s to all vertices in V.

w t h 4 V + {X, Y}) ~ f(edgdX, Y), p~th,(V + {X}))

( a e L,,,o,~,, W e V - (,), ~[~] = nil, and ~[,] = ~(~,~), length(e(s, s)) = 0 )

If we rewrite this top-down recursive program to a bottom-up iterative one and apply branch
and bound strategy to the search, we get the non-deterministic version of the labeling algorithm
of Ford [9]. Further, if we adopt the shortest first search as the special case for branch and bound
strategy, we get the Dijkstra's shortest path algorithm ([8]). Note that if we change the lattice
and the operation appropriately, the same program represents an algorithm which calculates the
minimum spanning tree of G.

5 Theoretical Background

Now we give the declarative and procedural semantics special to our framework. The theoretical
results are also presented, but the details of thei~ proofs([1]) are omitted owing to the space limit.

5.1 declarative semantics

An interpretation I of a lattice valued program P associates every gIound atom A with a value
in LA, which will be denoted by val(A,/). The order in interpretations is introduced as follows
: I i E 12 ¢ ~ val(A, I1) E val(A, I2) for any ground atom A. Using this order, the set of
interpretations of a program P forms a complete lattice.

Some notations are useful for convenience and readability. Let C be a clause A ~ B x , ' " , B n .
C + denotes the head A and C - denotes the body B 1 , " ' , B n . Let (A *-- B1,...,B,~, f) be
a clause pair and I be an interpretation. For example, f(val(Bl,I),..-,val(B,~,I)), can be
denoted by V A L ( ( B t , . . . , B n , f ) , I ) .

Let (C, f ) be a clause pair. The clause C is true in an interpretation I with ~espect to f iff
for every its ground instance C', vat(Ct+,I) _ V A L ( ( U ' - , f ) , I ) holds. In this case, I is called
 "104

modelfor (C, f). An interpretation M is called modelof a program P iff M is a model for every
clause pair in P. Let M(P) denote the set of all models of a program P, then its greatest lower
bound I-]M(P) also becomes a model, which is called the minimal model of P.

In logic programming, van Emden and Kowalski [5] associate a p~ogram P with a mapping Tp
over interpretations and show that models of P are fixpoints of Tp. Furthermore, by virtue of
mathematical resulis of Tp, Tp can be used to discover properties of models and provides the
link between declarative and procedural semantics. We also follows this method and define To
as follows. Let P be a program, A be an ground atom and I be an interpretation.

val(A, Tp(I))
= U{VAL((C'-, f), I) IC' is a ground instance of C in <C, f) and C '+ = A}
-~ / L ~ t (if there exists no clause whose head and A are unifiable)

As in logic programming, it can be proved that Tp is a monotone function over the complete
lattice of interpretations and all models can be characterized in terms of Tp. Hence we can
establish a useful connection between models and fixpoints, that is the least fixpoint of Tp is
equal to the minimal model of P. Furthermore, since Tp is a continuous mapping, the least
fixpoint of Tp is equal to Tp T w. This is the very result that connect declarative and procedural
semantics of lattice valued programs.

We will conclude t,his subsection by collecting the results.

Theorem.
Let P be a program and A be a ground atom. Then

val(A, nM(P)) = val(A, lfp(Tp)) = val(A, Tp T w)

5.2 procedural semantics

Now we give the procedural semantics, such as a derivation, a derivation tree and a derived
value. Then we will show one result that combines declarative and procedural semantics.

Let P be a program and R be a computation rule which selects at least one atom from a goal
if it is not an empty goal. A derivation of a goal pair {Go, Fo> via R is a sequence of tuples
{{Gi, Fi},(Ci, h),O~) (i = 0,1,--.), in which a goal (Gi+l, Fi+l) is derivedby a goal (Gi, Fi)
and a clause (Ci, fl} as follows. Let Gi be ~ A 1 , - - - , A~,/~ be g ( X i , ' " , X~), Am be the atom
selected by R, Ci be A *-- B I , ' - - , B q whose head and Am are unifiable by an mgu Oi. Then

Gi+l = ~ ( A 1 , . . . , A m - i , B I , . - . , Bq, A , , + I , . - . , Ak)el

Fi+l = g(X1,..., Xm-1, fi(Y1," • •, Yq), X,~+i, -. •, X~)
 105

Especially, if (C~, fl) is (A e--, f~) (fi E LA), then Gi+l is ~ (A1,---, Am-l, A m + I , ' " , Ak)~i
and F~+I is g ( X 1 , . . . , X m _ l , fi, X,~+l, oo.,Xk). Note lhat if the region of F0 is I , then the
region of each F; is also L.

A derivation can be finite o~ infinite. A finite derivation can be ~uccez~l or failed. A successfui
derivation ends in a goal pair which has the form (~---,Fn). A failed derivation ends in a goal
pair (G,,F,,), where G,~ is not empty and the selected atom from G,, does not unify with the
head of the clause in any clause pair. Let D be a derivation whose initial goal is (Go, F0). The
derived value, de~((G0, F0), D), of D is defined as follows:

• If D is a successful derivation tha:; ends in (~-, F,,), then der((G0, Fo), D) = F,~.

• I f D is failed or infinite, der((Go, Fo),D) = -LL, where I is the region of F0.

By the way, according to the definition of derivation, there are two choices in each step of
de~iving goals;

• Choice of computation rules which select an atom in a goal.

. Choice of a clause whose head and the selected atom are unifiable.

Due to these choices, we may have to consider a large number of derivations. Fortunately,
however, since the independence of computation rules can be proved as in the usual logic pro-
gramming [t0] , we can enumerate derivations by a fixed computation rule. This fact really
reduce the number of derivations. Furthermore, we can generate all derivations systematically
and make up them into a tree, which we call derivation tree.

Let P be a program, ( G , F ) be a goal pair and R be a computation rule. The derivation tree
for (G, F ) via R is defined as follows.

• Each node of the tree is a goal pair.

• The root node is (G, F).

* Let (G', F ' ) be a node of the tree ~nd suppose that A,,, is selected from G' by R. If there
is no clause whose head and A,,, are unifiable, then this node is called failure node and it
has no descendants. On the other hands, this node has a descendant for each clause pal1
(C, f) such that C + and Am are unifiable by an mgu 8. The descendant node is the goal
pair derived by (G', F'), (C, f), and 8.

• If a node is ( ~ , F ' ) , then this node is called succe~ node and it has no descendants.

A derivation tree T for a goal (G, F) contains all possible derived values. Each branch f~om the
root to a leaf node can be regarded as ~ derivation. A branch to a success node is a successful
 106

derivation and a branch to a failure node is a failed derivation. We also define the derived value
der((G,F),T) for a derivation tree T : der((G,F),T) = U{der((G,F),D)l D is a derivation for
(c, F) in T}.

Now we give will our main result which connects the declarative and procedural semantics of
lattice valued programs.

Theorem.
Let A be a ground atom of a program P. If a derivation tree T of a goal (~-- A , I ) ( I ( X ) = X)
via a computation rule is finite,

val(A, riM(P)) = der((÷- A, Z), T).

6 Conclusion

From the programming viewpoint, the optimization process in the sorting example is similar
to what is discussed as program transformation in Clark and Darlington [3] where the initial
program is like:

sort(List, SoriedList) ~ permutation( List, Sor~edList) & ordered( SortedLis~).

However, our algorithm derivation is much easier to understand than theirs, because lattice
programming enables the explicit representation of the sequences and the separation of the
ordering relation from the other aspects of the prograra. In lattice programming, as far as prop-
erly choosing the lattice of sequences and operations on sequences which satisfies requirement
(a), the programmer can concentrate on reducing the computational complexity of permutation
without paying attention to ordered. The system takes care of calculating the ordered solution.
This property makes it easy to evaluate how much the program's computational complexity is
reduced by each programming invention.

From the theoretical viewpoint, the notion of lattice valued programs is extended from a language
P r o B o o l e ([2]) which only considers free Boolean algebras as lattices. It is also a generalization
of works done by van Emden([4]) and Sakakibara([6]) who tried to give meanings to logic pro-
grams with uncertainties. The interval [0,1] of uncertainties is one instance of complete lattices.
O'Keefe([7]) proposed propositional Horn clauses over lattices having finite depths.

We are still in an early stage of the exploration of this new paradigm and much work is to be
done before claiming its usefulness. The following list shows the part of the future research
items.
 107

* Accumulation of algorithm design experiences to refine the theory for the natural repre-
sentation of various algorithms and to develop systematic programming methodologies.

• Design of a programming language which enables the declaration of lattices and functions
of interest and facilitates the verification of their correctness.

* Implementation of a language processor which enables the execution of the primitive lattice
operations in a fixed amount of time and storage.

References

[1] S.Morishita , Lattice-valued programs, TRL Research Report, TR87-0041, 1988

[2] S.Morishita et al. , Symbolical construction of truth value domain for logic program, Proc.
of 4th ICLP, 1987

[3] K.L. Clark and J. Darlington, Algorithm classification through ,ynthesi,, The Computer
Journal, vol.23, 1980

[4] M.H. van Emden, Quantitative ])eduction and it~ fixpoint theory, J. of Logic Programming,
1986

[5] M.H. van Emden and R.A. Kowalski, The semantics o] Logic as a Programming Language,
3.ACM, vol.23, pp733-742, 1976

[6] Y.Sakakibara, On semantics of Logic Programs with Uneertaintie% ICOT Technical Report,
1987

[7] R.A. O'Keefe, Finite fixed-point problems, Proc. of 4th ICLP, MIT Pzess, 1987

[8] E.W. Dijkstra, A note on two problems in connection with graphs Numer.Math.,1, 1959

[9] L.R.Ford et al., Flows in networka Pzinceton Univ.Press, Princeton NJ, 1962

[10] J.W.Lloyd, Foundation, o] logic programming, Springer, 1984

 A Simple Programming System Written in GHC
and Its Reflective Operations
Jiro Tanaka *
Institute for New Generation Computer Technology
4-28, Mita 1-chome, Minato-ku, Tokyo 108, JAPAN

Abstract
A programming system can be defined as an environment where one can input and
execute programs. In this paper, we try to describe a simple programming system written
in GHC. We try to capture the function of the metaeall first. Input/output problems in
GHC are also considered. After describing the shell, we try to assemble these parts into
a simple programming system, and discuss how to add new features such as reflective
operations to the programming system. This paper assumes a basic knowledge of parallel
logic languages such as PARLOG, Concurrent Prolog or GHC.

1. I n t r o d u c t i o n
The goal of Fifth Generation Computer Project exists in connecting knowledge in-
formation processing and parallel computer architecture [Fuchi 86]. Since we would like
to realize everything in logic, it was quite natural for us to try to develop a logic-based
machine and write an operating system in a logic programming language.
In the first stage of our project, ESP [Chikayama 84], which is the object-oriented
dialect of Prolog, has been chosen as our tentative Kernel Language. Logic-based machine
PSI [Taki 84] and logic-based operating system SIMPOS [Takagi 84] have been developed.
However~ SIMPOS was full of side effects. It might be an object-oriented system, but far
from logic programming which we expected. What we have learned is that we need more
elegant mechanisms which realize logic in natural way. Therefore, we next paid attention
to parallel logic languages which are based on and-parallel execution of programs.
Various kinds of parallel logic languages have been proposed so far. PARLOG [Clark 85],
Concurrent Prolog [Shapiro 83] and GHC [Ueda 85] [Tanaka 86] are examples of such lan-
guages. In these languages, we can create processes dynamically and express the synchro-
nization between processes quite easily.
Therefore, it seems to be quite natural to try to describe an operating system in these
languages. In fact, various proposals have been made for system programming from the
very beginning of parallel logic languages [Clark 84] [Clark 85] [Shapiro 83] [Shapiro 84].
PPS (PARLOG Programming System) [Foster 86] [Clark 87] and Logix [Silverman 86]
are the examples of such systems. However, they are kinds of empirical systems which
consist of huge amounts of codes, and their overall structures are not clear enough.
*Authors current address: International Institute (IIAS-SIS), Fujitsu Limited, 1-17-25 Shinkamata,
Ota-ku, Tokyo 144, e-mail: jiro%[email protected]
 109

In this paper, we try to describe a simple programming system written in GHC. Our
objective is not building up a practical programming system like PPS or Logix. Rather,
our interest exists in expressing a simple programming system more systematically. We
would also like to test new features of a programming system such as reflective operations.
We believe that our research will lead to the development of new programming techniques
in parallel logic languages and prove the expressive power of GHC in the long run.
The organization of this paper is as follows. Section 2 describes metacalls which
present the basic functions of our programming system. Section 3 discusses input and
output handling in the programming system. Section 4 describes the UNIX-like shell and
the overall structure of the simple programming system. Section 5 describes reflective
operations in the programming system.

2. Metacalls and their descriptions

User programs can be executed on a programming system. However, the program-
ming system must not fail even if a user program fails. Therefore, we need the metacall
mechanism which protects the programming system from failure [Clark 84] [Foster 87].
This metacall predicate executes the given goal and reports the execution result.

2.1. Various metacalls

Various kinds of metacalls have already been discussed in [Clark 84]. Here, we briefly
review how they work. The simplest metacall is the following single-argument metacall.

call (G)

This metacall simply executes goal G. However, this form of metacall does not help us
much because it is too simple. Therefore, the following two-argument metacall has been
proposed.

call(a,B)
This metacaI1 executes goal G and returns the result by instantiating R to success
when it succeeds and to failure when it fails.
The next extension is the following three-argument metacall, which has slightly been
modified from [Clark 84].

c a l l (G, I n , 0 u t )

Here, In is called input stream and used for communication from the system to the
metacall. Out is called output stream and used for communication from the metacall.
Goal execution can be suspended, resumed, or aborted by instantiating In to [suspl I n ' I ,
[resume I In' ' ] or [abort I In' ]. When the execution of the metacall finishes successfully,
Out is instantiated to [success]. When it fails, Out is instantiated to [ f a i l u r e ( R ) ] ,
where R is instantiated to the message which shows the cause of the failure.

2.2. Metacall and m e t a - i n t e r p r e t e r

The next problem is how to implement these metacalls. For efficiency, they should be
implemented as a built-in predicate. Metacalls are prepared as primitives in [Clark 84].
 110

Whereas, Shapiro takes the meta-interpreter and program transformation approach to

keep flexibility [Silverman 86].
If we disregard efficiency, it is possible to express these m e t a c a l l s as a m e t a - i n t e r p r e t e r .
The original notion of the m e t a - i n t e r p r e t e r comes from the self-description of EVAL in
Lisp. In Prolog, the following self-description is very famous [Bowen 83].

e x e c ( t r u e ) :- !.
exec((P,Q)) :- !, exec(P), exec(Q).
exec(P) :- clause((P:-Body)), exec(Body).
exec(P) :- P.

This meta-interpreter simply executes the goal which is given as the argument of
ezec. You may notice that this meta-interpreter corresponds to the implementation of
the single-argument metacall. The GHC version of this meta-interpreter can be written
as follows:

exec(true) :- true I true.

exec((P,Q)) :- true i exec(P), exec(Q).
exec(P) :- not_sys(P) l
reduce(P,Body), exec(Body).
exec(P) :- sys(P) I P.

This program is almost the same as the Prolog program except that every clause
definition includes the ] operator.
Two-argument metacalls can be written similarly by modifying the single-argument
meta-interpreter.

exec(true,R) :- true I R=success.

exec(false,R) :- true I R=failure.
exec((P,Q),R) :- true l
exec(P,R1), exec(Q,R2),
and_result(R1,R2,R).
exec(P,R) :- not_sys(P) i
reduce(P,Body), exec(Body,R).
exec(P,R) :- sys(P) i sys_exe(P,R).

2.3. Three-argument metacall implementation

It has already been stated that metacalls should be implemented as a built-in predicate
for efficiency. However, the three-argument metacall may be too complex to implement
it in this way. Its specification also needs to be flexible. Therefore, we adopted a kind of
hybrid approach. We split the metacall into two parts, the exec part which realizes the
basic function of metacall and the exec_server part which is in charge of other services.
The exec part should be implemented as a built-in predicate for execution efficiency and
exec_server part should be implemented as a software to keep the system flexible. However,
if we disregard efficiency, the exec can also be programmed as follows:
 111

exec(true,ln,0ut) :- true
0ut=[success].
exec(false(R),In,Out) :- true
0ut=[failure(R)].
exec((A,B),In,0ut) :- true
exec(A,In,Ol), exec(B,In,02),
merge_result(01,02,0ut).
exec(A,In,Out) :- sys(A),var(In)
sys_exe(A,In,0ut).
exec(A,In,Out) :- is_io(A),var(In) i
Out=[A].
exec(A,In,0ut) :- not_sys(A),var(In)
reduce(A,In,Body,0ut,New0ut),
exec(Body,In,NewOut).
exec(A,[suspiIn],0ut) :- true l
wait(A,In,Out).
exec(A,[abortiIn],Out) :- true l
Out=[aborted].

wait(A,[resumeIIn],Out) :- true
exec(A,In,0ut).
wait(A,[abortiIn],0ut) :- true I
Out=[aborted].

The unique features of this exec are In and Out which connect the object level and
the metalevel, Note that I/O is handled as a message to the metalevel. Also, v a t ( I n ) is
the special predicate which checks the absence of messages in the argument variable.
The exec_server part which is in charge of other services can be expressed as follows:

exec_server(State,G,El,[successIE0],In,0ut) :-
var(In) I
0ut=[output([success,goal=S])].
exec_server(State,G,EI,[failure(R)~E0],In,Out) :-
var(In)
EI=[abort],
Out=[output([failure,reason=B])].
exec_server(State,G,El,[GliEO3,In,Out) :-
var(In), is_io(Gl) I
Out=[GliOutl],
exec_server(State,G,EI,EO,In,Outl).
exec_server(State,G,El,[undefined(Gl) IE0],In,0ut) "-
vat(In) i
0ut=[input([tmdefined_goal=Gl,expected_result?S,NG) ioutl],
GI=NG,
exec_server(State,G,EI,EO,In,Outl).
exec_server(State,G,EI,E0,[CIIn],Out) "-
true
 112

control_receiver(C,State,G,EI,EO,In,Out).
This exec_server has six arguments. The first argument shows the internal state of
the metacall. The second argument keeps the initial goal, G, and is used for printing
computation result. The third and fourth arguments are connected to exec. The fifth and
the sixth arguments are used for communication to the system. When the exec_server
receives a success or failure message from exec, it transmits the message to the system by
adding the appropriate message. I/O messages are also forwarded to the system. When
an undefined goal G1 appears in exec, the exec_server sends the message to the user and
urges us to input the new goal, NG, which corresponds to the execution result. When the
exec_server receives a control message from outside, it invokes control_receiver.
exec and exec_server can be connected as follows:

call(G,In,Out):- truei
exec_server(run,G,EI,EO,In,Out),
exec(G,EI,EO).
Note that the pair of exec_server and exec works as a metacall as a whole.

3. Input and output

Handling input and output in logic programming is an important problem. We assume
virtual processes which correspond to the actual devices. We consider that there exists
a single stream which connects these virtual processes to the system. These virtual
processes are always consuming a stream. For output, we send a message in the form
output(Message). The input message has the format i n p u t ( M e s s a g e _ l i s t , X). In this
case, Message_list is printed first, then the user's input is instantiated to X. An other
possibility is to assume two streams which correspond to input and output. However, the
synchronization of input and output becomes difficult in this case.
In our approach~ virtual processes can be created by the create predicate in a program.
This predicate is a special one, which we can execute only once in our program for each
virtual process.

3.1. W i n d o w
The window is usually created on a bitmap display and we can input and output
messages from there. For example, a window is created when create(window,X) is
executed. The input and output to the window are expressed as messages to stream
X. (The actual input is completed when we move the cursor to the window and type
in messages from the keyboard.) The virtual processes which correspond to devices are
deleted by instantiating X to [].

3.2. Keyboard controller

As stated above, input and output are performed at the request of the program.
That is, the system does not accept keyboard input without a request from the program.
Therefore, we need to make a program which always generates this request. The keyboard
controller acts as such a program in a programming system. It can be written as follows:
 113

keyboard(Out,In) :- truel
Out= [ i n p u t ( [©] ,T) 1 0 u t l ] ,
keyboard(T,Outl,In).

keyboard(halt,Out,In) :- trueJ
Out = [ ] ,
In=[].

keyboard(T,Out,In) :- goal_or_command(T)
In=[TIInl],
Out= [input ( [@], TI) IOut 1],
keyboard(T1,0utl,Inl).

Here, the message input ( [©], T) outputs @first, and user's input will be entered into
T. Note that @is used as a prompt for user input.

3.3. Database server

In the programming system, we need a database capability which can add, delete and
check program definitions. The database capability is a kind of I/O in a b r o k e r sense.
Therefore, we imagine a virtual process which corresponds to the database. Operations
to the database can be realized by messages to the virtual process.
In fact, PPS tries to realize the database in this way [Clark 87]. However, every exec
needs to carry the stream to the database in that case. This is very complicated and this
may cause a database access bottleneck. Therefore, we have implemented rib_server using
a side effect as follows:

db_server([add(Code) i I n ] , r e a d y , O u t ) : - t r u e ]
add_definition(Code,Done,Out,Outl),
db_server(In,Done,Outl).
db_server([delete(Name) lIn],ready,Out):-truel
delete_definition(Name,Done,Out,Outl),
db_server(In,Done,Outl).
db_server([definition(Name) IIn],ready,Out):-truel
definition(Name,Done,Out,Outl),
db_server(In,Done,Outl).

The db_server predicate has three arguments. The first is the input from the system.
The second is used to sequentialize database access. The third is the output to the system.

4. Building a p r o g r a m m i n g s y s t e m
We have already discussed the description of metacall and the handling of I/O in the
programming system. The next step is the construction of a programming system~ In
this section, we first describe the shell which plays the central role in the programming
system. Then we try to assemble these parts into a simple programming system.
 114

4.1. Shell
The shell creates the user task or enters the program to the database, depending on
messages from the user. T h e following is a programming example for the shell.

shell ( [], Val,Db, Out) :-true I

Val= [],
Db--[ ] ,
Out= [].
shell([goal(Goal) lln]
Val, Db, Out) :-true I
Val= [record_dict (Goal, NGoal) IVall],
create (Window, WOut),
keyboard (KO, PI),
exec_ server (run, NGoal, EI, EG, PI, PO),
exec (NGo al, EI, EO),
shell(In,Vail ,Db,Out),
merge (K0, PO, WOut).
shell( [db (Message) IIn],
Val, Db, Out) :-true l
Db= [Message IDbl],
shell (In, Val, Dbl, Out).
shell([binding(Message) IIn],
Val, Db, Out) :-true [
Val= [Message IVall],
shell(In, Vai1,Db, Out).

The shell has four arguments; the first is the input stream, the second is the stream
to the variable dictionary, the third is the stream to the database server, and the fourth
is the output stream. This shell is connected to the variable dictionary where the user
can freely define variables and their bindings. This variable dictionary is a kind of user
interface and provides macro facilities. The shell program works as follows:

1. If the input stream is [], it means the end of input. All streams will be closed in
this case.

2. If goal (Goal) is in the input stream, Goal is sent to the variable dictionary. The
variable dictionary checks the bindings of every variable in Goal and creates NGoal
where every variable is bound to the current bindings. A window, a keyboard
controller, an exec_server and an exec are also created.

3. If a message to the database_server or the variable dictionary is received, the message

is sent to the appropriate stream.

T h e function of the variable dictionary is to memorize the value of variables as its

internal state. It works as a kind of user interface.
Figure 1 shows the snapshot where processes are created in accordance with the user
input.
 115

J she 1 1
fl
C exec ) C"exec ) Cexe¢ )
~/ ,~
Co~oo-~)C~oo-@ (proc--s)

000

I-n°o-I
Figure 1 Creation of processes in the shell

Corresponding to the input goal, four processes, exec, exec-server, keyboard controller
and window, are dynamically created.
Since the four processes each have their own window and keyboard, they can run
independently from the shell. The keyboard controller always sends the read request to
the window, and the user can input the control commands to the exec-server from the
window.

4.2. E x a m p l e of a p r o g r a m m i n g s y s t e m
This section shows an example of a programming system by connecting the components
described above.

create_world :- truel
create(window,Out),
keyboard(Out,In),
shell(In,Va,Db,Out2),
vr_dictionary(Va,[],Out3),
db_server(Db,ready,Out4),
merge4(Outl,0ut2,0ut3,0ut4,0ut).
This program can be illustrated as follows:
 116

f'w i n d o w

l }-
Figure 2 A simple programming system

Here, we create the system window, keyboard controUer, shell, variable dictionary and
database server. We connect the outputs of the shell, variable dictionary and database
server to the system window, together with the output of the keyboard controller.
Since the keyboard controller always generates the read request to the system window,
we can input goals from it. Commands to the database server or variable dictionary can
also be entered.
Note that this shows just an example of a programming system. A more sophisticated
programming system such as distributed programming system is also possible.

5. Reflective operations in the programming system

Implementing various reflective operations, such as seen in [Weyhrauch 80] [Smith 84]
[Maes 86] [Watanabe 88], is not too difficult, once we get the meta-interpreters.
We sometimes need to understand the current state of the system. We also need to be
able to modify and return the state to the system. These kinds of reflective capabilities,
seem to be very useful in the operating system. In 3-Lisp [Smith 84], we can easily obtain
the current continuation and environment from the program. Smith used metacircular
interpreters as a mechanism to obtain information from the program.
We extend our meta-interpreter in a similar way to Smith's approach. How do we
extend the meta-interpreter? It depends on what kind of resources we want to control.
We would like to control computation time, memories and processes in the programming
system. Therefore, we introduce a scheduling queue explicitly in our meta-interpreter.
Program continuation was explicit in the metacircular interpreter in 3-Lisp. We have
thought that the scheduling queue acts as a continuation in GHC. We also introduce the
reduction count to control computation resources. We assume that this reduction count
corresponds to the computation time in conventional systems.
The three-argument exec in 2.3 becomes the following six-argument exec when we
perform this modification.

exec(T,T,In,Out,Max/~C,RC) :- true I
Out= [success (reduction.count=RC)] .
 117

exec ( [true [H], T, In, Out, MaxRC, Re) •-

true l
exec (H, T, In, Out, MaxRC, RC).
exec ( [false (R) iH], T, In, Out, MaxRC, RC) "-
true
Out= [failure (R) ] .
exec([A~H] ,T,In,0ut,MaxRC,RC) :-
sys(A) ,var(In) ,MaxRC>RC i
sys_exe (A ,T, NT,RC ,RCl, Out, N0ut),
exec (H, NT, In, N0ut, MaxRC, RC1).
exec([AIH] ,T,In,0ut,MaxRC,RC) :-
is_io(A),var(In) ,MaxRC>RC l
Out= [A lN0ut],
RCI--RC+I,
exec (H, T, In, NOut, MaxRC, RC1).
exec([AIH] ,T,In,Out,MaxRC,RC) :-
not_sys(A) ,var(In),MaxRC>RC I
reduce (A, T, NT, RC ,RCI ,Out, N0ut),
e x e c (H, NT, I n, N0ut, MaxRC, RC I).
exec (H, T, In, 0ut,MaxRC, RC) :-
MaxRC=<RC
Out= [count_over] .
exec(H,T, [Mes~In] ,0ut,MaxRC,RC) :- true
control_exec (Mes, H,T, In, 0ut,MaxRC,RC).
The first two arguments of exec, H and T, express the scheduling queue in Difference
list form. The use of Difference list for expressing scheduling queue was originally invented
by Shapiro [Shapiro 83]. Goals are processed sequentially because we have introduced a
scheduling queue. However, it does not mean that the whole world is sequential. It simply
means that the world inside exec is sequential. The third and the fourth arguments are
the same as before. The fifth argument, MaxRC, shows the limit of the reduction count
allowed in that exec. The sixth argument, RC~ shows the current reduction count.
There is no notion of job priority in this exec. We sometimes need to execute goals
urgently. Therefore, we also introduce the express queue to execute express goals which
have the form, G@exp. This can be realized by adding two more arguments, E H and
ET, which correspond to the express queue, to the six-argument exec. The following
two definitions describe the transition between the six-argument exec and eight-argument
exec.

exec(In, Out, [G©expi H] ,T,RC,MaxRC) :-

var (In) l
exec([G ]ET], ET, In, 0ut, H, T,RC,MaxRC).
exec (ET, ET, In, Out, H, T, RC, MaxRC) :-
var (In) I
exec (In, 0ut, H, T, RC, MaxRC).
If we come across the express goal, we simply call the eight-argument cxec. The eight-
argument exec executes express goals first, and the reduced goals are also entered to the
 118

express queue. If the express queue becomes empty, we simply return to the six-argument
e.~ec.
The next thing is to realize the reflective operations. Here, we consider four kinds of
reflective operations, get_rc, put_rc, get_q, put_q. These can be defined as follows:

exec([get_rc(Max,C) [EH] ,ET,In,Out,H,T,RC,MaxRC) :- t r u e ]

Max := MaxRC, are
C := RC,
RCI := RC+I,
exec (EH, ET, In, Out ,H, T, RCl, MaxRC).
exec([put_rc(C) IEH] ,ET,In,0ut,H,T,RC,MaxRC):- truel
RC1 := RC+I,
exec (EH, ET, In,0ut ,H, T, RCI ,C).
exec ([get_q(NH,NT) ~EH] ,ET,In,Out,H,T,RC,MaxRC) :- true I
RC1 := RC+I,
NH = H,
NT -- T,
exec (EH, ET, In, Out ,H, T,RCI, MaxRC).
exec([put_q(NH,NT) JEH] ,ET,In,Out,H,T,RC,MaxRC) :- truel
RCI := RC+I,
exec (EH, ET, In,Out, NH, NT, RCI, MaxRC).
get_re gets MaxRC and R C from exec. put_rc resets MaxRC of exec to the given
argument, get_q gets the current scheduling queue of exec. put_q resets the current
scheduling queue to the given argument.
Here, we show an example which uses these reflective operations. This example shows
how to define the check_re predicate which checks the current reduction count of the
system and changes it if allowed fewer than 100 reductions.

check_rc :-true [
get_rc (MaxRC ,RC),
RestRC := MaxRC-I%C,
check(MaxRC, RestRC).

check(MaxRC,RestRC) :- lO0>=RestRC I
get_q(H,T),
input ( [reduct ion_ increment, @], AddRC),
NRC :-- MaxRC+AddRC,
put_rc (NRC),
T= [check_rc~exp INT],
put_q (H, NT).
check(MaxRC,RestRC) :- 100<RestRC I
get_q(H,T),
T= [check_rc@exp INT],
put_q(H, NT).
In a sense, these reflective operations are very dangerous because we can easily access
and change the internal state of the system. However, we can say that privileged users
 119

must have these capabilities for advanced system control. Our current interest exists in
examining the reflective operations in a programming system. Thus, we assumed these
it reflective operations in a very primitive manner. The more sophisticated handling of
reflective operations can be seen in may [Tanaka 88].

6. Conclusion
In this paper, we discussed various factors in the programming system and showed
an example of such a system. Reflective operations in the programming system were also
considered.
You may have noticed that these program fragments are quite simple and declarative.
Their simplicity mainly comes from the extensive use of streams, processes and meta-
interpretation techniques. The clear separation of the meta from the object also makes
the system very clean. Note that the program fragments shown here are the extremely
simplified version. The more complete version, workable on DEC 2065, is available from
the author.
We would like to answer the question of why we need to research the programming
system now. Looking back on the history of parallel logic programming, we designed
the language specification first. Then we made interpreters and compilers. Now we have
started working on programming environments and applications for those languages.
It is clear that we can program easily in GHC. Techniques for parallel problem solv-
ing, such as in [Ohwacla 87], are very similar to those of parallel programming systems.
Research in parallel programming systems leads to research in parallel problem solving.

7. Acknowledgments
This research has been carried out as a part of the Fifth Generation Computer Project.
! would like to express my thanks to Yukiko Ohta, Fujitsu Social Science Laboratory, for
her useful comments. I am indebted to her for part of this research. I would also like to
express my thanks to Koichi Furukawa, Deputy Director of ICOT, for his encouragement
and giving me the opportunity to pursue this research.

References
[Bowen 83] D.L. Bowen et al., DECsystem-10 Prolog User's Manual, University of
Edinburgh, 1983

[Chikayama 84] T. Chikayama, Unique Features of ESP, in Proceedings of the Interna-

tional Conference on Fifth Generation Computer Systems 1984, pp.292-
298, ICOT, 1984
[Clark 84] K. Clark and S. Gregory, Notes on Systems Programming in Parlog, in
Proceedings of the International Conference on Fifth Generation Com-
puter Systems 1984, pp.299-306, ICOT, 1984
[Clark 85] K. Clark and S. Gregory, PARLOG, Parallel Programming in Logic, Re-
search Report DOC 84/4, Dept. of Computing, Imperial College of Science
and Technology, Revised 1985
 t20

[Clark 87] K. Clark and I. Foster, A Declarative Environment for Concurrent Logic
Programming, Lecture Notes in Computer Science 250, TAPSOFT'87,
pp.212-242, 1987
[Foster 86] I. Foster; The Parlog Programming System (PPS), Version 0.2, Imperial
College of Science and Technology, 1986
[Foster 87] I. Foster; Logic Operating Systems, Design Issues, in Proceedings of the
Fourth International Conference on Logic Programming, Vol.2, pp.910-
926, MIT Press, May 1987
[Fuchi 86] K. Fuchi and K. Furukawa; The Role of Logic Programming in the Fifth
Generation Computer Project, in Proceedings of the Third International
Conference on Logic Programming, Lecture Notes in Computer Science
225, pp.l-24, Springer Verlag, July 1986
[Maes 86] P. Maes; Reflection in an Object-Oriented Language, in Preprints of the
Workshop on Metalevel Architectures and Reflection, Alghero-Sardinia,
October 1986
[Ohwada 87] H. Ohwada and F. Mizoguchi; Managing Search in ParalM Logic Program-
ming, in Proceedings of the Logic Programming Conference '87, pp.213-
222, ICOT, June 1987
[Shapiro 83] E. Shapiro; A Subset of Concurrent Prolog and Its Interpreter, ICOT
Technical Report, TR-003, 1983
[Shapiro 84] E. Shapiro; Systems programming in Concurrent Prolog, in Conference
Record of the 11th Annum ACM Symposium on Principles of Program-
ruing Languages, pp.93-105, ACM, January 1984

[Silverman 86] W. Silverman, M. Hirsch, A. Houri and E. Shapiro; The Logix System
User Manual, Version 1.21, Weizmann Institute, Israel, July 1986

[Smith 84] B.C. Smith; Reflection and Semantics in Lisp, in Proceedings. of in Con-
ference Record of the 1lth Annual ACM Symposium on Principles of Pro-
gramming Languages, pp.23-35, ACM, January 1984
[Takagi 84] S. Takagi, T Yokoi, S. Uchida, T. Kurokawa, T. Hattori, T. Chikayarna,
K. Sakai and J. Tsuji; Overall Design of SIMPOS, in Proceedings. of the
Second International Logic Programming Conference, UppsMa, Sweden,
pp.1-12, July 1984
[Taki 84] K. TaM, M. Yokota, A. Yamamoto, H. Nishikawa, S. Uchida, H.
Nakashima and A. Mitsuishi; Hardware Design and Implementation of
the Personal Sequential Inference Machine (PSI), in Proceedings of the
International Conference on Fifth Generation Computer Systems 1984,
pp.398-409, ICOT, 1984
 121

[Tanaka 86] J. Tanaka, K. Ueda, T. Miyazaki, A. Takeuchi, Y. Matsumoto and K.

Furukawa; Guarded Horn Clauses and Experiences with Parallel Logic
Programming, in Proceedings. of FJCC '86, ACM, Dallas, Texas~ pp.948-
954, November 1986
[Tanaka 88] J. Tanaka; Meta-interpreters and Reflective Operations in GHC, in Pro-
ceedings of of the International Conference on Fifth Generation Computer
Systems 1988, pp.774-783, ICOT, November 1988
[Ueda 85] K. Ueda; Guarded Horn Clauses, ICOT Technical Report, TR-103, 1985
[Watanabe 88] T. Watanabe and A. Yonezawa; Reflection in an Object-Oriented Con-
current Language, in Proceedings. of ACM Conference on OOPSLA, San
Diego, September 1988
[Weyhrauch 80] R. Weyhrauch, Prolegomena to a Theory of Mechanized Formal Reason-
ing. In Artificial Intelligence 13, pp.133-170, 1980
 Implementation of Full GHC by Communicating
Processes
Minoru TERADA

Department of Mathematical Engineering and Information Physics

Faculty of Engineering, University of Tokyo

Abstract
In this paper an algorithm for interpreting the parallel logic language GHC
(Guarded Horn Clauses) is described. The algorithm interprets full GHC programs
using several communicating processes, which do not share data objects. To control
unification and suspension, an integer (called commit value) is attached to each
external references. An implementation of the algorithm on a concurrent lisp is also
reported. This algorithm is applicable to conventional languages.

1 Introduction
In this paper an algorithm for interpreting the parallel logic language GHC (Guarded
Horn Clauses)Ill is described, and an implementation of the algorithm on a concurrent
lisp mUtilisp[2, 3] is reported.
In GHC, a clause is separated into the guard part and the body part by a commit op-
erator. Several clauses start execution for a goal-literal, and only one of the clauses which
succeeded to solve its guard part gains the right to affect the outer world (the environment
belonging to the caller). This is the fundamental mechanism to achieve synchronization
in GHC, and called commit mechanism. A clause which has not committed yet (not suc-
ceeded to solve its guard part) is not permitted to affect the caller's environment, and if
the clause tries to instantiate an unbound variable belonging to the caller, it is suspended
until the variable is instantiated by another clause.
There exists a subset of GHC, called Flat GHC (or FGHC), which allows only system-
defined predicates (e.g., arithmetic comparison) on the guard part. The execution algo-
rithm for GHC is quite different from that of FGHC, especially in unification. Because
the overhead of GHC is certainly heavy compared to FGHC, FGHC is regarded to be
enough for practice, and most implementations of GHC were those of FGHC[4, 5].
When we first consider the implementation of a parallel logic language by the interpro-
cess communication, we found that full GHC can be implemented without much effort.
So we decided to implement full GHC, and tried to found the basis for the comparison of
their overheads.
Our algorithm has the following characteristics:
 123

1. The execution of GHC program is divided into several processes, and their interac-
tion (such as unification) is achieved by interprocess communication with messages;

2. The sharing of data between processes is inhibited as far as possible;

3. To control unification and suspension, an integer (called commit value) is attached

to every external references;

4. The direction of external reference is constrained from child to parent.

In chapter 2, we describe the restriction for the language. The detail of the algorithm
is described in chapter 3. In chapter 4, our implementation on mUtilisp is reported briefly.
Various problems of the algorithm are discussed in chapter 5.

2 Language specification and restrictions

Our algorithm handles full GHC, and the following three data types are admitted: integer,
symbolic atom and list (or cons).
We posed the following restrictions on the language:
. The execution of a clause is separated into three distinct parts (head, guard and
body) and they are not overlapped, although these parts could be executed in par-
allel according to the language specification. (Of course even the body part cannot
affect the caller's environment until the clause commits.) This restriction is essential
for our algorithm.

2. We support only list (cons) for structured data types. This restriction is not essen-
tial, and the addition of general compound-term is not so difficult to do.

3 Algorithm
3.1 Control
In our algorithm, programs are executed by communicating processes. To execute one
goal-literal, several clauses that match the goal start execution in parallel (called or-
parallel). And each goal-literal included in a clause is executed in parallel (called and-
parallel). Corresponding to those two parallelisms, there are two types of process: OR
and AND. Figure 1 illustrates a simple example of execution.
The followings are the control algorithm for OR and AND:
O R (argument: one goal-literal)
1. Collect clauses that match the argument. If no clause is found, go to step 6.
2. Fork AND as a child process for each clause collected in step 1.
3. Wait for reports of commitment from the children.
4a. If any child reports success, grant the child to commit, and tell other children to
terminate.
 124

a:-b,cle.
a:-dle, a :-b_, , Ic II.. e
b:-e. 1""t .....
d. [! b :- e .I
e.
--[ ell
L'E'E'~ r':Ek3
:-a. -- ~ ~L~-- I~"Ul
i H, i,'

a :- d 1 e .!
Isuc( 'ed I
r d .I [ e .! [ failed I
Figure 1: Example of execution

4b. If all the children report failure, go to step 6.

5. Wait for the completion of the granted child, and report the result to the parent,
and terminate.
6. (Failure) Report failure to the parent and terminate. []
A N D (arguments: one goal-literal, one clause)
1. Unify the goal and the head of the clause.
2. Fork OR as a child process for each goal-literal in the guard part of the clause.
3. Wait for reports from the children.
4a. If any child reports failure, go to step 9.
4b. If all the children report success, report the success of guard to the parent.
5. Wait for the parent to grant the commitment.
6. If the parent grants the commitment, fork OR as a child process for each goal-literal
in the body part of the cla,use.
7. Wait for reports from the children.
8a. If any child reports failure, go to step 9.
8b. If all the children report success, report success to the parent and terminate.
9. (Failure) Tell other children to terminate, report failure to the parent and terminate.
[]
Actually processes perform other works than described above. Those works are initi-
ated by messages as follows:
Instruction of termination Whenever a process receives a message to terminate, it tells its
children to terminate and kill itself.
Inquiry of value of variables and instruction of unification These messages are sent from a
child (or any descendent), and the process should take appropriate action.
 125

3.2 Value
3.2.1 T y p e of values
Variables and conses occurred in a clause belong to the AND process which is executing
the clause. The process is responsible for the values of variables and the elements of
conses. There exist 8 types of value in a cell (figure 2):

(1) Unbound
(2) Unified with another cell (left value)

(2a) with a variable in the same process

(2b) with a variable in another process
(2c) with a car/cdr cell in the same process
(2d) with a car/cdr cell in another process

(3) Already instantiated (right value)

(3a) to a constant (integer, symbolic atom)

(3b) to a cons in the same process
(3c) to a cons in another process

(2b) (2d) (3c)

(1) unbound
(3a) const
remote vat remote car/cdr remote cons

(2a) (2c) (3b)

IXl--4- ? YLZ_.L.J
NN
local var local car/cdr local cons

Figure 2: Type of values in a cell

3.2.2 R e f e r e n c e to a n o t h e r process
Pointers may refer objects in another process (as shown (2b), (2d) and (3c) in last section).
Such pointer (called external reference) is not implemented by direct pointer, in order
to prevent object sharing between processes. Rather, it is represented by a record of the
following four fields:

t y p e The type of referred object, which is one of four types RVAR, RCAR, RCDR and RCONS.
 126

owner The name of owner process

id The name of the variable or the serial number (within the owner process) of the cons.
commit value Explained in 3.3.2.

3.2.3 Direction of pointers

In our algorithm, the direction of external reference is constrained to be from-child-to-
parent. To achieve this, we need some extra operation on unification (see 3.3.4 Copy-up
of cons).
This restriction means that if a process has no child, there exists no external reference
to the process. One of the consequences of this is that when a process is about to
terminate, because the children of the process are all terminated, no object in the process
is necessary any longer. So the entire workspace for the process can be freed, reducing
the cost of garbage collection significantly.

3.2.4 Propagation of values

Unification of two unbound cells creates a directed link between cells. We call the series
of such links variable chain. When a variable is instantiated (by getting unified to a
constant or a cons), the value should be propagated through the variable chain in which
the variable is included. There are several ways to do the propagation:

1. Broadcast to all processes;

2. Sending message to the owner of each variable in the chain;
3. Keep the value in one cell of the chain, and when a variable needs its value, access
the cell by message.

The first has a problem because it needs broadcasting. The second may send a message
to the processes which don't need the value. (For example, a clause which receives an
argument and passes it immediately to the subgoal does not need the value.)
We adopted the third strateg}: By the constraint to the direction of pointers, the
value of a variable chain is held by the variable belonging to the highest process.

3.3 Unification
3.3.1 Difference between FGHC and GHC in unification
GHC and FGHC differ most in unification. See the following GHC program:

[pl] a :-true ] b(X) ....

[p2] b(Y) :-c(Y) t ...
[p3] c(Z) :-true I Z=3.
Now p3 is about to instantiate the variable Z, which is linked to the variable X of pl
via variable chain. Because p2 is not committed yet, this unification must be suspended.
 127

This example means that the information necessary to determine the suspension is
distributed among the hierarchy of processes. Neither the process trying the unification
(p3) nor the owner of the variable (pl) can determine that by itself.
On the other hand, because FGHC does not allow user-defined predicate in the guam
part, there is no child for not-committed process. So the suspension is fully determined
by the state of the process trying the unification. That is, an attempt to instantiate an
unbound external variable in the head or the guam part of a clause causes a suspension.

3.3.2 C o m m i t Value
In order to determine whether a unification causes an instantiation or a suspension, any
external reference (to cell or structure) maintains an integer, called a commit value. A
commit value is the number of processes between the referring process and the referred
process that have not committed yet. (The status of referred process is not counted while
that of referring process is counted.)
In the example in 3.3.1, the variable Z of p3 refers X of pl. The commit value for
this external reference is 1, because p2 has not committed yet. Y of p2 also refers X of
pl, and its commit value is 1. In general, commit values that reference the same cell (or
structure) may differ if referring processes differ.
The commit value makes it very easy to determine the suspension. In a unification,
any instantiation of a cell which is designated by an external reference whose commit
value is not zero must be suspended.
The cell may be designated by a chain of references, not a single direct reference (figure
3). In such cases, the commit value for the cell is the sum of the commit values of the
references.
The commit value for a structure is provided to distribute it to the components. See
the following example:

[pl] a :- b ( X ) I . . .
[p23 b(Y) :- c([Y I nil]) I ...
[p3] c(Z) : - Z = [A I B] I ...

The commit value for A of p3 (to X of pl) is 2 as shown in figure 4.

The calculation of commit value is simple:

1. When a process delivers a term to the child as an argument of the goal literal, unless
the term is not atomic (i.e. not a number nor a symbol), the commit value for the
term is incremented. If the term is local to the delivering process, it should be
transformed into an external reference with commit value 1. (That is, the commit
value for a local reference is 0 logically.)
2. When a process commits, it must decrement the commit values of all external ref-
erences held by the process.

The separate execution of the guard and body (mentioned in chapter 2) plays crucial
role for the calculation of the commit value. An intermediate process (a process waiting
for its children to terminate) never changes its commit status while the execution of the
 128

[p2] [p2] GD

,,,,J I '1
Figure 3: Chain of references Figure 4: Chain through a cons

leaf children. So once a process has received an external reference, the commit value of
the reference is valid until the process terminates.
And the restriction of the direction of external reference (in 3.2.3) is also very impor-
tant for the algorithm. When a process is about to commit, the process has no children.
So the commit values affected by the commitment are only those attached to the external
references held by the process.

3.3.3 E x a m p l e s of unification
Unification is divided into two stages:

FETCHVAR Acquire the actual value of one term

UNIFY Tell the other term to unify with the result of FETCHVAR
If any term includes external reference, those stages may need interprocess commu-
nication, which is implemented by remote procedure calls. The unifying process sends
a message containing the name of the request (FETCHVAR or UNIFY) to the referred
process, and waits for the reply. Receiving the request, the referred process examines the
message and takes one of the following actions:

1. Send reply to the unifying process;

2. Forward the request to another process;
3. Queue the request and defer the reply.

In FETCHVAR stage, 1 or 2 may occur. If the reference is chained, the request is

forwarded. In UNIFY stage, any of the three actions may occur. The reply is SUCCEED
or FAIL.
Now the algorithm for unification is explained using examples.
 129

[p2] I~
c!D

Figure 5: Relationship of variables in [example 1]

[example 1] First an example without suspension is presented.

[pl] a :- true [ B = 3, b(A, B), ...

[p2] b(P, Q) :- true I R = Q, c(P, R),
[p3] c(X, Y) : - t r u e I X = Y.

p3 is trying to unify X with Y. (Figure 5 illustrates the relationship of variables.)

stage 1 (FETCHVAR)In order to acquire the value of X, p3 sends p2 a message.

Because the reference is chained to A of pl, the message is forwarded to pl, and the final
value (A of p1, commit value is 0) is returned to p3. Notice that the commit value is the
sum of the commit values along the forwarding path.

stage 2 (UNIFY) pS tries to unify Y with the result of stage 1, sending a message
to p2. p2 relays the request to pl, in which the chain of reference ends. p1 is requested
to unify the final value of Y (constant 3) with the final value of X <A of pl, commit value
is 0). Because the commit value is 0, the variable is instantiated to the constant 3, and
the reply of SUCCEED is returned to p3.

[example 2] See how suspension occurs during unification. Suppose the p2 of the last
example is not committed yet:

[pl] a : - t r u e [ B = 3, b(A, B),

[p2'] b(P, Q) : - R = Q, c(P, R) I D ~ 4

[p3] c(X, Y) :- true [ X -- Y.

stage 1 (FETCHVAR)The result p3 gets is (A of p1, commit value is 11, because

the reference from P of p2' to A of p! has commit value 1.
 I30

,J

~ ~ <1>
[pl] '* [pl] ~b, " <2>"

'1¥11<1>1 IYI I I --,1Yi1<1>1

[p2] -IY21 > I [p2] "]Y21 <2>1
(a) before (b) after

Figure 6: Copyup of a cons

s t a g e 2 (UNIFY)As in example 1, pl is requested to unify A to constant 3. But

because the commit value for A is not zero, this instantiation must be suspended, pl
stores the request into a queue attached to the variable A, and terminates the service for
the message. So p3 cannot get any reply for its request, and keeps waiting for the reply.
(This realizes the suspension.)
When the variable A is instantiated later, the queue is scanned and suspended unifica-
tions are continued. For example, if A is instantiated to constant 3, SUCCEED message
is sent to p3 as a reply, and p3 completes unification. On the other hand, if A is bound
to another variable V which belongs to an ancestor of pl, contents of the queue of A are
transferred to the variable V, and wait for the instantiation of V.

3.3.4 C o p y - u p o f cons
As mentioned in 3.2.3, all the external references must be from-child-to-parent. When a
pointer is created as a result of unification, its direction must be considered carefully. In
almost all cases this is easy because the direction of a link between two cells is determined
by the level of the cells. But the only exception is the case when a cell is instantiated to
a cons belonging to the lower process.

[pl] a :- t r u e [ b(X).
[p2] b(Y) : - Y I - - I , Y2=2 I Y=KYI I Y2].

This program unifies the variable X of pl and a cons of p2. (Those two terms are
indicated by boldfaced rectangles in figure 6(a).) It is inhibited to make a pointer from
X to the cons, because the pointer is from-paxent-to-child. To solve the problem, a new
cons whose fields are unbound is allocated in p1 as a value of X, and these two eonses are
unified in turn (figure 6(b)).
As a result, the cons of p2 was copied to pl, saving the data which will be necessary
after the termination of the owner process (p2). That may be seen as a local garbage
collection by copying.
 t81

(defunwait-for ((protocols))
(cond ((msgq-lookup p r o t o c o l s ) )
(t (do ((msg (receive)(receive)) (action))
((memq (protocol msg) protocols)
(exit msg))
(setq action (get (protocol msg) 'g:protocol-action))
(cond ((null action) (msgq-inser% msg))
(t (funcall action msg)))))))

(putprop 'TERMINATE (function terminate) 'g:protocol-action)

(putprop 'FETCHVAR (function fvd) 'g:protocol-action)
(putprop 'UNIFY (function ud) 'g:protocol-action)

Figure 7: Function wait-for

4 Implementation
A prototype for the algorithm is implemented on a concurrent lisp mUtilisp, mUtilisp is
an extension of Utilisp[6] to include concurrent processing. It has the following charac-
teristics:

1. Processes are created dynamically (by the function c r e a t e - p r o c e s s ) ;

2. Processes are not allowed to share lisp objects (such as symbols or lists);
3. Processes communicate by sending and receiving messages.

Any lisp object can be used as a message. But it is copied in order to avoid sharing.
Message is sent by the function send with an argument specifying the receiver process.
No broadcast is available.
To receive a message, the function r e c e i v e is used with an optional argument desig-
nating the sender. If there is no messages to receive, the calling process is blocked until
a message arrives.
Our algorithm requires to receive only messages of specified types, such as the result
of the guard part. And furthermore, the message for termination or unification must be
received and processed at any time. Because the language does not support the selection of
message by its contents, function w a i t - f o r was defined (figure 7). The function receives
any message to the process, inspecting the type (called protocol) of the message. If a
message of specified protocol arrives~ the function returns it. If the protocol requires
instant action (e.g., termination), the function internally calls the appropriate service
routine.
The code contains about 700 lines in the form of lisp source. Although the current
implementation lacks arithmetic operations and input-output predicates, it processes fun-
damental GHC programs. The sample session is shown in figure 8.
 132

;...... Example "union".

; mem(A,[A{R]) :- true I true.

; mem(A,[B{R]) :- A!=B I mem(A,R).
(putprop 'mem '(
((A (A . R)) () ())
((A (B . R)) ((!= A B)) ((mem A R)))) 'two)

; notmem(A,[]) : - true ] true.

; notmem(A,[BIR]) : - A!=B ~ notmem(A,R).
(putprop 'notmem '(
((A ()) 0 ())
((A (B . R)) ((!= A B)) ((notmem A R)))) 'two)

; union([],X,Y) :- true [ X=Y.

; union([HJT],X,Y) :-mem(H,X) J union(T,X,Y).
; union([HlT],X,Y) :-notmem(H,X) I union(T,X,Y2),Y=[H[Y2].
(putprop 'union ' (
( ( 0 x Y) 0 ((= x Y)))
(((H . T) X Y) ((mem H X)) ((union T X Y)))
(((H . T) X Y) ((notmem H X)) ((union T X Y2)(= Y (H . Y2))))
) 'three)

> (g:top ' ( ( u n i o n (1 2) (3 2) Y)))

"orO" (union (rcons "and" I I) (rcons "and" 3 i) (rvar "and" Y I))
"andO0" (union (rcons "and" I I) (rcons "and" 3 I) (rvar "and" Y 1))
"andOl" (union (rcons "and" I i) (rcons "and" 3 i) (rvar "and" Y I))
"and02" (union (rcons "and" I 1) (rcons "and" 3 I) (rvar "and" Y I))
"orOlO" (mere (rcar "and" I 2) (rcons "and" 3 2))
"or020" (notmem (rcar "and" 1 2) (rcons "and" 3 2)~
.... (62 lines) ....
"or021" SUCCEED
"and02" SUCCEED
"orO" SUCCEED
((union (icons ? ? 1 nil nil) (icons ? ? 3 nil nil) (ivar ? Y nil)))
> result-vars
((Y (1 3 2)))

F i g u r e 8: S a m p l e session
 133

[pl]

[p2]
Figure 9: Copyup of multiple conses

5 Discussion
5.1 M a i n t e n a n c e of C o m m i t V a l u e s
As mentioned in 3.3.2, when a clause commits it must decrement all commit values held
by the clause. There will be two ways to implement that:

1. Actually decrement all commit values scanning all external references;

2. Only set a flag at the time of commitment, but after that when an external reference
is sent to another process its commit value is decremented 'on demand'.

The former seems to be too naive, but it is not so bad because external references at.
the time of commitment are usually only those given by the parent as the arguments of
the goal.
On the other hand, the overhead of the latter will be heavy because there may exist
many requests of FETCHVAR or UNIFY from child processes.
(Certainly pathological cases exist; the following program is one of such cases.

[pl] a :-b([1, 2, 3 . . . . ] ) .
[p2] b(X) : - c ( X , Y) I . . .
[p3] c(L, A) : - L = [] I A = [ ] .
[p4] c(L, A) : - L - - [H I T] A = [L] A l l , ¢(T, A1).

When p2 commits, its variable Y holds a list which is composed from the list of p1. Each
cons in the list contains an external reference to the cons cell of pl. Notice that although
X of p2 points the list of pl, it causes only a single external refeience.)

5.2 Overhead of Copy-up conses

Although the explanation of copy-up ~n 3.3.4 is focused to single cons, several conses may
have to get copied up if the first cons includes a reference to another cons belonging to
a lower process (figure 9). Those conses are necessary for the upper process into which
they are copied, so the copy-up is inevitable anyway.
But in some situation, unnecessary intermediate copy-up may occur. See the following
example:
 134

[pl] a :-make_list(X, I00).

[p2] make_list(X, N) :- N -- 0 I X = nil.
[p3] make_list(X, N) : - N > 0 ] X = [N I XI], make_list(Xl, N-I).
The order of the execution of the two body literals in p3 is important. If the second
literal (make_list(Xl, N-I)) completes first, a list of length N - 1 is copied up to the
process. After that the first literal of p3 unifies X to a cons, resulting a copy-up of the
cons. But the copy-up is followed by N - 1 copy-ups of the list cells already copied to
the process. Thus, in this situation O(N 2) copy-ups are necessary.
On the other hand, if the first literal (X = [N I Xl]) completes first, the result of the
second literal (a list of length N - 1) is copied up directly to pl, the target process. In
this situation, only O(N) copy-ups are necessary.
Probably this problem should be solved as a matter of programming style. One of the
remedy for that is to execute unifications (e.g., the first literal) prior to other literals.

5.3 Optimization
The current version of the algorithm lacks optimizations. One of the most necessary op-
timizations is the tail recursion optimization, or more generally the tail call optimization.
It saves the number of processes by not forking a subprocess for the last goal-literal
of a clause. The parent process itself executes the literal. But in our algorithm it is not
easy to implement this because even for the last literal there may be several clauses to be
executed in parallel (OR parallelism).
One of the solutions is to merge the child process for the last literal when it is about
to commit, because at the commitment only one child process can survive. The rest of
its clause (i.e. body part) is sent to the parent along with its local variables and conses.
Several problems are left such as the conflict of variable names, the overhead of process
merging.

5.4 Deadlock
Various program errors are often detected as a deadlock in GHC, so its detection is
necessary for a practical implementation. Although current implementation lacks the
detection, it will be not so hard to implement it.
Because the suspension of the process is implemented using the blocking by receiving
messages, the process never becomes runnable again until the suspension is released. So
if there exists no runnable process, it is considered that a deadlock has occurred.
If the suspension were implemented by busy wait (the suspended process itself exam-
ines the variable periodically), more complex method would be necessary.

5.5 Possibility of implementation

Although the prototype implementation was on a lisp, it is possible to use other lan-
guages. Because the algorithm does not need garbage collection, it may be implemented
on a conventional language if it supports multiprocessing and basic communication. For
example, the multiprocess package on the unix C language[7] can be used for that.
 135

It will be a good property for multiprocessors that no data is shared between processes.
But the implementation on multiprocessors is not investigated yet.

6 Conclusions
An algorithm for interpreting full GHC program is presented, and a prototype implemen-
tation of the algorithm on a concurrent lisp is reported.
The algorithm can be implemented on conventional languages, and has features suit-
able for multiprocessors.

A cknowledgement s
The author would like to thank Professor Eiiti Wada for giving us the opportunity to do
this work and Dr. Hideya Iwasaki, Mr. Tetsuro Tanaka and Mr. Koichi Konishi for their
helpful discussions and useful advices.

References
[1] K. Ueda: Guarded Horn Clauses, Logic Programming '85, E. Wada (ed.), Lecture
Notes in Computer Science 221, Springer-verlag (1986).
[2] H. Iwasaki: Programming and Implementation of a Multi-processing Lisp (in
Japanese), Trans. IPS Japan, Vol. 28 No. 5 (1987).
[3] H. Iwasaki: mUtilisp Manual, METR 88-11, Department of Mathematical Engeneering
and Information Physics, Faculty of Engeneering, University of Tokyo (1988).
[4] T. Miyazaki and K. Taki: Flat GHC Implementation on Multi-PSI (in Japanese), The
Login Programming Conference '86 (1986).
[5] K. Fujimura, M. Kurihara and I. Kaji: An Implementation of GHC compiler on the
LISP Systems (in Japanese), Trans. IPS Japan, Vol. 28 No. 7 (1987).
[6] T. Chikayama: Implementation of the Utilisp System (in Japanese), Trans. IPS Japan,
Vol. 24 No. 5 (1983).
[7] Y. Tada and M. Terada: Primitives for Testing Environment of Concurrent Processes
(in Japanese), IPS Japan WGOS, No. 33-2 (1986).
[8] M. Terada, H. Iwasaki, T. Tanaka, K. Konishi, K. Shirakawa and E. Wada: Imple-
mentation of GHC on the concurrent lisp (in Japanese), IPS Japan WGSYM, No. 44-3
(1988).
 Inference Methods and Semantics
on Or-type Knowledge Bases

Hiroshi SAKAI
Information Science Center, Kyushu Institute of Technology
680, Kawazu, Iizuka, Fukuoka 820, Japan

ABSTRACT
Or-type knowledge is information which has some disjunctive conclusions,
for example, (AvB)<--C. We often face with this type of knowledge in our everyday
lives. There are several problems to deal with such type of knowledge in computer
systems. In general, it is difficult to represent the or-type knowledge in Horn logic.
In this paper, we extend first order formulas and define the interpretations for them

in a three valued logic. According to the discussion, a Horn clause is also extended
so as to deal with the or-type knowledge. We show a fixpoint theorem, the soundness

and completeness of a resolution.

1. Introduction
Let us consider the following example in order to clarify the problem of or-
type knowledge. In reality, the following rulel is not a Horn clause, and this kind of

knowledge are rarely dealt with in the usual logic programming.

( r u l e l ) place(X, k i t a k y u s h u city)Vptace(X, iizuka city) :- staff(X, kit)
( rule2 ) address(X, fukuoka) :- place(X, fukuoka city)V
place(X, k i t a k y u s h u city)vplace(X, iizuka_city)
( factl ) staff(taro, kit)
First rule means that if a person is on the staff of KIT ( Kyushu Institute of
Technology ), the place of his employment is in either Kitakyushu City or Iizuka
City. We call this kinds of knowledge or-type knowledge. The second rule means
 137

that if the place of his employment is in Fukuoka City or Kitakyushu City or Iizuka
City, his address is in Fukuoka Prefecture. The fact1 means Taro is on the staff of
KIT.
First we consider an aspect of logical semantics of our problem. In the above
example,
( fact2 ) place(taro, kitakyushu_city)vplace(taro, iizuka_city)
is logically true by the rulel and the factl, and
( fact3 ) address(taro, fukuoka)
is also logically true by the rule2 and the fact2.
Now consider the aspect of the formal inference on the same example. We can
formally infer fact2 by the r u l e l and the factl. However, neither
place(taro, kitakyushu city)
nor

place(taro, iizuka_city)
can formally be inferred. In this case, using the closed world assumption (CWA) [1]
would lead to a contradiction. Furthermore, in most logic programming languages,
the rule2 means that one of the following clauses holds:
address(X, fukuoka) :- place(X, fukuoka_city)
address(X, fukuoka) :- place(X, kitakyushu_city)
address(X, fukuoka) :- place(X, iizuka city)
Hence, formally fact3 can not be inferred. Some logically true facts can not be
formally inferred in this example. The logically true fact3 can not be proved by
using the SLD-resolution [2], either.
The importance of the or-type knowledge and the hardness of handling it
have been pointed out in [3]. Loveland [4] has also studied the related problems, but
has not discussed the semantics of his system. Lloyd [5] has studied a system which
allows any first order formula in the body of program clauses. However, he has not
discussed the head of the program clause. Incomplete information has also been
studied in [6,7,8], but these discussions are based on relational database theory, and
 138

the incompleteness of facts is only d e a l t with. We deal w i t h n o t only t h e

incompleteness of facts but also the incompleteness of rules. In this paper, we study
and develop a system in which or-type knowledge can easily be represented, and

discuss semantics of the system.

2. D e c l a r a t i v e S e m a n t i c s of O r - t y p e K n o w l e d g e B a s e s
In this section, we extend first order formulas, and give the interpretations
for the formulas. According to the extended formulas, an or-type knowledge base is

formally defined in Section 2.3.

2.1. E x t e n d e d F o r m u l a s a n d T h e i r I n t e r p r e t a t i o n s
DEFINITION 1. An alphabet consists of the following classes of symbols:

(1) variables
(2) constants

(3) predicates
(4) connectives : "V", "A", "-1 ", "'~-" and "~"

(5) quantifiers : V ( the universal quantifier ) and 3 ( the existential quantifier )

(6) punctuation symbols : "(", ")", ". " a n d ","

In this paper, we do not consider function symbols for simplicity, so a term is a

variable or a constant.
DEFINITION 2. If p is an n-ary predicate symbol and tl ..... tn are terms, then
p(tl,"',tn) is called an atomic formula ( or an atom ). An atom that contains no
variables is called a ground atom.
Now, we introduce a set of formulas called elementary formulas.
DEFINITION 3. An elementary formula is defined inductively as follows:
(1) An atom is an elementary formula.
(2) If PI,"" ,Pm are distinct finite atoms, then P1;P2;"" ;Pm is an elementary
formula.
 139

We denote the set of all elementary formulas by EF, and call an elementary

formula which does not contain variables a ground elementary formula. The set of

all atoms which are used to construct an e l e m e n t a r y formula P is denoted by

Atom(P). For example,

Atom(P1;P2;"" ;Pro)--{ P1, P2,"" , Pm }.
Intuitively a ground elementary formula P is True if at least one element of Atom(P)

is True; however it is sometimes impossible to decide which specific elements of

Atom(P) are True.
We extend formulas by the elementary formulas.

DEFINITION 4. An extended formula is defined inductively as follows:

(1) An elementary formula is an extended formula.

(2) I f P and Q are extended formulas, then so are (~P), (PAQ), (PVQ), (P--~Q),
(P~Q).
(3) I f P is an extended formula and x is a variable of P, then (VxP) and (3xP) are
extended formulas.

First order formula is a special extended formula, and in the rest of this paper

we use a "formula" instead of "extended formula". A formula which does not con-

tain free variables is called a closed formula. A formula with free variables is as-

sumed to be universally quantified, and thus we always consider only closed formu-
las.

We define an interpretation for ground elementary formulas, and give the

interpretations for formulas.

DEFINITION 5. For any domain D and a set of elementary formutas EF,

Base(D) ={ Q I Q is a ground elementary formula constructed by

substituting an element of D for any argument of an
elementary formula P( ~EF ) }.
DEFINITION 6. For any domain D, a mapping from Base(D) to { T(true), F(false),
±(undefined) } is called a relation on the Base(D).
 t40

PROPOSITION 1. An a s s i g n m e n t ~/, which depends on I ( E2Base(D) ) and satisfies

the following conditions, is a relation on the Base(D).
(1) ~ / ( P ) = T ( P E I )
(2) ~/(P) = T (Atom(Q)C_Atom(P) for some Q ( E l ) )

(3) ~/(P) = F ( P is a ground atom, and PC~Atom(Q) for any Q( EI ) )

(4) ~ / ( P ) = F ( ~j(R)=Ffor any ground atomR(EAtom(P) )
(5) ~I(P) =_L (Atom(P)C_Atom(Q) for some Q( ~I ) and Atom(R)~=Atom(P) for any
R(El) )
(6) ~/(P) = A_ ( otherwise )
[Sketch of Proof] For any P(EBase(D) ), there are two cases to be considered.

(1) There exists Q( EI ) such thatAtom(P)NAtom(Q)~ 0.

(2) There exists no Q ( E I ) such t h a t Atom(P)NAtom(Q)~ 0. !
A c c o r d i n g to t h i s p r o p o s i t i o n , we d e f i n e i n t e r p r e t a t i o n s for g r o u n d

e l e m e n t a r y formulas.
DEFINITION 7. An interpretation forground elementary formulas M is a quadruple

M = < D , F, I, ~I>,
where
D is a domain,
F is an a s s i g n m e n t from a constant to an element of D,

I is a subset of Base(D),
(4 is a relation defined in Proposition 1.
Since a relation ~ is u n i q u e l y defined by I, from now on we identify I with ~/.
DEFINITION 8. Let M be an interpretation for ground elementary formulas. For
the r e l a t i o n / i n M, Ker(D is as follows:
Ker(I) = { A EI 1there is no B( EI ) s.t. Atom(B)cAtom(A) }
Ker(M) is also defined as follows:
Ker(M) = < D, F, Ker(I), ~er(I) >
By Proposition 1 and Ker operation,
{ AEBase(D) IA is assigned T by the relation Ker(I) }
 141

= { A (Base(D) IA is assigned T by the relation I }

holds. Therefore, we call Ker(I) non-redundant relation with respect to the assign-

m e n t value true, and call Ker(M) non-redundant interpretation.

EXAMPLE 1. For simplicity, take propositional variables gl, g2 and g3, and any
domain D. In this case,

Base(D)={gl, g2, d3, gl;g2, g2;g3, g3;gl, gl;g2;g3}.

Let a relation I be {gl, gl;g2 }, then

ger(1) : { gl }, ~,I(gl) : ~I(gl;g2) : ~I(gl :g3) :~I(gl;g2;g3): T,

~/(g3) = F , ~/(g2) = ~/(g2;g3) = _l_.
DEFINITION 9. A mapping p is defined by the following tables and equations:

-~P PAQ T F _k
T F T T F _I_
F T F F F F
i i 'i i r i

~Q T F ± !P~Q T F ± PoQ T F ±
T TTT iT TF± T TF±
F T E l !F TTT F FT±
± T ± ± ± T ± ± ± ± ± ±

VxP(x) = AdED P(d), 3xP(x) = Vd~D P(d).

In the above tables, the first lines mean the truth values of Q and, the first columns
mean the truth values of P.

The logic defined by the above truth table is generally called Kleene's three
valued logic [9].
DEFINITION 10. An interpretation for formulas N is a quintuple
N=<D,F,I, 6j, p > ,
where D, F, I and ~/are the same definitions as in the interpretation M, and p is the
mapping in Definition 9.
 142

From now on, we also use "interpretation" for "interpretation for formulas".

Ker operation for an interpretation N is the same as in Definition 8. We show a

proposition for the subsequent discussion.

PROPOSITION 2. By A =B we denote t h a t A and B have the same t r u t h table. The

following relations hold:

(1) (P<---Q)=(~PVQ)
(2) ( ~(PVQ))=('~PA ~Q)

(3) (~(P/kQ))=('~PV~Q)
(4) ((PvQ)AR) = ((PAR)v(QAR))

(5) ((PAQ)vR) = ((PvR)A(QvR))

2.2. M o d e l T h e o r y o f E x t e n d e d F o r m u l a s

In this section, we discuss the model theory of formulas.

DEFINITION 11. Let P be a closed formula.

(1) An interpretation which assigns T to the formula P is called apositive model

of P.
(2) An interpretation which assigns F to the formula P is called a negative model

of P.
(3) An interpretation which assigns _k to the formula P is called an incomplete

model of P.
For a n y formula P and a n y interpretation N, if the relation I in N does not

have enough information to interpret P, t h e n the interpretation N becomes an

incomplete model for formula P.
EXAMPLE 2. U n d e r the same conditions as in Example 1, put a formula P=gl;g2.
Then, P is a closed formula. An interpretation with a relation { gl;g2;g3 } is an

incomplete model of P. An interpretation with a relation { gl } or { gl, g3 } is a

positive model of P.
DEFINITION 12. Let S be a finite set of closed formulas.
(1) An interpretation which assigns T for any formula in S is called apositive
 143

model orS.
(2) An interpretation which assigns F for some formulas in S is called a negative

model of S.
(3) Another interpretation, except (1) and (2), is called an incomplete model of S.

PROPOSITION 3. Let S be a finite set of closed formulas.

(1) An interpretation N is a positive model if and only i f N is a positive model of

Ap~sP.
(2) An interpretation N is a negative model if and only if N is a negative model

of Ap(sP.
(3) An interpretation N is an incomplete model if and only if N is an incomplete

model of Ap~ 8P.

DEFINITION 13. Let S be a finite set of closed formulas.

(1) We say S is satisfiable if there exists a positive model of S, especially we say
S is valid if there is no negative model of S.

(2) We say S is unsatisfiable, if there is no positive model orS.

EXAMPLE 3. Under the same conditions as in Example 1, let a formula P be

( g l V ~ g l ) . P is a closed formula. Any interpretation which assigns T or F to the

propositional variable gl is a positive model of P. An interpretation which assigns _L

to gl is an incomplete model of P, but there exists no negative model of P. Namely,

P is valid.

Now we define two types of logical consequences.

DEFINITION 14. Let S be a finite set of closed formula and P be any closed formula.
(1) We say P is a [] -logical consequence orS if, any positive model for S is also a
positive model for P.

(2) We say P is a ~-logical consequence of S if, there exists a positive model for S
and P.

The definition of a []-logical consequence is the same as that normal logical

consequence. The []-logical consequence defines some formulas which strictly hold
 144

under the or-type knowledge. The <~-logical consequence defines some formulas,
which may be true in some cases.
{ P ]P is a [] -logical consequence ofS } __ { P I P is a ~b_logical consequence ofS }

is obvious from the definition.

EXAMPLE 4. Under the same conditions as in Example 1, let S be {gl;g2 }. Then,
{ P [ P is a D -logical consequence of S }= {gl;g2, g l ;g2;g3 },

{P IP is a ~-logical consequence ofS }={gl, g2, gl;g2, gl;g3, g2;g3, gl;g2;g3}.

Now we discuss interpretations with a fixed domain and a fixed assignment

for constants.
DEFINITION 15. Let N and N' be interpretations which have the same domain and

the same a s s i g n m e n t for constants. Let I a n d /' be r e l a t i o n s in N a n d N',

respectively. N' is called an extended interpretation of N and denoted by N ' ~ N if the

following two conditions are satisfied.

(1) For any P'( E/" ), there exists P ( E I ) s.t. (O¢)Atom(P') C_Atom(P).
(2) For a n y P ( E I ) , there exists P'( E/' ) s.t. (O~:)Atom(P') C Atom(P).
Intuitively the knowledge in N' is more detailed than that in N.
D E F I N I T I O N 16. Suppose N and N' are two interpretations with the same domain

and the same assignment for constants. Let I a n d / ' be a relation in N and N', respec-
tively. We say N' is included by N if/'C_I holds, and denote it by N'CN.
EXAMPLE 5. Under the same conditions as in Example 1, take four interpretations

N1, N2, N3, N4 with the following relations, respectively,

I1 -" { g l }, I2 --'--{ gl, g3 }, 13 = { gl;g2, g2;g3 }, 14 = {gl;g2;g3 }.

Then,
NI_-<N4, N2 ~ N3 ~ N4 and N1 C N2
hold, but
N I ~ N2, N2CN1 and N2C_N3
do not hold.
P R O P O S I T I O N 4. Let S be a finite set of closed formulas.
 145

(1) If N is a positive model of S, then any extended interpretation of N is a

positive model orS.

(2) If N is a negative model of S, then any extended interpretation of N is a

negative model of S.
[Sketch of Proof] By Proposition 2, any closed formulas can equivalently be trans-

formed to a disjunctive normal form. !

According to Proposition 4, it is convenient to deal with a set of positive

models which have more incomplete relations.

DEFINITION 17. Let S be a set of closed formulas, D be a domain and F be an

assignment for constants. We define a Maximum model Maxm(S,D,F) and a

Maximum relation Maxr(S,D,F) as follows:

Maxm(S~DJV) = { Ker(N) IKer(N) is a positive model of S on D with
the assignment F, and there is no positive model Ker(N')

( ~ Ker(N)) for S such that Ker(N) ~ Ker(N') }.

Maxr(S,D,F) = { I~ 2Base(D) t I is a relation of a positive model in Maxm(S,D,F) }.
PROPOSITION 5. Let P be a closed formula, D be a domain and F be an assignment

for constants. Then the following two propositions are equivalent.

(1) Any element in Maxm(S,D,F) assigns T to a closed formula P.

(2) Any positive model N ( = <D, F,/, 4, P > ) for S assigns T to a closed formula
P.

[Proof] Trivial by Proposition 4 and Definition 17. ,

We have only to consider the set Maxm(S,D,F) for the decision of a []-logical
consequence by Proposition 5 and Definition 14.

2.3. S e m a n t i c s o f O r - t y p e K n o w l e d g e B a s e s
Now we discuss a class of closed formulas.

DEFINITION 18. Or-type knowledge base is a finite set of closed formulas made by
substituting an elementary formula for an atom in Horn clauses.
 146

Clearly, a set of Horn clauses is an or-type knowledge base. The following

incomplete knowledge can be represented in the or-type k n o w l e d g e base as

C e--A1;A2;A3, B1;B2;B3.
If at least one element of A1, A2 and A3 hold, and
if at least one element orB1, B2 and B3 hold, then C holds.
PROPOSITION 6. Let S be an or-type knowledge base and suppose S has a positive
model. Then S has a positive model with the Herbrand universe US as its domain
and an assignment F for constants which assigns any constant to itself.

[Proof] L e t N be a positive model of S and I b e its relation, and let

P = { A ~Base(Us) I A is assigned T with respect to I }.
An interpretation N( = < US, F, Y, ~4, P >) is clearly a positive model. !
According to this proposition, we define an interpretations for an or-type

knowledge base.
DEFINITION 19. An interpretation N for an or-type knowledge base S is a quintuple

N = < Us, F,I,~4, p > ,

where
US is the Herbrand universe of S,
F is a constant assignment which assigns any constant to itself,
I, ~, p are the same as in Definition 10.
PROPOSITION 7. Let S be an or-type knowledge base, and
N = <Us, F , I , ~ , p >
be an interpretation for S. Then, an interpretation with a relation

NI(Maxr(S,U~,F) I
is a non-redundant positive model of S.
We denote the relation( = nI(Maxr(S,Us,I01~ by RS, and say the interpretation
NS( = < US, F, RS, ~Rs, P >) a least positive model orS.
Now we show the relation between RS and the [] -logical consequence.
THEOREM 8. Let S be an or-type knowledge base. Then
RS =Ker( { AeBase(Us) I A is a 5 -logical consequence o r s } ).
 147

[Proof]

A EKer( { A EBase(Us) IA is a [] -]ogical consequence of S } )

¢~ A is assigned T b y any relation IEMaxr(S, Us,F). ( by Proposition 5 )
ca there exists Q( EI ) s.t. Atom(Q)cAtom(A) for any IEMaxr(S,Us,F).

¢~ there exists Q(EI) s.t. Atom(Q)=Atom(A) for any I(Maxr(S,Us,F).

( by Ker operation )
Therefore, A = QEI for any Maxr(S,Us~. !
We define two mappings.
DEFINITION 20. Let S be an or-type knowledge base, and US be a H e r b r a n d
universe of S. We define two mappings [] TS and (~ TS from 2Base(vS) to 28as~vS) as
follows:

(1) [ ] T s ( / ) = { A EBase( Us) I

(a) A<--P1,P2,"',Pn is a ground instance [2] of a formula in S,
(b) For any Pi (1_-<i ~ n), there exists Q(El) s.t. Atom(Q)C_Atom(Pi),
(c) For a n y R ( E I ) , Atom(R)~=Atom(A) }.

(2) ~Ts(I)= {AEBase(Us) I

(a) A<--P1,P2,"",Pn is a ground instance of a formula in S,
(b) For any Pi (1~ i =
< n), there exists Q(EI) s.t. Atom(Q)nAtom(Pi)~ 0,
(c) For a n y R ( E I ) , Atom(R)~Atom(A) }.

PROPOSITION 9. Two mappings [~]T8 and ~ T 8 are continuous, and there exist
least fixpoints of them. Furthermore,

Ifp([~3TS) = D T8 ~ ~,
Ifp(C>Ts)=~Ts ~ co,
ifp(5 Ts)Clfp( ~ T8)
hold.
[ Proof ] Similar to Proposition 5.1 in [2]. ,

PROPOSITION 10. Let S be an or-type knowledge base. An i n t e r p r e t a t i o n

N( = < US, F, I, ~, p > ) for S satisfies IEMaxr(S,Us,F) if and only if [] TS(I)C_I.
[Proof] Similar to Proposition 6.4 ir.~ [2]. !
 148

We show that NS is characterized by the least fixpoint of [] TS.

THEOREM 11. Let S be an or-type knowledge base, and US be a Herbrand universe
of S.
(1) RS
= Ker({A~Base(Us) IA is a [] -logical consequence ofS })

= Ifp(• TS)
= • T s ~ co.
(2) { A EBase(Us) IA is assigned T by the relation Ifp( ~ 7"8)}
= {A~Base(Us) IA is assigned Tby the relation ~ T s t co}
C {A(Base(Us) IA is a ~-logical consequence ofS }

Base(Us)
/ ~<~-logicalconsequence~

Fig.1 Relationship between rules.

In Fig. 1, (Rs)* and (<> TS 1' co)* imply the ground elementary formula assigned T by
the relation RS, and ~ TS T co, respectively.
[Proof]
(1) By Theorem 8,
RS =Ker( { AEBase(Us) IA is a []-logical consequence ofS } ).
Then, by Proposition 9,
Ifp([]Ts)= []Ts T co, lfp(OTs)= (~Ts ~ co, Ifp([3Ts)C_Ifp(~Ts)
R S -- nI(Maxr(S,UsgF) [

=glb{ IE 2Ba~(US)I [] TS(I)C_I} ( by Proposition 11 )

 149

= lfp(N TS) ( by Proposition 5.1 in [2] ).

(2) is trivial. I

3. Procedural Semantics of Or-type Knowledge Bases

This section studies the procedural semantics of or-type knowledge bases. We

omit some definitions, which depend on [2].

3.1. Two Resolutions

We discuss an aspect of proof theory of the or-type knowledge bases. First, we
extend the SLD-resolution.

DEFINITION 21. T a k e a goal G and a rule C,

G: <-P1,P2,"" ,Pn,

C: Q <-R1,R2,"',Rm,

and let CR be a computation rule [2]. We define a D-resolution and a <>-resolution

as follows:

(1) If there exists a substitution 0 which satisfies (a) and (b):

(a) PI is the selected atom given by the computation rule CR,

(b) Atom(QO) c Atom(PlO),

t h e n we derive a new goal G'

G' : <--(P1," ",P I-t ,R1,R 2, " ",R m,P t + I,"" ,P n)O.

G' is a []-resolvent of G and C. We say 0 a []-unifier, and the condition

(b) specifies a [] -unification.

(2) If there exists a substitution O which satisfies (a) and (b):
(a) Pl is the selected atom given by the computation rule CR,
(b) Atom(QO) n Atom(PlO) ¢ ~,
t h e n we derive a new goal G'

G': <--(P1,"" ,PId,R1,R 2,"" ,Rm,PI + 1,"" ,Pn)O.

G' is a <~-resolvent of G and C. We say 0 a <)-unifier, and the condition (b)
specifies a ~-unification.
 150

DEFINITION 22. Let S be an or-type knowledge base, G be a goal and CR be a

computation rule.
(1) A [] -refutation of SU{G} via CR is a finite sequence of []-resolutions which
has the empty clause as the last goal in the sequence.
(2) A ~-refutation of SU{G} via CR is a finite sequence of [] -resolutions or O-
resolutions which has the empty clause as the last goal in the sequence and

at least one ~-resolution is in the sequence.

DEFINITION23. Let S be an or-type knowledge base, US be the Herbrand universe.
(1) A [] -success set is the set of all ground elementary formula A in Base(Us)
such that SU{~-A} has a [] -refutation.
(2) A @-success set is the set of all ground elementary formula A in Base(Us)

such that SU{,--A} has a ~-refutation.

3.2. S o u n d n e s s of t h e [ ] - r e s o l u t i o n
We show the soundness of the [] -resolution in this section.

THEOREM 12. Let S be an or-type knowledge base, and a goal G be

G: ¢-P1,P2, °'" ,Pm.
If there exists a D-refutation for SU{G} via a computation rule CR, and 0 is the
composition of all []-unifiers, then v((P1,P2,"" ,Pro)O) is a []-logical consequence of

S.
[Proof] The result is proved by induction on the length of the []-refutation.

[Base step]
In this case, G is a goal of the form ~P1, and there exists Q ( ( S ) and substitution 0
such that Atom(QO)C_Atom(PlO). Since Q0 is an instance of a unit clause of S, v(P10)
is a []-logical consequence of S.
[Induction step]
Suppose that G: e-P1,P2,"" ,Pro has a D-refutation of length k, and 0 1 , ,Ok are
substitutions. Since G has a [] -refutation, there exists a rule C,
C: Q e--RI~2,"" ,RI,
 151

and a selected atom Pi which satisfies Atom(QOk)CAtom(PiOk). By the induction

hypothesis,

v((PI,"" ,Pid,R 1/'" ,Rt~i + 1V- ,Pm)Ol,"" ,Ok)

is a [] -logical consequence of S. Therefore,

V((R1,"" ,Rl)O1,"" ,Ok)

is a []-logical consequence of S. Consequently, V(Q01,".,0k) and V(Pi01,'-" ,Ok) are

[]-logical consequence of S. Hence, v((P1,P2,'" ,Pro)01,"" ,Ok) is a Ill-logical conse-

quenceofS. ,
COROLLARY 13. Let S be an or-type knowledge base. Then

[] -success set of S c { A (Base(Us) I A is a [] -logical consequence of S }.

[Proof] Take any A~( []-success set o r s ). By the definition of success set, SU{e-A}
has a []-refutation with respect to some computation rule. Since A is an element of
Base(Us), A(= VA) is a D-logical consequence of S. !

3.3. C o m p l e t e n e s s of t h e [ ] - r e s o l u t i o n
We show the completeness of the [] -resolution.
THEOREM 14. Ker( []-success set ofS ) = RS.
[Proof] By Theorem 11, RS= [] TS ~ ~). By induction on the length, we show t h a t if

A ( [] TS ~ n, then S U{(--A} has a []-refutation. If it is proved, the following relation

holds,

Ker( { A (Base(Us) I A is a []-logical consequence of S } )

= RS (byTheorem8)
= •Ts~o (byTheoremM)
C_ []-success set of S.
By the definition of the [] -refutation, for any element B in {A~Base(Us) t A is a D-
logical consequence }, there exists an element A in Ker( { A(Base(Us) I A is a []-
logical consequence } ), which satisfies

Atom(A) C Atom(B).
 152

If the ground e l e m e n t a r y formulas A has a [] -refutation, t h e n the ground elementa-

ry formula B also has a [] -refutation. Namely,

{ A (B ase(Us) ]A is a [] -logical consequence of S } c_ [] -success set of S

holds. Hence by Corollary 13,
[]-success set of S = { A (Base(Us) ]A is a [] -logical consequence of S }.

By Ker operation,

Ker( []-success set o f S )

= Ker({AEBase(Us) IA is a O-logical consequence o r s } )
= RS.
Now, we show the proof of "A ( [] TS ~ n, then SU{(--A} has a [] -refutation".

[Base step]
If A E [] TS 1' 1, then A is a u n i t clause and SU{*-A} has a []-refutation.

[Induction step]
IfAE [] T8 ~ k, t h e n there exist a rule C : Q ~- PI,'"Pm which satisfies (a) and

(b).
(a) There exist a substitution 0 such t h a t A = Q0,

(b) For each i ( 1 ~ iG m ), there exists a ground e l e m e n t a r y formula

RiO ( D T s T (k-l) such thatAtom( RiO )c Atom( PiO ).
By the induction hypothesis, SU{e-PiO} has a [:]-refutation for each i ( 1 ~ i --< m ).

Therefore, SO{e-RiO I Atom(RO)CAtom(PiO) } also has a El-refutation, and SU{(--A}

has a [3-refutation. !
LEMMA 15. Let S be the or-type knowledge base, A be an e l e m e n t a r y formula. If
V(A) is a g-logical consequence of S, then there exist a [:]-refutation of SU{~A}

with the identity substitution as the computed answer substitution.

[Proof] By Theorem 14 and L e m m a 8.5 in [2]. !
LEMMA 16. Let S be an or-type knowledge base, G be a goal a n d 0 be a
substitution. If there exists a []-refutation with respect to SU{G0}, t h e n t h e r e
exists a []-refutation with respect to S U{G}.
[Proof] By L e m m a 8.2 in [2]. !
 153

THEOREM 17. Let S be an or-type knowledge base, G a goal and 0 be a substitution.

If G0 is a []-logical consequence of S, then there exists a []-refutation of SU{G} via

a computation rule CR.

[Proof] Let goal G be 4- P1,'"Pm. Since G0 is a D-logical consequence of S,

v((P1,"'Pm)0) is a []-logical consequence of S. Namely, V((Pi)O) is a []-logical

consequence of S for any i( 1 -<

_ i<_
-- m ). By Lemma 15, there exists a []-refutation of

SU{~(Pi)0} with the identity substitution as the computed answer substitution for

any i ( 1 -< - m ). By combining these []-refutations of SU{~---(Pi)O }( 1<= i <--m ), we

- i -<

can make a [3-refutation of SU{Ge} with the identity substitution. Lemma 16

shows there exists a []-refutation of S U{G}. !

4. D i s c u s s i o n

In this paper, we gave a definition of a system, where or-type knowledge can

easily be represented. We have introduced an interpretation of formulas in the

three valued logic, and discussed the declarative semantics, fixpoint semantics and

procedural semantics. Furthermore, we have shown the [:]-logical consequence,

[] TS T co and the E] -refutation characterize the same class of formula. We have also
shown that the []-resolution can be used as a D-theorem prover in the or-type
knowledge bases. However, there are some problems to be solved.
(1) unification algorithm

(2) properties of the ~-resolution

(3) computation rule

(4) selection of input clause

ACKNOWLEDGMENT
I am deeply grateful to Prof. Setsuo Arikawa of K y u s h u University for his
constant support and encouragement.
 154

REFERENCES
[1] R. Reiter : On Closed World Databases, In Logic and Databases, H. Gallaire
Eds., 1979.
[2] J.W. Lloyd : Foundation of Logic Programming, Springer-Verlag, 1984.
[3] Tanaka and Katsuno : Ko-sui-jun Database (unpublished manuscript in
Japanese ).
[4] D.W. Loveland : Near-Horn Prolog, Proc. 4th Int. Conf. on Logic Programming,
1987.
[5] J.W. Lloyd and R.W. Toper : Making Prolog More Expressive, Journal of the
Logic Programming vol.3, 1984.
[6] E.F. Codd : Extending the Database Relational Model to Capture More Meaning,
ACM Transaction on Database Systems, vol.4, no.4, 1979.
[7] W. Lipski : On Semantic Issues Connected with Incomplete Information
Databases, ACM Transaction on Database Systems, vol.4, no.3, 1979.
[8] W. Lipski : On Databases with Incomplete Information, Journal of the ACM,
vot.28, no.l, 1981.
[9] R.Turner :Logics for A rtificial Intelligence, Ellis Horwood, 1984.

A P P E N D I X " S i m p l e R e a l i z a t i o n of a [ ] - p r o v e r in Prolog
The following program is the or-type knowledge base in introduction. We
simply realized the following interpreter for the or-type knowledge base. We have
used the unification algorithm in Prolog.
Program
or2(stay(X, kitakyu), stay(X, iizuka)) :- staff(X, kit).
address(X, fukuoka) :- or3(stay(X, fukuoka), stay(X, kitakyu), stay(X, iizuka)).
staff(taro, kit).
Interpreter
box(X) :- write('START BOX RESOLUTION!!'), nl, X, unique(b(X)),
write('FIND BOX REFUTATION!!').
 155

box(X) :- write('END OF BOX RESOLUTION!!').

set :- retract(b(X)), fail.
set :- write('READY, PLEASE ENTER BOX QUERY!!').
unique(X) :- not(X), assert(X).
or2(X,Y) :- X;Y.
or3(X,Y,Z) :- or2(X,Y);or2(Y,X);or2(X,Z);or2(Z,X);or2(Y,Z);or2(Z,Y).
Execution of Program by the I n t e r p r e t ~
t ?-set.
READY, PLEASE ENTER BOX QUERYII
yes
I ?-box(live(X, fukuoka)).
START BOX RESOLUTION!!
FIND BOX REFUTATION!!
X = taro;
END OF BOX RESOLUTION!!
X = X;
no

I ?-set.
READY, PLEASE ENTER BOX QUERY!!
yes
t ?-box(or2(staff(X, kyushu), staff(X, kit))).
START BOX RESOLUTION!!
FIND BOX REFUTATION!!
X = taro;
END OF BOX RESOLUTION!!
X = X;
no

i?-
 ACCESS PROGRAM TO MINIMIZE REDUNDANT
REFUTATIONS ON THE NETWORK DATABASE SYSTEM

Makoto TAKIZAWA, Motoshi KATSUMATA, and Shinichi NAGAHORA

3 e p t . o f Information and Systems Engineering
Tokyo Denki U n i v e r s i t y
[shizaka, Hatoyama-cho, Hiki-gun, Saitama 350-03, JAPAN

This paper presents how to generate access programs which get a l l answer s u b s t i t u -
t i o n s from t h e network d a t a b a s e w i t h o u t d u p l i c a t e r e f u t a t i o n s from Prolog goal
c l a u s e s and a Prolog view of the conventional network d a t a b a s e system. We show an
improvement of t h e Prolog r e f u t a t i o n procedure by taking advantage of the network
model. In our system, redundant r e f u t a t i o n s , which occur w h i l e f i n d i n g a l l answer
substitutions f o r given goals, are prevented in a d d i t i o n to preventing meaningless
backtracking to find one r e f u t a t i o n by i n t e l l i g e n t backtracking. We show t h a t our
system i s s u p e r i o r to the conventional Prolog system in performance.

1 • I NTRODUCTI ON

Information systems are composed of heterogeneous database systems i n t e r c o n n e c t e d by

communication networks. Users want to m a n i p u l a t e t h e s e database systems indepen-
dently of t h e i r h e t e r o g e n e i t y and d i s t r i b u t i o n [ 2 0 ] . Data s t r u c t u r e s and p r o c e d u r e s
can be defined r e c u r s i v e l y and uniformly in Prolog[6,16] while conventional database
languages l i k e 5QL[9] cannot. Our d i s t r i b u t e d database system[20,21,22,23] provides
users with common Protog views on heterogeneous database systems. This view i s named
a f a c t base system (FSS) which p r o v i d e s a s e t of Prolog ground u n i t c l a u s e s , and
operations for m a n i p u l a t i n g them. That i s , u s e r s can m a n i p u l a t e h e t e r o g e n e o u s
database systems in Prolog as i f they were f a c t base systems. How to provide Prolog
views on t h e r e l a t i o n a l d a t a b a s e systems[8] has been discussed in [ 4 , 1 3 , 2 5 . e t c . ] .
Also, the l o g i c a l r e l a t i o n between Prolog systems and the network database systems
have been discussed in [21,22]. [24] d i s c u s s e s how to find e f f i c i e n t access path to
the r e l a t i o n a l database system and shows the way of t r a n s l a t i n g Prolog goal c l a u s e s
into relational s e t - a t - a - t i m e operations. In t h i s paper, we d i s c u s s how to find e f -
ficient navigational, i.e. r e c o r d - a t - a - t i m e , o p e r a t i o n s on the network database f o r
P r o l o g goal c l a u s e s and d i s c u s s how to reduce the number of redundant answers and
n a v i g a t i o n a l accesses to N. I n t e l l i g e n t backtracking[3] aims at avoiding meaningless
backtrackings to find a r e f u t a t i o n . In t h i s paper, we present a r e f u t a t i o n procedure
 157

which r e d u c e s r e d u n d a n t r e f u t a t i o n s in a d d i t i o n to conventional intelligent back-

tracking. Also, we show t h e e v a l u a t i o n by c o m p a r i n g o u r s y s t e m and c o n v e n t i o n a l
Prolog system.

In S e c t i o n 2, we p r e s e n t a P r o l o g view on t h e c o n v e n t i o n a l neteork database system.

In S e c t i o n 3, we show p r o b l e m s in c o n v e n t i o n a l P r o l o g r e f u t a t i o n procedure, tn sec-
t i o n 4, we show how t o c o m p i l e a goat c l a u s e t o an a c c e s s p r o c e d u r e t o t h e network
database system. In s e c t i o n 5, we show t h e e v a l u a t i o n of our s y s t e m .

2. PROLOG VIEWS ON NETWORK DATABASE SYSTEM

An a b s t r a c t model o f t h e c o n v e n t i o n a l network m o d e l [ 7 , 1 9 ] is presented.

2.1 Network Data Structure

The network d a t a s t r u c t u r e i s composed o f entity (E) t y p e s and f u n c t i o n (F) t y p e s .

An E t y p e i s a t o t a l l y ordered set of E t u p l e s , each of which i s i d e n t i f i e d by i t s
identifier. Each E t u p l e i s composed of a t t r i b u t e values. The E t y p e c o r r e s p o n d s t o
a record type, E tuple to a record occurrence, and i d e n t i f i e r to a database (db)
key in t h e c o n v e n t i o n a l network m o d e l [ 7 ] . An F t y p e S r e p r e s e n t s a partial function
from an E type El to E~. Each element <a,b> of S is an F tuple. For every b in E2, a
subset R[b] of F tuples <a,b> in R is totally ordered. The F type corresponds to a
set type and E[b] a set occurrence in [7], Fig. 1 shows an example of a network data
structure on suppliers (s), parts (p), and parts-subparts relations (b).

Tuples in N are navigationally accessed according to the ordering by the data opera-
tions, which correspond to the conventional DHL operations[19].

2.2 Prolog View of Network Database

A Prolog view D f o r the network database N i s composed o f a f a c t b a s e (FB) and a

r u l e b a s e (RB). The FB i s a c o l l e c t i o n o f ground u n i t c l a u s e s , and t h e RB i s a c o l -
lection of rule clauses. There exists an ( m + l ) - a r y p r e d i c a t e symbol A f o r each E
t y p e A which h a s m a t t r i b u t e s kx . . . . . Am, a b i n a r y p r e d i c a t e symbol C f o r each F t y p e
C, and a c o n s t a n t symbol c f o r each v a l u e c in N. Each c l a u s e A(Oa,a~ . . . . . am) in t h e
FB d e n o t e s an E t u p l e <Sa,a~ . . . . . am> w i t h identifier @a and m a t t r i b u t e values
a~ . . . . . am i n an E s e t A, and C($a,Sb) d e n o t e s an F t u p l e <<@a. . . . >,<@b . . . . >> in an F
t y p e C from an E t y p e A t o B. F i g . 2 shows t h e FB f o r Fig. 1. Here, constant symbols
prefixed by 8 d e n o t e t h e i d e n t i f i e r s of E tuptes. For example, a clause sp(@pl,Ssl)
d e n o t e s an F t u p l e < $ @ p l , c > , < $ s l , a > > i n Fig. 1.
 158

(I) network schema

.... v-T
. . . . . .

(2) network database N

__+___+ . . . . . + - - + - - - + . . . . . + - - + - - - +

s l@s ]snamel p_ _ + _{@p

__+
{pname
. . . . . +
b- - + l@b
- - - +
[
____+______+ . . . . . +

@bl{
@p2 ~
+---+ . . . . . +
@p3 ~
@p

+_-_+ . . . . . + ÷---+

_--+ . . . . . . . + . . . . . . . + ---+ . . . . . + . . . . . . . + _--+ . . . . . + . . . . . . . +

sp ] . . . . P. . . +I . . . . s. . . +l~ pb bp I b l p t
___+ _ _ _ + ] . . . b. . +] . . . . p. . . +[ .... + . . . . . + . . . . . . . +

<~pl c>l<Osl a > l | <@bl> <@pl,c> ]<@bl>t<@p2,d>[

<@p2'd>l<@slla>]} sp[<@sl,a>] l<@b2>l<Opl,c>l + . . . . . 4..........
I<@b2>l<@p3,e>[
÷

< >~ '=<' ~e s ~ = v~J =

, a>/~ +..... +....... + + . . . . .
+ . . . . . . . +
<@p4:f>t<@sl,a> ) l<$b3>l<@p2,d>[
÷ . . . . .
+ . . . . . . . +
<@b3>l<@p5 g>I
<@pS,g>l<@s2 b>]~ sp[<@s2,b>] <@b4>[<@p3 e>] <@b4> <@p5 g>
l<@p6,h>l<@s2:b>l) l<@bS>l<@p3:e>l+ <@b6> + <. @. P. S. ; g. >. .
+ . . . . . +
+ . . . . . . .
+ . . . . . . . + + . . . . . + . . . . . . .

t<Sb6>l<@p6,h>[ I<@bS>l<@p6,h>l+
+ . . . . . + ........
+ . . . . . + . . . . . . . +

Fig. l Network Database

D ={sl) s ( @ s l , a ) , s2) s(@s2,b), pl) p(@pl,c), p2) p(@pZ,d). p3) p(@p3,e>.

p4) p(@p4,f), pS) p(@pS,g), p6) p(Op6,h), bl) b(@bl). b2) b(@b2).
b3) b(@b3), b4) b(@b4), bS) b(@bS), b6) b(@b6). spl) sp(@pl,@sl).
sp2) sp(@p2,@sl), sp3) sp(@p3,@sl), sp4) sp(@p4,@sl). sp5) sp(@p5,@s2).
sp6) sp(@p6,@s2), pbl) pb(@bl,@pl), bpl) bp(@bl,@p2). pb2) pb(@b2,@pl).
bp2) bp(@b2,@p3), pb3) pb(@b3,@p2), bp3) bp(Ob3,@pS). pb4) pb(@b4,@p3).
bp4) bp(@b4,@pS), pbS) pb(@b5,@p3), bpS) bp(@bS,@p6). pb6) pb(@b6,@p6).
bp6) bp(@b6,@pS). }

Fig.2. Fact Base(FB) of Fig. 1

3. PROBLEMS IN PROLOG REFUTATION PROCEDURE

Let ?-(Bx . . . . . B,)O be a goal c l a u s e H where 0 is a substitution. For an atom Bj

s e l e c t e d by a computation r u l e and an input c l a u s e B:- A~. . . . . AK ~here t h e r e e x i s t s a
u n i f i e r a such t h a t Bj8 a = B a , ?-(B2 . . . . . Bx-~,Ax. . . . . AK,B**t. . . . . BQO a i s a r e s o l v e n t
[ F i g . 3 ] . This i s an SLD r e s o l u t i o n [ 2 , 1 6 , 1 7 ] . Variables which denote i d e n t i f i e r s in
the network database N are primary. Underlined atoms denote s e l e c t e d ones.

Let us consider a goal G = ? - s ( S , ? X ) , p ( P , Y ) , s p ( P , S ) for N, where X i s a t a r g e t v a r i -

able, i.e. users want to get s u b s t i t u t i o n s of X. V a r i a b l e s S and P are s a i d to be
primary s i n c e both d e n o t e s i d e n t i f i e r s . F i g . 4 shows an SLD t r e e [ 1 7 ] by the Prolog
 159

r e f u t a t i o n procedure where atoms are s e l e c t e d from l e f t to r i g h t in G. From Fig.2,

there are two clauses of s in D, and six ones of p. Also, s i n c e sp denotes a func-
t i o n from p to s, there i s a t l o s t one clause for a p a i r of s and p clauses. Atoms
s(S,?X) and p ( P , ¥ ) s h a r e no v a r i a b l e . Hence, 2.6=12 d e d u c t i o n s are done and
2+2*(6+6)=26 t u p l e s are derived from N to obtain a l l r e f u t a t i o n s . Next, an SL3 t r e e
by another computation r u l e which s e l e c t s atoms in an order, s, sp, p, i s shown in
Fig,5. Since for each s($k . . . . ), only clauses of sp( . . . . $k) can be accessed, t o t a l l y
six deductions are done, and 14 t u p t e s are derived. Therefore, i f atoms are s e l e c t e d
based on the data s t r u c t u r e , a search space to find r e f u t a t i o n s i s reduced.

?- (B1 . . . . . B~-~,B~,B~+~. . . . . BO0

B : - A~. . . . . AK where BjO a = B

/

?- (BI ..... Bj-I,A~ ..... AK,B~+I..... B.) 0

Fig.3. SLD Resolution.

?-s(S,?X), p(P,Y), sp(P,~;).

S, |s2>
?-p(P,Y), sp(P,~!sl). ?-p(P,Y), s p ( P , @ s 2 ) .

?-s_E}~Pl,$sl). ? - ~ ( $ p 2 , ~ s l ) . . . . ?- ($pl,~s2). ?- ..
{spl) sp2) ~ ~('P2"s2) ' i

Fig.4 SLD Tree by Prolog Procedure.

ca> 7- ,p(P,Y),sp(P,S t.
Is2>
(c) ?-p (P, Y ) , ~ . ?-p (P, Y), sp (P, ~s2).
s~l) / |sp2) lsp3) sp4) [spS) sp6)]

|pl> lp2> Ip3> Ip4> |ps> Ip >

Fig.5 SLD Tree by Another Procedure.

Refutations (1)---(4) in Fig.5 have the same answer s u b s t i t u t i o n s {X/a). Thus, i f

m u l t i p l e r e f u t a t i o n s have the same answer s u b s t i t u t i o n s , they are said to be redun-
dant. ( 1 ) ~ ( 4 ) are redundant. In order to get a set of answer s u b s t i t u t i o n s for t a r -
 160

get v a r i a b l e s , i t is meaningless to obtain the r e f u t a t i o n s ( 2 ) ~ ( 4 ) a f t e r (1) is ob-

t a i n e d . Another s u b s t i t u t i o n {X/b} can be obtained by backtracking from (d) to (a).
In t h i s paper, we t r y to prevent these redundant r e f u t a t i o n s .

4. ACCESS PROGRAM

We consider a program which d e r i v e s a l l the answer s u b s t i t u t i o n s f o r a goal clause G

by navigationally accessing tuples in N.

4.1 Basic Behaviors of C e l l s

In t h i s paper, we assume t h a t e v e r y goal c l a u s e G contains no r u l e atoms. In our

system, i f G i n c l u d e s a r u l e R, R i s r e p l a c e d with t h e body of t h e r u l e c l a u s e
u n i f i a b l e M t h R. In t h i s paper, r e c u r s i v e r u l e s are not considered.

An access program f o r a goal c l a u s e G is composed of c e l l s i n t e r c o n n e c t e d by one-way

channels. Each c e l l C denotes an atom C in G. C i s composed of an ordered s e t D(C),
i.e. E type or F t y p e , in N, s t a t e variable P(C), and two i n p u t ports, i.e.
C:FS(FirSt) and C:NX(NeXt), and two o u t p u t ports, i.e. C:SC(SuCcess) and C:FL
(FaiLure). There are s p e c i a l c e l l s named ST and OUT. C e l l s A and B are connected by
a one-way channel A:O-~B:I from an A ' s o u t p u t port A:O to a B's input port B:I.
Data u n i t named a token i s t r a n s m i t t e d through the channel. C e l l s behave as follows.

[Cell ST] (1) F i r s t , ST sends an i d e n t i t y s u b s t i t u t i o n ~ as a token at ST:SC.

(2) On r e c e i p t of a token at ST:NX, ST makes a l l c e l l s t e r m i n a t e .
[Cell C] (I) On r e c e i p t of a token 0 at C:FS, D(C) is s e q u e n t i a l l y accessed based
on i t s ordering relation. That i s , r e c o r d types and s e t types [7] are naviga-
t i o n a l l y accessed in t h e c o n v e n t i o n a l network d a t a b a s e system. If a tuple t,
such t h a t t ~ = COa, i s found, a token 6 a is s e n t a t C:SC and t is s t o r e d in
P(C). I f not found, 0 is sent at C:FL.
(2) On r e c e i p t of 0 at C:NX, D(C) is s e q u e n t i a l l y a c c e s s e d from a t u p l e t in
P(C). C behaves l i k e (1).
[Cell OUT] On r e c e i p t of 6 at OIJT:FS, an answer s u b s t i t u t i o n of 0 is output to the
user, and 0 i s sent a t OUT:FL.VI

In Fig.6, an access program for a goal ?- p(P,?Y),pb(B,P),p(?PP,X),b(B),bp(B,?PP)

is shown, which represents the Prolog refutation procedure. For example, if a tuple
t unifiable with p(P,?¥) is derived from the E type p in N in the cell (I), the sub-
stitution 0 such t h a t p(P,?Y) O= t i s sent to (2) v i a (1):SC-*(2):FS. I f (4) f a l l s ,
(4) sends a token to (3) v i a (3):FL-~(2):NX. However, s i n c e s u b s t i t u t i o n s obtained
 161

by (3) i n c l u d e no b i n d i n g of t h e v a r i a b l e B in (4), backtracking from (4) t o (3) i s

meaningless. Further, this program does not accesses N by making use of t h e d a t a
structure o f N.

+ output port * input port S: SC F: FL

F i g . 6 P r o l o g Access Program.

Bere, i f C:FL-+A:NX and C:SC-+B:FS, let parent(C) and n e x t ( C ) d e n o t e c e l l s A and B,

respectively. Also, let prior(A) d e n o t e C. p a r e n t ( C ) shows a c e l l t o be b a c k t r a c k e d
from C and n e x t ( C ) shows a n e x t s e l e c t e d atom. In F i g . 6 , (3) = p a r e n t ( 4 ) and (3) =
next(2). If a cell A h a s no c e l l B such t h a t A = parent(B), A i s s a i d t o be a l e a f .
(5) i s a l e a f in F i g . 6 .

4.2 Computation Rule Based on the Network Data Structure

First, we c o n s i d e r a c o m p u t a t i o n r u t e which s e l e c t s atoms a c c o r d i n g t o t h e u n d e r l y -

ing data structure i n N. Relationships among t y p e s in N are denoted by s h a r e d
primary variables among t h e atoms which d e n o t e t h e t y p e s . For atoms A and B, and a
substitution O, if AO and BO include t h e same v a r i a b l e X, A and B a r e s a i d t o
s h a r e X on O .

Now, a c o m p u t a t i o n r u l e SP f o r a goat c l a u s e G i s g i v e n a s f o l l o w s . Bere, c o s t ( A O )

g i v e s an e x p e c t e d number of t u p l e s t o be a c c e s s e d t o f i n d an i n p u t c l a u s e whose head
is unifiable with A#. It is defined later. A function vinst(A,O) returns O{X/cx
where X i s an u n i n s t a n t i a t e d variable X in AOand cx i s an unused c o n s t a n t symbol.

[Simple c o m p u t a t i o n r u l e SP] (0) I n } t i a l l y , A = C = NULL, Oo = Ox = c . and j = O.

( I ) S e l e c t an atom B whose c o s t ( B O o ) i s t h e minimum in G. Go t o ( 3 ) .
(2) For a c e l l P = parent(A), select an atom B in G whose c o s t ( B O ~ i s t h e minimum
among o n e s which s h a r e v a r i a b l e s w i t h P on Op.
(3) [B i s found] Construct a cell B. I f A = NULL, t h e n make a c h a n n e l B:FL-+ST:NX,
else B:FL-+P:NX and C:SC-+B:FS. C = B, ~ = vinst(B,O~), remove B from G, A =
B, j = j + l , 0 j = 0 ~-lO B. Go t o (2).
(4) [B i s n o t found]
4-1) I f 6 i s empty, make C:SC-+OUT:FS and OUT:FL-+ST:NX, and t e r m i n a t e .
4-2) I f p a r e n t ( A ) i s n o t S?, A = p a r e n t ( A ) and go t o ( 2 ) .
 162

4-3) I f parent(A) = ST, then s e l e c t an atom B whose cost(BO~) is the minimum.

Make C:SC-*B:FS and B:FL-~A:NX. C = B, Oe = v i n s t ( B , O ~ ) , A = B, and remove B
from G. j = j + l , 8a = Oj-18e. Go to ( 2 ) . [ ]

A sequence of atoms o b t a i n e d by the SP i s r e p r e s e n t e d by an o r d e r e d t r e e where

d e p t h - f i r s t order of c e l l s g i v e s a s e l e c t i o n sequence of atoms[21,22]. Fig.7 shows
the t r e e obtained by the SP f o r Figo6, and Fig.8 shows a deduction by the SP. From
F i g . 7 and Fig. l, it is clear that the sequence of s e l e c t e d atoms r e p r e s e n t s an
access path in N.

(I>I<t
YE ?-(p(P, ?¥),pb(B,P),b(B),bp(B,?PP),
p(~PP,X)) Oo
1
b(...).
?-(p(P ?Y~,pb(B P~_,bp(B,?PP),p(?PP,X)) 81
1 I
• pb(...).
?- (P_~ _?Y_i,bp (B, ?PP), p (?PP,X) ) 02
Oa=e Ol=0o{B/c.} 02:01 {P/e~} p(...).
•

0~= Oz{Y/cj 0.= 03{PP/epp} ?-(bp_~ ?~]_,p(?PP,X)) 0

/
bp(.. , ) .
Fig.7 Navigational Tree. ?-p__(?P, X~O
p(.. ,).
/
Os Os= O~{X/cx}

F i g . 8 Deduction by SP.

F i g . 9 shows an a c c e s s program o b t a i n e d by t h e SP, which a c c e s s e s n a v i g a t i o n a l l y

t u p l e s in N. Since a c e l l (1) is a parent of (4) in Fig.7, t h e r e e x i s t s a channel
(4):FL-*(1):NX. If (4) f a i l s , (I) i s r e s o l v e d by our r e f u t a t i o n procedure as sho~n
in Fig.9 although (3) i s resolved in the Prolog. Thus, our program not only accesses
N by taking advantage of the access path supported by N but a l s o prevents meaning-
l e s s backtrackings.

ST
<I <I IJ <I
~-§ b(B) ~-~ P) SC-~ ?Y) SC-~ ?PP) SC-~ X ) SC-~ OUT....
I
Fig.9 Access Program.

A rightmost cell which includes target variables in the access program is said to be
a target cell. In Fig.9, (4) is a target cell of a variable PP. There exists a func-
 163

tion from (I) to (2). Bence, even if ~e backtrack from (3) to (2), (2) fails since
for every clause unifiable with (I) there exists at most one clause unifiable with
(2).

[Improvement] For a cell C, let B be parent(C) and A be parent(B). If there exists a

function from a type denoted by A to a type denoted by B in the network database N,
then remove C:FL-~B:NX and add C:FL-~A:NX.[[]

By t h e above improvement, an a c c e s s program S as shown in Fig. lO i s o b t a i n e d .

I....i iilil....iiill......iill......iiill......iliill......IT
ST ~ - ~ b(B) ~ - ~ P) SC-~ ?Y) SC-~ ?PP) SC-~ X) SC-~ OUT

Fig. lO Improved Access Program $.

4.3 Reduction of Redundant R e f u t a t i o n s

Next, we try to get all answer substitutions without redundant refutations. Although
the access program S in Fig. 10 can find a refutation without meaningless backtrack-
ings by taking advantage of the underlying data structure, all the answer substitu-
tions cannot be obtained. After one refutation is found, the same backtrackings as
the Prolog program are required, ~hich are denoted by edges +++> as shown in Fig. ll.

-4.- + -+- - + - "--+. -~-- - + - '--4:. -+-- -~" 4-+-.$

NX~% ~L+*~x
+ % k+ + IN×
% k++ % I~x k+ % +~NX k++ *NX
% ++FL
++÷÷4÷+++ ++÷+÷++÷+ ++÷++t++++ +++÷+++++ ++++++++4+ +++++4++++
(1) (2) (3) (4) (5)

Fig. ll Access Program with Prolog backtrackings.

An access program T is considered to be a sequence of cells <ST,Cx ..... C~,OIJT> where

Cj~= next(Cj) for j=1 ..... m-1. For example, S in Fig. t0 is <ST, b. pb, p, bp, p, OUT>. Let
root(T) be Cx and last(T) bc Cm. Cj is said to be greater than Cm written as C, <TCJ,
iff j>h. A ceil B is a descendent of a cell A in T if A= parent(B) or parent(B) is a
descendent of A. tail(A) is an A's descendent B if there exists no A's descendant C
such that B <TC. Let a subtree of A, Sub(A), be a substring <A,A~ ..... An> of T where
A,= tail(A). According to the definition of the SP, it is clear that Sub(A) includes
 164

only and all descendants of A. The following notations are introduced.

Itarg(A)= a greatest target cell B where B <TA.
rtarg(A)= a least target cell B where A <:B.
join(A,B)= a least common ancestor of A and B.
njoin(A,B)= an ancestor D of A such that join(A,B) = parent(D).
rjoin(A)= njoin(rtarg(A),A).
ljoin(A)= njoin(A, Itarg(A)).
In Fig. 11, join(3,4) = 1, njoin(3,4) = 3, rjoin(2) = 3, ljoin(5)= 5, rtarg(3) = 4,
and tlast(T) = 4.

For a cell A in T, suppose that there are n cells gz..... B, where A = parent(Bj) for
j=l ..... n. Let T~ be Sub(B,). Suppose that a substitution 6 A is obtained by the
resolution of A. Let Ans(B,) be a set of refutations obtained from Tj6A. According
to the property of T, any two T~8~ and TKS~ have no common uninstantiated variable.
This means that Ans(g~) can be obtained independently of other TK(k~j). In the
Prolog program, for each ASA, a cartesian product CAA= Ans(B~) x...x Ans(B~) is ob-
tained. Here, let T~ be a subtree which includes targets in Tt ..... T~ (k=1 ..... p,
pen). In our method, only a projection of CA~ on T~ ..... T~, i.e. AA~= Ans(B,~) x...x
Ans(B~) is accessed. Since it is clear that ICA~I ~ IAA~I, we can get all the answer
substitutions in less accesses to N than the Prolog program.

In one method to get CAA, each Ans(B~) is stored in the intermediate storage and
then the cartesian product is taken by accessing them. If Ans(Bj) is so small to be
stored in a main memory, the cartesian product can be efficiently taken. In a case
of database, since Ans(B~) is too large to be stored in the main memory, it has to
be stored in the files. However, only sequential access method is supported on the
intermediate files. So, we adopt a nested-loop method in which Ans(B~) is obtained
in the outermost loop ..... and Ans(Bn) is obtained in th e innermost loop, where
tuples are accessed in every loop.

T T,,~~

NX+ ~ .... I ,LFL

+++++++ ++++++++ + +~++'~-4-+ +

Fig.12 Access Program

Suppose that the resolution of B,.~ f a i l s in the deduction, i . e . all answer s u b s t i t u -

t i o n s in Ans(B,+~) are already obtained, and that Tj includes a target cell 0
[Fig.12]. By backtracking from B~.~ to D, resolutions for Sub(O) are executed. When
the resolution of t~il(O) succeeds, a new answer substitution for Sub(B,) is ob-
 165

rained. Now, r e s o l u t i o n s a r e forwarded to Bj+~ from t a i l ( O ) . Here, Ans(Ba) i s ob-

tained again by r e s o l u t i o n s f o r Tj+~. I f 0 f a i l s , we backtrack to C = parent(O) and
r e s o l u t i o n s of S u b ( O a r e done. On s u c c e s s of t h e r e s o l u t i o n of t a i l ( C ) , T~+, i s
r e s o l v e d by forwarding to B,÷x. By t h i s program, AAA can be obtained instead of CAA.
Bere, f o r every c e l l C in Tj, rjoin(C) is Bj÷~ and ltarg(Bj+~) is O. To r e p r e s e n t the
b a c k t r a c k i n g from Ba÷~ to 0 and t h e forwarding from c e l l s in Tj to B~+~, new output
p o r t s LFL and RSC are introduced. The b a c k t r a c k i n g and f o r w a r d i n g a r e denoted by
channels B~+~:LFL-*O:N× and tail(O):RSC-~Bj÷x:FS, r e s p e c t i v e l y .

We show a general computation r u l e ~P f o r reducing redundant r e f u t a t i o n s by adding

t h e s e new channels.

[General Computation Rule GP] GP is obtained by modifying SP as follows.

(1) Remove OUT:FL-~ST:NX, and add OUT:FL-~tlast(T):NX.
(2) For each c e l l A, i f f o r some C, A = r j o i n ( O , then add A:LFL and A:LFL-*
ttarg(A):NX.
(3) For each l e a f A, i f S u b ( l j o i n ( A ) ) includes a t a r g e t c e l l , then add A:RSC, and
A:RSC-*rjoin(A):FS i f rjoin(A) e x i s t s , A:RSC-~OUT:FS o t h e r w i s e . [ ]

A general access program is one generated by the 6P, Fig. 13 shows a general access
program T for F i g . lO. Since tlast(T) is (4), OUT:SC-~ (4):NX. Since (4) =
nrjoin(3,4), (4):LFL-*(3):NX and (3):RSC~(4):FS.

; *+++++++; ; ; ++++++++++++++++++++~ +
+ + + +

Fig. 13 General Access Program.

Two kinds of tokens have to be d i s t i n g u i s h e d , one f o r f i n d i n g the f i r s t refutation

and one f o r f i n d i n g a l l answer s u b s t i t u t i o n s . A token is a 4 - t u p l e <Type, Begin,
End, S u b s t i t u t i o n > , where Type i n d i c a t e s a kind of token, i.e. F shows t h e f i r s t
type and N the second type, Begin and End i n d i c a t e a subtree <Begin . . . . . End> in T.
First, Begin = root(T) and End = l a s t ( T ) . When a token t is sent via A:LFL-~B:NX,
t's Begin i s B and End i s C = t a i l ( B ) . A f t e r r e s o l u t i o n s of Sub(B), a token has to
be sent from C v i a C:RSC-*A:FS. On success of r e s o l u t i o n of C, C decides at which
p o r t SC or RSC a token i s s e n t . I f End i s C, i t i s s e n t at RSC, otherwise at SC.
Each c e l l behaves as follows.
 166

[Cell ST] (1) F i r s t , ST s e n d s a t o k e n < F , r o o t ( T ) , l a s t ( T ) , ¢ > a t ST:SC.

(2) On r e c e i p t of a t o k e n a t ST:NX, ST makes a l l cells terminate.
[Cell A] (1) On r e c e i p t o f < N , B , E , O > a t A:FS or A:NX, i f B = E = _, l e t B be A and
E be t a i l ( A ) . I f A <r B, l e t B be A and E be t a i l ( A ) .
(2) D(A) i s a c c e s s e d as t h e same a s SP. I f an u n i f i a b l e t u p l e t in D(A) w i t h AO
is obtained, where AO 6 = t o ,
(2-1) I f E = A, t h e n <N . . . . . 66> i s s e n t a t A:RSC.
(2-2) I f B iT A <T E, <N,B,E, O 6 > i s s e n t a t A:SC.
(3) I f no u n i f i a b l e tuple is obtained,
(3-1) I f B <T A ~TE, t h e n < N , B , E , O > i s s e n t a t A:FL.
(3-2) t f A=B and r j o i n ( C ) = A , t h e n <N. . . . . O> i s s e n t a t A:FL.
(3-3) O t h e r w i s e , <N. . . . . 0> i s s e n t a t A:FL.
[Cell OUT] On r e c e i p t o f < . . . . O> a t OUT:FS, an answer s u b s t i t u t i o n from 0 is out-
p u t and <N. . . . . O> i s s e n t a t OUT:FL.~

[Proposition] By t h e GP, a l l answer s u b s t i t u t i o n s f o r a goal G a r e o b t a i n e d by l e s s

redundant refutations than the Prolog refutation procedure.
[Proof] Let ? he a g e n e r a l a c c e s s program f o r G. S u p p o s e t h a t for a cell A i n T,
there are n cells B~ such t h a t A = parent(Bj) in T ( j = l . . . . . n ) . Let T~ be a s u b t r e e
sub(B~), Let 0 be a s u b s e t {0~ . . . . . 0.} £ {B~. . . . . Be} where Sub(OK) i n c l u d e s target
c e l l s ( k = 1 . . . . . m). Suppose t h a t T j O . i n c l u d e s Nm r e f u t a t i o n s . In t h e P r o l o g s y s t e m , N
= Nsx*,..*NB. r e f u t a t i o n s are required to get all answer s u b s t i t u t i o n s of Sub(A) O . .
On t h e o t h e r hand, in T, s i n c e o n l y one r e f u t a t i o n i s o b t a i n e d f o r T j 6 ~ where B~ i s
n o t i n O, M = Nm*...*Na, r e f u t a t i o n s a r e done. I t i s c l e a r t h a t H i s l e s s t h a n N.II

4.4 aeeess Cost

Let us c o n s i d e r t h e c o s t f u n c t i o n s c o s t ( A O ) for a cell A and a s u b s t i t u t i o n O. Let

O~ be an o r d e r e d s e t in N d e n o t e d by A and aA be t h e p r o b a b i l i t y t h a t each t u p l e in
OA i s u n i f i a b l e with A. aA i s computed based on t h e s e l e c t i v i t i e s [ 1 4 ] of a t t r i b u t e s
in the set. L e t FSA be the expected number of t u p l e s which a r e a c c e s s e d in O~ t o
g e t aI1 u n i f i a b l e tuples w i t h A. FSa depends on t h e a c c e s s p a t h s u p p o r t e d by 0~. I f
O~ h a s no a c c e s s path, FSR i s I0~1. If the j-th attribute of O~ has an a c c e s s p a t h
and i t is instantiated i n A, FS~ i s a~*LO~t where a ~ is the selectivity of t h e j -
th attribute.

For two c e l l s A and B, t h e c o n n e c t i v i t y CN~ from A t o B i s t h e a v e r a g e number of

tuples accessed in OB t o f i n d t u p l e s by which comlon u n i n s t a n t i a t e d variables of A
and B a r e instantiated. If B denotes an F t y p e and A i s its range, EN~B i s an
a v e r a g e number o f t u p l e s < r , p > f o r each t u p l e p. In t h e c o n v e n t i o n a l network model,
 167

the connectivity[20] r e p r e s e n t s an a v e r a g e number of member r e c o r d o c c u r r e n c e s f o r

each owner r e c o r d o c c u r r e n c e and a p r o b a b i l i t y that e a c h member h a s an o ~ n e r in a
s e t type [7]. No~, we c o n s i d e r t h e c o s t o f a g e n e r a l a c c e s s program T.

First, let S be a s u b t r e e o f a c e l l A where A = p a r e n t ( B ~ ) ( j = l . . . . . n) in T. Let S~

be Sub(B~). Let Cs be an e x p e c t e d number o f t u p l e s a c c e s s e d in o r d e r t o o b t a i n a l l
t h e a n s w e r s f o r each t u p l e in A, and Ss be an e x p e c t e d number o f answers f o r S. Cs
i s g i v e n by (N~ + F~ a ~ +1. Ns and F~ a r e t h e e x p e c t e d numbers o f t u p t e s a c c e s s e d f o r
s u b t r e e s T~ which c o n t a i n t a r g e t c e l l s and none, r e s p e c t i v e l y ,
Fs = Zj-x ......FCC~*Cs~.
FCC~sj = CN~m/(I + C N ~ * a ~ j ) if T~ includes no target cells, 0 otherwise.
FCCAm is an expected number of tuples in Bj to find one tuple which satisfies the
condition of Be.
Ns = NCC,m*Cm + SSsx*( NCC~2*Cs~ + SSs2*( NCC~*C~ + . . . + SS~-x*(NCC~*Csn) . . . ) )
NCeas~ = CN~s= i f S= i n c l u d e s t a r g e t c e l l s , 0 otherwise.
SS~ = S~ i f S= i n c l u d e s t a r g e t c e l l s , 1 otherwise.

The e x p e c t e d number Ss o f a n s w e r s i s given as f o l l o w s .

Ss = a 4 * H j.~ ..... , A C C ~ * S S ~
ACC~j = CN~ i f Sj i n c l u d e s t a r g e t c e l l s , 1 otherwise.
As a r e s u l t , an e x p e c t e d number CTT o f t u p l e s a c c e s s e d f o r T in o r d e r t o o b t a i n a l l
t h e answers i s a s f o l l o w s .
CTr = FS,o~<r>*CTr.

The t o t a l c o s t COSTT i s a w e i g h t e d summation o f CTT and ST.

COSTT = WC*COSTT + WS*ST.

For a goal G, i t is difficult t o compute Cr and Sr f o r e v e r y p o s s i b l e a c c e s s program

T. So, we use t h e h e u r i s t i c s based on f o l l o w i n g a c c e s s f u n c t i o n s . Here, s u p p o s e t h a t
f o r an atom A t h e r e a r e n atoms B~ ~ h i c h s h a r e common v a r i a b l e M t h A ( j = l . . . . . n) on
O. T~o c o s t f u n c t i o n s a r e c o n s i d e r e d .
cost1(B~8) = CN~j
c o s t z ( B ~ 8 ) = CN~j*(1 + ~ * F ( a , b ) )
Here, F(a,b) = a*(1 - b"-~)/(1 - b), a = ZK.~..... ~-~.m...... C N ~ / ( n - 1 ) , and b = EK-z..... 5-
~J+1...... a s J ( n - 1 ) . In c o s t a , c e l l s a r e s e l e c t e d among t h e n e a r e s t n e i g h b o r c e l l s . In
c o s t z , t h e e f f e c t o f r e s o l u t i o n o f T~ t o o t h e r Tk i s t a k e n a c c o u n t t o .

5. EVALUAT I ON

Our s y s t e m named LIP (Logic I n t e r f a c e P r o c e s s o r on d a t a b a s e s y s t e m ) has been a l r e a d y

 "~68

implemented in IJTI-Lisp[5] on M-380Q, whose f a c t b a s e and r u l e b a s e a r e implemented

by t h e r e l a t i o n a l database management s y s t e m AIM/RDB[1][Fig. 14]. S i n c e a network
database s y s t e m AIM/DB c a n n o t be used in our university, a network database
s i m u l a t o r i s implemented as a F o r t r a n a p p l i c a t i o n program o f AIM/RDB. Fig. 15 shows a
network d a t a b a s e used by t h e e v a l u a t i o n .

t Prolog

I.... l.... IP: i n f e r e n c e p r o c e s s o r

[ F B [ FB: f a c t b a s e
[ ~ RB: r u l e b a s e

Fig. 14 LIP

I!iiiiiiiit :- - iiiiiiii!iiiiI<L Iiiii!iiiiI::tii!iiiiii : liiiiiiiiiiiiI

Fig. 15 Network D a t a b a s e

For goal clauses, the following points are measured.

icount = the number of tuples accessed.
t c o u n t = t h e number o f t u p l e s o u t p u t as answer s u b s t i t u t i o n s .
crime = t i m e f o r g e n e r a t i n g an a c c e s s program T[msec].
atime = e x e c u t i o n t i m e of T[msec].
etime = crime + etime.
We e v a l u a t e t h e LIP i n o r d e r t o make c l e a r (1) how many m e a n i n g l e s s b a c k t r a c k i n g s
a r e r e d u c e d and (2) how good a c c e s s p a t h i s s e l e c t e d by our c o m p u t a t i o n r u l e GP. In
(1), we measure b o t h t h e LIP and t h e P r o l o g program f o r t h e same goal c l a u s e s whose
atoms a r e s e q u e n c e d a s s e l e c t e d by t h e 6P.

I'_!t-'I'_!_t-'t!:t-'I_!_I
->I!'_t-`I'_It
->I!:I-`I':t->I!,_t-`yt->Ii'`!I +
++++++++++++++++++3
+

Fig. 16 Access Program

For a g o a l G = ?-St(S,_),Ls(S,L),L(L,_),TI(T,L),T(T,_),Ss(SB,S),Sb(SB,?SNAME),
 t69

Sp(P,SB),P(P,_), a g e n e r a l a c c e s s program T as shown in Fig. 16 i s g e n e r a t e d . Results

of e x e c u t i o n of Prolog program and T a r e shown in Table 1, From t h i s table, icount
and t c o u n t in our method a r e less than Protog. T h i s means t h a t our method can
d e c r e a s e t h e number of b a c k t r a c k i n g s and redundant r e f u t a t i o n s . Our method needs an
additional c o m p i l a t i o n time. In our s y s t e m , a large amount o f d a t a a r e a c c e s s e d .
Problem i s how to reduce t h e number of t u p l e s a c c e s s e d . Access time to the d a t a b a s e
dominates t h e t o t a l time. The c o m p i l a t i o n time, i.e. ctime, can be n e g l e c t e d com-
pared with the access time, atime. Hence, even i f t h e network data s t r u c t u r e and
g o a l s a r e more c o m p l i c a t e d , our method can be a p p l i e d .

Table 1 Result Table 2.

- - + . . . . . . . + . . . . . . . + . . . . . . . + . . . . . . . .

. . . . . . . . + . . . . . . . . . . . . + . . . . . . . . . . { t GP1 J GP2 { P r o l o g
__+ . . . . . . . ÷ . . . . . . . +. . . . . . . +. . . . . . . .
I Our system { Prolog
. . . . . . . . + . . . . . . . . . . . . + . . . . . . . . . . {icount [ 511 16231 5621
icount I 609 1 48333 Itcount I 101 8801 1060
tcount { 88 { 880 tlctime t 8901 38941
ctime { 1321 t Iatime { 93221 253169{
atime } 34436 I letime I 10212{ 257063{ 178145
- - + . . . . . . . + . . . . . . . + . . . . . . . + . . . . . . . . .

etime [ 35757 I 171649

. . . . . . . . + . . . . . . . . . . . . + . . . . . . . . . . {icount } 28211 5211 655
Itcount { 880t 88{ 88
2{ctime } 892} 1041t
latime { 2423661 833411
letime I 2432581 843621 69665
- - + . . . . . . . + . . . . . . . + . . . . . . . + . . . . . . . . .

{icount { 45811 48331 1710

ltcount I 880{ 8801 880
3{ctime [ 8131 11171 3976
latime ! 2486891 4999771 253776
{etime t 2494991 5010941 257752
- - + . . . . . . . + . . . . . . . + . . . . . . . + . . . . . . . . .

Next problem i s which a c c e s s path our system g e n e r a t e s . In Table 2, cost~ i s used in

GP1 and cost2 in GP2. P r o l o g i n d i c a t e s t h a t a goal c l a u s e w r i t t e n by an d a t a b a s e e x -
pert i s p r o c e s s e d by t h e P r o l o g program. Table 2 shows t h a t GP does not s e l e c t al-
ways good a c c e s s paths. An a c c e s s program i s composed of two p a r t s , one f o r f i n d i n g
a refutation and t h e o t h e r for finding answer s u b s t i t u t i o n s . The f o r m e r aims a t
decreasing i c o u n t and t h e l a t t e r aims a t d e c r e a s i n g t c o u n t . Decreasing i c o u n t does
not n e c e s s a r i l y mean d e c r e a s i n g t c o u n t . So, we t h i n k a method t h a t i f a l a r g e number
of answer s u b s t i t u t i o n s are retrieved, an a c c e s s program i s g e n e r a t e d so as t o m i n i -
mize t h e number of answer s u b s t i t u t i o n s , i.e. tcount, otherwise so as t o m i n i m i z e
t h e number of t u p l e s a c c e s s e d , i.e. ~count.

6. CONCLUDII~G RHdARKS

In t h i s paper, we p r e s e n t a Prolog interface on t h e c o n v e n t i o n a l network d a t a b a s e

 170

system, This system is now operational on M-380Q. Users can retrieve data from the
network database system as if it were a Prolog system.

In our system, r e d u n d a n t r e f u t a t i o n s which give the same answer s u b s t i t u t i o n s are

prevented by accessing n a v i g a t i o n a l l y the network d a t a b a s e w i t h o u t making i n t e r -
mediate f i l e s , Our method can not only be applied to the netuork database system but
also the r e l a t i o n a l database system. The r e l a t i o n a l database system provides t u p l e -
a t - a - t i m e i n t e r f a c e like cursor [9] in a d d i t i o n to the non-procedural SQL i n t e r f a c e ,
Our method can apply to the t u p l e - a t - a - t i m e i n t e r f a c e , in a n o t h e r method, i n t e r -
mediate r e s u l t s are created, we think t h a t i t is d i f f i c u l t to e f f i c i e n t l y access the
intermediate files because they do not provide e f f i c i e n t access path other than a
s e q u e n t i a l access method.

Our system cannot necessarily generate an optimal access program. We are considering
the method by taking usage of knowledge of database a p p l i c a t i o n e x p e r t s on access
paths. At p r e s e n t , we t r y to provide a t r a n s a c t i o n functions[23] which manipulate
the network database system atomically.

REFERFaNCES

[1] "FACOMOS IV/F4 MSP AIM/DB," F u j i t s u .

[2] Apt,R. and van Emden,M.H., "Contributions to the Theory of Logic Programming,"
JACM,Vol.29,No.3, 1982, pp.841-862.
[3] Cambell,J.A., "Implementations of Prolog," E l l i s horwood limited, 1982.
[4] Chang,C.L., "On E v a l u a t i o n of Queries Containing Derived Relations in a Rela-
t i o n a l Data Base," Logics & Database, Ptenum Press, 1981.
[5] Chikayama, "FACOMUTILISP Manual," Univ.of Tokyo.
[6] Clocksin,W.F. and Mellish,C.S., "Programming in Prolog," Springer-Verlag, 1984.
[7] CODASYLDDL Committee, "CODASYL Data Description Language," Journal of Develop-
ment, 1973, 1978.
[8] Codd,E.F., "A R e l a t i o n a l Model of Data f o r Large Shared Data Bank," CACM,
Vol.13,No.6, 1970, pp.337-387.
[9] Date,C.J., "An Introduction to Database Systems," Addison-wesley, 1981.
[I0] Dayal,U., et al., "Query Optimization for CODASYL Database Systems," Proc. of
the ACM SIGMOD. 1982, pp.138-150.
[11] £ndertom,H., "Mathematical Introduction to Logic," Academic Press, 1972.
[12] Gallaire,fl., Minker,J,, and Nicolas,J.M., "Logic and Database: h deductive Ap-
proach," ACM Computing Survey, vol.16, No.2, 1984, pp.153-185.
[13] Henschen,L.J. and Naqvi,S.h., "On Compiling Queries in Recursive First-order
Databases," JACM, VoI.31, 1984, pp.47-i07.
 171

[14] Hevner,h. and Yao,S.B., "fluery Processing on a Distributed Databases," Proc. of

the 3rd Berkeley Workshop, 1978, pp.91-107.
[153 K o b a y a s h i , I . , " C l a s s i f i c a t i o n and Transformations of Binary Relationship
Schemata," Sunno I n s t i t u t e of Business Administration, 1985.
[16] Kowalski,R., "Logic for Problem Solving," Research Studies Press, 1979.
[17] Lioyd,D., "Foundation of Logic Programming," Springer-verlag, 1985.
[18] Ohsuga,S., "Developing a Deductive Relational Database for Uniform Handling of
Complex Queries," JIP(IPSJ), 1983, pp.123-137.
[19] 011e,T., "The CODASYLApproach to Data Base Management," John Wiley & Sons,
1978.
[20] Takizawa,M., "Distributed Database System-JDDBS," JARECT, Ohm-sha and North-
holland, Vol.7, 1983, pp.262-283.
[213 Takizawa,M,, "Deductive Network Database System," Journal of JSAI, Vol.2, No.2,
1987, pp.182-191.
[223 Takizawa, M., et a l . , "Logic Interface System on Navigational Database System,"
Lecture Notes in Computer Science, Springer-Verlag, No.264, 1987, pp.70-80.
[233 Takiza~a,m M., "Transaction Management by Prolog," Proc. of the Logic Program-
ming Conf., ICOT, 1987.
[24]~arren, D. H. D., " E f f i c i e n t Processing of I n t e r a c t i v e Relational Database
Queries Expressed in Logic," Proc. of the VLDB, 1981, pp.272-281.
[253 Ullman,J.D., "Implementation of Logical Query Languages for Databases,"
TODS,Vol. IO,No.3, 1985, pp.289-321.
 EUODHILOS: A General-Purpose Reasoning Assistant System
- - Concept and Implementation-

Toshiro MINAMI, Hajime SAWAMUI~A

International Institute for Advanced Study of Social information Science (IIAS-SIS),
FUJITSU LIMITED, 140 Miyamoto, Numazu, Shizuoka 410-03, Japan
E-mail: {tos, haj ime}7.iias, f u j i t s u , junet ~aunet. uu. n e t
and
Kaoru SATOH, Kyoko TSUCHIYA
Software Development Laboratory, FUJITSU LABORATORIES LTD.,
1015 Kamikodanaka, Nakahara-ku, Kawasaki, Kanagawa 211, Japan

Abstract

The concept of the general-purpose reasoning assistant system is presented. We

have developed the first version of a system named EUODHILOS as a prototype of that
kind of system. It assists the user in describing the syntax of the logical expressions,
defining a variety of logics which consists of axioms, inference rules, and rewriting rules,
and also in constructing proofs for theorems under the logic so defined. Proofs are
constructed under the support of the facility named "sheet of thought", which is an
interactive and visual proof editing environment. Proofs are expressed and edited in
the tree structured forms of the natural deduction-like style.

1. I n t r o d u c t i o n

In these days, logics play important and even essential roles in many fields like
mathematics, computer science, artificial intelligence, and so on. In fact, various logics
such as first-order, higher-order, equational, temporal, modal, intuitionistic, and type
theoretic logics are used in these fields, and human reasoning activities based on these
logics appear in daily work.
ttere the phrase "human reasoning" is used to denote the process consisting of the
following three phases.
 173

(1) Making mental images about the objects and concepts.

(2) Making logical models which describe the mental images.
(3) Examining the models to make sure that they are sufficient.
The process begins with the first phase when one becomes aware of some mental
images of objects and concepts having some structures and Mso wants to clarify what
they are. To clarify the mental images, one has to describe them formally. A formal
framework for describing objects and concepts is called a "logic" in this paper. A
"logical model" is a logical description which specifies some mentM image.
The second phase is for making logical models. Since the language is important
for describing objects, one defines the syntax of the language at the beginning. In this
phase, one has to determine what objects, concepts, and relations appear in the universe
of discourse. The logical structure of the world of objects can be described by giving the
axioms and derivation rules. The derivation rules are given in forms of such as inference
and rewriting rules.
In the third phase, one derives results from the logical model. One comes to know
many formal properties of the model. At the same time, one examines the correctness
of the model. The model is insufficient if some properties which are expected to hold
by the image of the objects fail to prove in the model. In this case, one has to modify
some or M1 of the logical expressions about the objects. Sometimes one has to modify
not only the logical expressions, but also the definition of the language used for the
modeling.
Since human beings reason in various fields, it is valuable to help them by a
computer-based system. The purpose of the reasoning assistant system is to make the
human reasoning process accurate and being more efficient with the aid of the computer.
Two major subjects are pursued for realizing a reasoning assistant system. The
first one is the "generality" of the system. Here the phrase "general-purpose system"
indicates that it is free from any specific logic. As S. K. Langer told [Langer 25],
we recognize that "Every universe of discourse has its logical structure." That is a
thought that for each object which we mention, there must be a logic best suited for
expressing about it. In order to assist human reasoning for the object, the reasoning
assistant system must have the ability to allow the user to describe all the existing logical
structures and manipulate the expressions under those logics. Our system EUODHILOS
(pronounced "you-oh'-dee-los") is named as an acronym of the phrase by Langer. This
reflects our intention to emphasize the generality of the system.
The other subject is the "user-friendliness" of the system. We investigate the
reasoning-oriented human-computer interface that can be established as an aspect of rea-
soning supporting facilities. A system having good interface so that it can be used easily,
is helpful for one to conceive ideas in reasoning. Furthermore, a reasoning methodology,
which often reminds us of programming methodology, needs to be investigated.
We aim at building an ideal reasoning assistant system; which can be used for
general-purpose and is used easily. EUODHILOS is a prototype of such a system. We
intend to clarify the concept of the ideal general-purpose reasoning assistant system
through developing and using EUODHILOS.
In Section 2, the reasoning assistant system is characterized through the compar-
isons to several other types of system which can be used for assisting human reasoning,
 174

specifically to theorem provers, proof checkers, and proof constructors. In Section 3, an

overview of EUODHILOS is illustrated. In the succeeding two sections, its two signifi-
cant features are described. In Section 4, the feature of defining logics is explained. In
Section 5, the feature of assistance for constructing proofs, called "sheet of thought", is
explained. In Section 6, some of the other features of EUODHILOS are illustrated. In
Section 7, some concluding remarks and the directions for future research are stated.

2. C h a r a c t e r i z a t i o n of the Reasoning Assistant System

We consider the following four as the system-types which can be used for assisting
human reasoning:
(i) reasoning assistant system
(ii) automated theorem prover
(iii) proof checker
(iv) proof constructor
In the rest of this section the reasoning assistant system is comparatively charac-
terized with other types of system. In Subsection 2.1 the reasoning assistant system is
introduced, and its characteristic features are exhibited. In Subsections 2.2 to 2.4, it is
compared with each of the other types of system.

2.1 Reasoning Assistant System

The purpose of the reasoning assistant r Symbol ...............

system is to support the whole phases of hu- I / sy~ . . . .
man reasoning process. The concept of the [l <~mt~a>- .
f l Internee R u , e - -
system comes from the philosophy of Langer
[Langer 25] introduced in Section 1, and the
recognition that the (mathematical) reason-
ing proceeds through "Proofs a~ld Refuta~
tions" [Lakatos 76].
The following two major characteristic
features of the system reflect these underly-
ing concepts respectively:
(i) General-purpose
(ii) Possession of user-friendly proof editing
environment
Figure 2.1 is an illustration of how the ,:%
reasoning assistant system is used. In the
upper half of the figure which corresponds to
the feature (i), the user specifies each of the F i g u r e 2.1 Reasoning Assistant System
 I75

components of a logic, i.e., symbols, syntax of expressions including (logical) formulas,

inference rules, etc. In the lower half of the figure, corresponding to (ii), the user tries
to construct proofs of theorems under the logic defined in the previous step. Proofs are
searched by trial and error.
EUODHILOS is the only one reasoning assistant system. In the system, partially
constructed proofs, which are called proof fragments, appear scatteringly on a sheet of
thought. The user edits these proof fragments by the editing commands such as create,
delete, extend (derive), connect, separate, and so forth. The sheet of thought is the
environment for creating theorems and their proofs. The theorems on the sheet can be
saved so that they may be reused in the later proofs for other theorems. They can be
used just in the same way as axioms.

2.2 Automated Theorem Prover

An (automated) theorem prover is a system which searches a proof of a given

formula. Figure 2.2 illustrates how the theorem prover is used. In the upper half of the
figure, a user tries to prove a formula. The user starts by putting an assumption of the
formula. But he has no idea how to proceed for completing the proofi In the lower half
of Figure 2.2, the user gives the formula to the theorem prover. The system finds out a
proof, and displays it on the screen.
The theorem prover is different from the
other types in that the proofs of theorems are
given by the system. In other types of sys-
tems, including the reasoning assistant sys-
tem, proofs are given by the user.
The situation in Figure 2.2 is, in fact, an
ideal one. Considering the current state of
the art of the proof-finding power of theorem
provers, there is little hope of finding practi-
,U,
cally useful proofs automatically by comput- Theorem
ers. From this observation, theorem provers O 0 implies , ~
Proof
are to be used for finding only those proofs
which fill small gaps which appear in a large
proof. If the gap in a proof is small, the
prover can find the proofs effectively. On
one hand human beings are good at making
a plan how to find a proof; on the other, they
are not good at doing things accurately. Ma-
F i g u r e 2.2 Theorem Prover
chines are on the contrary to human beings.
This is the reason why we pay attention to support the process of human reasoning by
the computer. The capacity of computers for finding proofs can be used as a part of a
reasoning assistant system.
 176

2.3 P r o o f C h e c k e r

In this paper, a system is called a proof

checker if it checks the correctness of a proof
described by the user. Figure 2.3 is an illus-
tration of how the proof checker is used. In
the upper half of the figure, the user makes a oo xx.l
proof of a theorem. A human proof may con- I
tain some careless mistakes including small
gaps in a proof. The checker provides a lan-
guage for describing human proofs. By using
this language, the user describes the proof.
In the lower half of the figure, the user gives
the description of the proof to the checker. If
the checker finds errors in the proof, it shows O 0 imp ~.~,
them on the screen. We can see an error in- Proof

dication in the figure.

When a user has a proof and wants to
verify its correctness, a proof checker is one
__m i
ERR08 ~

of the best tools for him. But when one be-

gins to find a proof for some formula, proof
constructors described in the next section is
more suitable for his purpose of proof finding. F i g u r e 2.3 Proof Checker
Many proof checkers have been devel-
oped up to now. AUTOMATH [de Bruijn 70] is a proof checker in which the user
can specify how the proofs are constructed. PL/CV2 [Constable 82] is used for proving
the correctness of P L / I like programs. CAP-LA [ICOT 86] deals with the proofs on
linear algebra.

2.4 Proof Constructor/Editor

A proof constructor is a system which
supports a user to construct proofs as well
as theorems through the interaction between {10} O 0 imp I::11:3
the user and the system. The proof con- (11} Assume ~:~
struction is, in other words, a "proof edit-
ing." Users edit proofs, precisely proof frag- (20) II
ments, by inputting, deleting, and combining
the proofs. From this point of view, a proof
constructor is a proof editor. The function
of the proof constructor is included in a rea-
soning assistant system as an aid to proof F i g u r e 2.4 Proof Constructor
constructions.
 177

Figure 2.4 is an illustration of how the proof constructors are used. The user
interactively uses~ a proof constructor by entering some commands and tries to find a
proof of some theorem.
Many proof constructors have been developed up to now. For example, LCF [Gor-
don 79], FOL [Weyhrauch S0], EKL [Ketonen 84] and Nuprl [Constable 86] are well
known.
The purpose of Nuprl is very similar to ours. It aims at providing the proof con-
struction environment for various logics. But the approach to the realization of it is
different from those of the reasoning assistant systems. Nuprl has a fixed underlying
logic and other logics are defined by using the expressions in this logic. In the ap-
proach of reasoning assistant system, even the syntax of the logic is specified by the
user. It aims at the complete realization of logic free systems which can assist human
reasoning in the various fields. Proof construction methodologies are also different. In
Nuprl, proofs are constructed only by refinement, while in the reasoning assistant sys-
tem, proofs are constructed by three types of deductions; i.e. by forward, backward
(same as refinement), and interpolating deductions. The interpolation is a deduction
which fills a proof gap between two formulas. Another difference is on user-interfaces;
that is, Nuprl has been based on character displays, while EUODHILOS uses bit-map
displays so that symbols can be expressed as its natural shape on a paper.
The significant difference between the proof constructors and the reasoning as-
sistant systems is that in the former, underlying logics are fixed, while in the latter,
underlying logics can be defined by the user. There are merits and demerits for fixing
the underlying logics. As a merit, it is easy to introduce some specific procedures suited
to the logic. As a demerit, if the system is applied for general cases of human reasoning,
the fixation of logic may restrict the reasoning about some objects under consideration.
In such a case, a generM framework treating a variety of logics is required.

3. O v e r v i e w of E U O D H I L O S

EUODHILOS is a prototype of the general-purpose reasoning assistant system. Its

first version is now working on the sequential inference machine PSI on an operating
system named SIMPOS.
It is designed by considering the following issues:
(1) Realization of a logic-free reasoning system, based on the philosophy in [Langer
25].
(2) Provision of an environment to be used easily to those who are not necessarily
familiar with computers.
(3) Support of logical thought, symbolic or logical manipulations done by human rea-
soners.
(4) Environment for experimenting logical model construction based on the philosophy
in [Lakatos 76].
 178

The issue (1) is reflected to EUODttlLOS in its logic definition feature which is
shown in Section 4. For (2), it is implemented under the overlapping window system
with bit-map display. It can be used easily because most of the operations are achieved
by selecting the menu item with the mouse. For (3) and (4), the major tools used
in EUODHILOS are "editors." The user can modify the language definitions, logic
definitions, proof constructions, and so forth directly through the visual editor windows.
Considering the importance of the proof construction, the system provides a special
proof editing environment called "sheet of thought", which is an interactive, easy to
use, and visual proof constructor.

_EUODHILOS
,~-Logical System S h e e t of T h o u g h t

--Language System ~ ~

User

I~...~om ...................... : l[ I I L--4,,I___J

Iit-.... . . . . .

i IT1 I
'i' i i

F i g u r e 3.1 An Overview of EUODHILOS

Figure 3.1 shows an overview of EUODHILOS. The system consists of two major
parts, one for defining a user's logical system and the other for constructing proofs ill
a sheet of thought. The former is concerned with the second phase of human reasoning
discussed in Section 1. It provides the feature for defining the syntax of logical expres-
sions called (well-formed) formulas and the logical structure of the mentioned objects,
as well. The latter is concerned with a proof construction which corresponds to the third
phase of human reasoning. It is a proof editing environment, where the user constructs
the proofs by trial and error. On the sheet of thought, proofs are expressed in the form
reflecting their derivation structures. In EUODHILOS proofs are represented in natural
deduction style, i.e. a tree structure, they are shown in their familiar representation.
Among the various features of EUODHILOS, we put stresses on the following topics:
(i) Defining Logics
(ii) Constructing Proofs (Sheet of Thought)
In the following Sections 4 and 5, the features of defining logical expressions and
supporting to construct proofs on the sheet of thought are explained respectively mainly
based on the specification for the current (the first) version of EUODHILOS.
 179

4. Defining Logics

4.1 Language Description Feature of EUODHILOS

In EUODHILOS, a language system to be used is designed and defined by the user
at the first stage. The language system consists of the specifications of the syntax of
the logical expressions (i.e. formulas). The syntax of the expression is given by using
definite clause grammar (DCG)[Pereira 80] in the first version. We intend to modify
DCG by adding the operator-declarations in the second version in order to decrease
the amount of descriptions. From the description in DCG, a bottom-up parser called
BUP[Matsumoto 83] can be automaticMly generated. We modify the BUP algorithm
by adding the function so that the augmented operators can be treated.
The system generates not only a parser, but Mso an unparser for the defined lan-
guage. The unparser translates from the internal expressions into external ones which
can be understood by the user. The parser and unparser are used in all the following
phases of symbol manipulations. When an expression is entered, the parser is invoked
for checking the validation of it. At the same time the internal structures of the ex-
pression of the language are constructed as well. When derivation commands are given
by the user, the internal expressions of the formulas are manipulated and new internal
expressions are generated. These expressions are presented to the user after translated
into the external ones by the unparser.
Figure 4.1 in the next page is an example description of the first-order logic in DCG
with operator and symbol declarations.

4.2 Derivation Description Feature of EUODHILOS

A derivation system in EUODHILOS consists of axioms and derivation rules which

are given as inference and rewriting rules. A finite set of formulas is given as the axiom
system. Inference rules are given in a natural deduction-like style presentation by the
user. An inference rule consists of three parts; the first one is the premises of a rule,
each of which may have an assumption, the second is the conclusion of a rule, and
finally the third is for the restriction that is imposed on the derivations of the premises,
such as variable occurrence conditions (eigenvariable). Well-known typical styles of logic
presentations such as Hilbert's style, Gentzen's style, equational style can be treated
within this framework.
Schematically, inference rules are given in the natural deduction style format as
follows:

[Assumption1] [Assumption2] ... [Assumptionn]

Premisel ,,, Premise2 -.- Premisen ....

Conclusion
 180

Syntax description:
:formula('=>' (F1,F2))---+formula(F1), "=:>", formula(F2)
formula('h' (FI ,F2))-+formula(Fl), "A", formula(F2)
formula('V' (FI,F2))-~formula(Fl), "V", formula(F2)
formula( '-~'(F))-+"-~", formula(F)
formula (F)-+at omit_formula (F)
atomic_formula (F)-+" (", formula(F), ")"
atomic_formula (F)-+predicat e_symbol (P), "( ", term_list (TL), ") ",
{F:.. [PI TL] }
t e r m - l i s t ( [T] )-~t erm (T)
t e r m _ l i s t ( [T I TL] )-+term(T), ", ", t e r m . l i s t (TL)
term (T)-+funct ion-symbol (F) ,''(", term_list (TL), ") ", {T=.. [FI TL] }
term (T)-+variabl e_symbol (T)
term (T)--~const ant_symbol (T)
Operator declaration:
"~" : xfx, I00
"A": y:fx, 50
"V": yfx, 50
"~": fy, 30
Symbol declaration:
predicate-symbol : "p .....r"
function_symbol: "f ....'h"
variable_symbol : "x"-"z"
constant-symbol : "a"-"e"
Meta symbol declaration:
formula: "P"-"T"
%erm: "A"-"D"
Figure 4.1 A description of the first-order logic

In this format, each of the assumption parts is optionM. If a premise has the as-
sumption, it indicates that the premise is obtained under the assumption, and otherwise
it is obtained unconditionally. An inference rule may have a condition on the eigenvari-
able. An inference rule is applied if all the premises are obtained in this manner, and
the restrictive condition is satisfied. Then~ the conclusion is obtained by the application
of the rule.
Rewriting rules are presented in the following format:

Pre_Expression
Post-Expression
 181

A rewriting rule indicates that it is applied to an expression when the formula has
a subexpression which matches to the pre_expression part of the rule. The resultant
expression is obtained by replacing the subexpression with the expression corresponding
to the post_expression part of the rule. Rewriting rules have no condition of application
in the first version.
Iterating the applications of the derivation rules described above, one can get a
derivation (tree). The following Figure 4.2 is an example of the derivation using both
inference and rewriting rules.

wyz.( + y)z = + yz (rE)

(a ÷ O)b = ab + Ob
(~ + 0 ---, x)
ab-ab+Ob (m=y-..-*x+z-y+z)
0--0b (m=y---*y-x)
ob = o (vI)
Vx.Ox = 0
F i g u r e 4.2 A derivation in EUODHILOS

5. Constructing Proofs

In EUODHILOS an environment called the "sheet of thought" provides the assis-

tance to find proofs of theorems by trial and error. This originates from a metaphor of
work or calculation sheet and is apparently analogous to the concept of sheet of asser-
tion due to C. S. Peirce [Peirce 74]. It allows one to draft a proof, to compose proof
fragments, detach a proof, to reason by using lemmas, and so on.
On the sheets of thought, proof fragments (in other words, partially constructed
proofs) are the elementary units for manipulation. Proof fragments are newly created
as assumptions, axioms, or theorems of the theory, which are composed, and deleted
according to the operations given by the user.
In a sheet of thought, applications of inference and rewriting rules are possible
in the same style as those users use on the paper. This naturally induces that the
appearance of a proof structure on the sheet is also the same as that on the paper. This
way of treating is considered as an example of the proof visualization.
It is desirable that reasoning during proof construction can be done along the
natural way of thinking of human reasoners. Therefore EUODItILOS supports the
typical method for reasoning, that is, forward (or bottom-up) reasoning, backward
(ortol>-down) reasoning, interpolation (Le. filling the gap between proof fragments) and
reasoning in a mixture of them. They are accomplished interactively by manipulating
the fragments on a sheet of thought. It is planned to incorporate not only such a proving
methodology but also methodology of science (e.g., Lakatos' mathematical philosophy
of science [Lakatos 76], Kitagawa's relativistic logic of mutual specification [Kitagawa
63], etc.).
 182

As an example of deduction process on a sheet, we will illustrate how one can

proceed the forward deduction. In order to deduce forward by applying an inference
rule, one has to start by selecting the formulas used as premises of the rule. The
selection is achieved with the mouse by clicking the desired formula. Then the user may
select the inference rule by calling the operation menu of the sheet, or he may enter a
formula as the resultant formula. If the user selects a rule, then by the "do it" command
(action) the system applies the rule to the premises and deduces the resultant. If the
user indicates the resultant formula, then the system searches the list of deduction rules
and tries to find one which is applicable to this deduction.
Among the editing functions on a sheet of thought such as delete, copy, move, undo,
separate, etc., we will illustrate two of them; connection and separation.

(1) Connection

The user can connect two proof fragments if they have equal formulas, where one of
them must be a resultant of a fragment and the other a hypothesis of another fragment.
In order to connect the two proof
fragments, one selects the two formulas
by clicking them with the mouse at first.
One of the two is a resultant, and the
other is a hypothesis which must be equal [A] [A~B]
to the resultant. And then, he invokes B [A] [A~B]
the operation menu and selects the "con- B
nect" item. As the result, the proof frag- [B] c [D]
ments are connected into the one frag- C [D] E
ment. Figure 5.1 is an illustration of this E
operation. The two fragments in the left
part of the figure are connected at the for- F i g u r e 5.1 Connection at B
mula "B", and the fragment in the right
part is obtained.

(2) Separation

This operation is just the inverse of the connection. When the user indicates a
formula in a proof fragment and selects the command "separate" in the operation menu,
then the fragment is separated into two parts at the position of the selected formula.
Suppose, for example, we have a proof fragment like the right one in the Figure 5.1.
If we point the formula "B" in the proof fragment and select the separate command in
the command menu, then we can get the two proof fragments which appear at, the left
part of Figure 5.1.

An example of the screen image of sheet of thought in EUODHILGS is indicated

in Figure 5.2.
 183

•II~IIIILI

FONT
eq~tl
SO~_~'Y8OARD
SYNTAX zerm
--->
--->
"-':
b_~arm; let
IppI~ ----> "°';
REWRI T I N G - R V ~ I

DERIV~_RLA.~ b_~e~l ----> v l r i ~ b l e ;

AXIOM ~ - - -

4 5 b s'c O ) I~lll

I,. . . . . I
(AZB) ox-A. (B. X) M.X~X*X
('~bs'~ O ) • , (sbs'~ O)
(A~8) • (A,B) - A - (B • ( A Z B ) ) M . (A~B) - (AZB) • (A~B)
($b~z 0 ) ('~ b,~'c 0 )

. u (, o )
"'°' '' I!i

..x..
......

(A*I~D . x - A . (M-X)
(sbsz O)
III IIII

M,x-x-x
........... (,~bsz O )
I+ ]"llI H~,~

(eq O) (A~9) - x - A - O8-x)

(~.~. (~.~ -A. ( (~,~ (~,~ ~ ii

.... ~ ....... III IIIJIII IIIIMl+lllH I /ii]

F i g u r e 5.2 Sheet of Thought

6. O t h e r Facilities for Reasoning

In order to make the system user-friendly and easy to use, we have to pay much
attention to the visualization of things. For this purpose, the bit-mapped display with
multi-window environment, mouse, icon, pop-up menu, etc. are exploited in the imple-
mentation of EUODHILOS. The followings are available as the user-interface facilities
for inputting formulas, visualizing formulas, and assisting reasoning.

6.1 Software Keyboard and Font Editor

The software keyboard and the font editor are used to design and input special
symbols often appearing in various formal systems. It is a matter of course that provision
of special symbols which reasoners a~e accustomed to use makes it possible to reason
in a natural way of thinking on a computer. The user designs the fonts of the symbols
in the dot matrix pattern of them by using the built-in font editor. These fonts are
assigned on the software keyboard and are used with the corresponding keys. In Figure
6.1, one can see a software keyboard on which user-defined symbols such as V, A, ~ , 3,
etc. are attached.
 184

i.tu-tvpe li: SYNTAX*" intu_t~pe " ', . . . . . . . . . . I ,)/.',%v,'z,v/;:,',',%v;,',v://;:.v:/:,v.%%*;

SOFT_KEYBOARD .v
FONT W E.~me.t - - > term~.eont&Ln.
SOFT_KEYBOARD I:: f e r r e t - - >
| m b d a . V ~ r | ~ b | e . ".
~erma~
SYNTAX I:* *'?"~ " ' >
,.|( t e ~ m ~ - - > ~ u n © t J o n . ope. term~
INFERENCLRULE
REWRITING-RULE I': tt eerm;~
r~-->
m e t a _ ~ l , "{",met~
- - > |er~2~
!i:
AXIOM I:"
"(', terent,")";
:i:i
:.)
SHEET_OF_THOUGHT t e r ~ - - > not. termS;
.................................. term3 - - > In. "(", termL~ ")";
~erm3 --> fU,c t ion;
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
~ I~ ~ : 3 [ ] © :i:i
%,A':.-,'A%',%'o*/;/.'.'.'/.'.'.'.',':A'2.,.'.'22 ~ __~
:,:.:.:-:-:°:-:-:-:°:.:-:.:°:°:-:-:.:. SHEET_OF_THOUGHT : i n t u _ ~ p e
~f~=-adi ~ar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [':'
0
SHEET_OF_T~ 0
!i!i!iii!i!i
O ::::':?:::: I 0
¢
{ ~ Ii ¢
o 1
¢ :i::.:i::'i:i
::::::::::::
i tnr
l 4 .; ::::?:::::: Ln
[xgP]
--(l.lI{ "g'l¢
tnl(x)~;~/(P
~::~ '

:i:!:i:i:i:i x O

kX. t O t , l ( x ) S P : l ::.V:; f inl

lnr((DC.I~tnI(x)))EPV(P=£)
(tnrI{2})
[f~(PV(PO~)lDi] %
(:E(2}) fl)Inr'(1 ~ . fDlnt (x)) 15.1.
~Utnr ((k~. ~lll.I (x)) )m/~

~uamplal

F i g u r e 6.1 Using the user-defined fonts in the system.

6.2 Formula Editor

The formula editor is a structure editor for logical formulas. It a~ms to simplify
inputting formulas by displaying complicated formulas. In addition to ordinary editing
functions, it provides some formula rewriting functions. It can be used not only for
entering formulas, but also for just displaying the structure of formulas clearly. This is
a visualization of formulas. You can see the formula editor in the bottom-right corner
of Figure 6.1.

6.3 Stationery for Reasoning

Independently of a logic under consideration, various reasoning tools such as de-
cision procedures become helpful and useful in reasoning processes. In a sense it may
also play a role of a model which makes up for a semantical aspect of reasoning. You
can display and erase the stationery at any time within the system. Currently, a logic
calculator for Boolean logic is realized as a desk necessary. Figure 6.2 indicates what it
looks like on the screen.
 185

=~nkax..~di~or name : =~
| i n f e r e n c e _ r u 1 e _ e d i ~or
| rewr~ing_rule.edikor
| axzom-a~i tor
I proof_mditor ~

:il/ l l ~ ~
• i .........................
I1 {FI )

....... ~L (F21
}roo~_eait=r

" - ' " " []

2 f or'm~ 1a4 (F2:~ ~r~l*4(F$) !

I~(a) ]
--(31 ( ~ ) ) 3 = a l l ( u n t f!~(F2, { O P . V . F I } ) } ;
f ~ r m u l ~ 4 ( l ~ ) - * > I~mllt~_formU|ll(l~F);
- - * ( - . E {2, 3~ ) f~rm~),a4(SF) ~ > t l l r m ( B F ) ;
J.
(~I {3} )
($, " ~ ' } } ,
(~I {31 ). 4
(~(x) } [O(x)VO ( x ) ]
(:1 ( } ) [v(4) > b~sl=.(~PaluI ~(V$}
(S, "G" } ] ,
--> pred£=~te_s~q~oi (S}. ¢811 (equ~J . s t r ~ n ~ l l
(~ (~x~ (x) })::O/X ( - . 4 ( x ) ) ) O(x)
~(', tor~(Tl)0 ~}',
c a ! I ( tm~fu(V~, ( ' ~ ' , T I } } ) I
~est

Figure 6.2 Boolean logic calculator

7. Concluding Remarks and Directions for Future Research

The first version of EUODHILOS is now available and the second version is under
development. The system is being improved by reflecting the experience of using it.
So far, we have dealt with logics, such as first-order logic (NK), propositional modal
logic (T), intensional logic (IL), combinatory logic, and Martin-LSf's type theory. Many
logics can be treated in the current version. Though, some logics such as tableau method
can not be treated in the current framework for logic description. We intend to extend
the framework so that these "irregular" logics can be treated in the system.
From the experiments so far in EUODHILOS, the followings become aware.
(i) Describing the syntax of logical expressions is difficult at first. But, by making
good use of the experience, it becomes easier. If the system keeps descriptions for
typical logics as a library, the description of a new logic is an easy task even for
beginners.
(ii) On a sheet of thought, users are free from deduction errors. On the paper, they
may make mistakes in deriving a new formula when deduction rules are applied.
The difference is important, because the users have to pay attentions only to the
decision how to proceed the proof on the sheet of thought.
 186

(iii) The reasoning assistant system can be used as a tool for CAI. In the system, users
can deM with a variety of logics.
The current state is the first step toward the realization of a practical reasoning
assistant system. To put the step forward, we have to investigate various subjects
including the followings:
,, Treatment of relationships between meta and object theories
• Maintaining dependency relations among various theories
• Opening up various new application fields of reasoning
,, Improvement and refinement of human-computer interface for the reasoning system
The experiments with different logical systems have shown the potential and use-
fulness of EUODHILOS in the realm of logics appearing in various fields.

Acknowledgements

The authors are grateful to Dr. T. Kitagawa, the president of IIAS-SIS, and Dr.
H. Enomoto, the director of IIAS-SIS, for their ceaseless encouragements. We wish to
thank also to Mr. T. Hal for his contribution to our project and to Dr. K. Kobayashi
and Mr. Y. Takada for their many valuable comments on the draft of the paper. This
work is a part of the major research and development of FGCS project conducted under
the program set up by MITL

References

[Constable 82] R. L. Constable, S. D. Johnson and C. D. Eichenlaub : An Introduction

to the PL/CV2 Programming Logics, LNCS 135, Springer-Verlag, 1982.
[Constable 86] R. L. Constable et al. : Implementing Mathematics with the Nuprl Proof
Development Systemj Prentlce-Hall, 1986.
[de Bruijn 70] N. G. de Bruijn : The Mathematical Language AUTOMATH, its Usage,
and some of its Extensions, In M. Laudet, D. Lacombe, L. Nolin and M. Schutzen-
berger (eds.), Symposium on Automated Demonstration, Springer-Verlag, pp.29-
61, December 1970.
[Gordon 79] M. J. Gordon, A. J. Milner and C. P. Wadsworth : Edinburgh LCF, LNCS
78, Springer-Verlag, 1979.
[Goguen 83] J. A. Goguen and R. M. Burstall : Introducing Institutions, LNCS 164,
Springer-Verlag,pp.221-270, 1983.
[Griffin 87] T. G. Griffin : An Environment for Formal Systems, ECS-LFCS-87-34,
University of Edinburgh, 1987.
[Harper 87] R. Harper, F. Honsell and G. Plotkin : A Framework for Defining Logics,
ECS-LFCS-87-23, University of Edinburgh, 1987.
[Ketonen 84] J. Ketonen and J. S. Weening : EKL - - An Interactive Proof Checker,
User's Reference Manual, Dept of Computer Science, Stanford University, 1984.
 187

[Kitagawa 63] T. Kitagawa : The Relativistic Logic of Mutual Specification in Statistics,

Mere. Fac. Sci. Kyushu University, Ser. A, 17,1, 1963.
[Lakatos 76] I. Lakatos : Proofs and Refutations - - T h e Logic of Mathematical Discovery
--, J. Worrall and E. Zabar (eds.), Cambridge University Press 1976.
[Langer 25] S. K. Langer : A Set of Postulates for the Logical Structure of Music, Monist
39, pp.561-570, 1925.
[Matsumoto 83] Y. Matsumoto, H. Tanaka, H. Hirakawa, H. Miyoshi and H. Yasukawa
: BUP : A Bottom-Up Parser Embedded in Prolog, New Generation Computing
1, pp.145-158, 1983.
[Peirce 74] C. S. Peirce : Collected Papers of C. S. Peirce, Ch. Hartshorne and P. Weiss
(eds.), Harvard Univ. Press, 1974.
[Pereira 80] F. C. N. Pereira and D. H. D. Warren : Definite Clause Grammars for Lan-
guage Analysis - - A Survey of the Formalism and a Comparison with Augmented
Transition Networks, Artificial Intelligence 13, pp. 231-278, 1980.
[Smullyan 85] R. Smullyan : To Mock a Mockingbird, and Other Logical Puzzles In-
cluding an Amazing Adventure in Combinatory Logic, Alfred A. Knopf Inc.,
1985.
[Weyhrauch 80] R. W. Weyhrauch : Prolegomena to a Theory of Mechanized Formal
Reasoning, AI Journal 13, pp. 133-179, 1980.
[Mizoguchi 85] F. Mizoguchi et al. : Prolog and Its Applications 2, Souken Syuppan,
1985. (in Japanese)
[ICOT 86] ICOT CAP-WG : The CAP Project (1)-(6), Proc. 32nd Annual Convention
IPS Japan, 1986. (in Japanese)
[Satoh 86] K. Satoh et al. : Well-Formed Formulas Editor for the Reasoning Assistant
System, Proc. 33rd Annual Convention IPS Japan, 1986. (in Japanese)
[Minami 86a] T. Minami and H. Sawamura : Proof Constructors for the Reasoning
Assistance, Proc. 33rd Annual Convention IPS Japan, 1986. (in Japanese)
[Minami 86b] T. Minami and H. Sawamura : A Construction of Computer-Assisted-
Reasoning System, Proc. 3rd Conf. JSSST, 1986. (in Japanese)
[Tsuchiya 87] K. Tsuchiya et al. : Well-Formed Formulas Editor for the Reasoning
Assistant System, Proc. 35th Annual Convention IPS Japan, 1987. (in Japanese)
[Minami 87] T. Minami and H. Sawamura : Proof Construction with Working Sheets
- - A Consideration on the Methodology for Computer Assisted Reasoning --,
Proc. 35th Annum Convention IPS Japan, 1987. (in Japanese)
[Sawamura 87] H. Sawamura and T. Minami : Conception of General-Purpose Reasoning
Assistant System and Its Realization Method, Proc. WG of Software Foundation
IPS Japan (87-SF-22), 1987. (in Japanese)
 Logic Based Lexical Analyser LAX
RySichi Sugimura
Matsushita Electric Industrial Co.,LTD.
K6ji Akasaka
Yukihiro Kubo
Institute for New Generation Computer Technology (ICOT)
Yfiji Matsumoto
Kyoto University

Abstract
This paper describes a method of Japanese morphological analysis
which runs in a sequential or parallel logic programming language.
In the method presented here, a morphological dictionary is compiled
into a logic program which executes lexical analysis deterministically.
Morphemes and their categories are written in a morphological
dictionary. The categories are used as inputs to a non-deterministic
finite state automaton (NFSA).
The dictionary is compiled into a TRIE structure in which mor-
phemes can be quickly looked up using an index search mechanism
such as those provided in logic programming languages like ESP or
Prolog.
The layered stream technique, which records search space in recur-
sively structured data, is applied in analysis. All solutions are given
at the end of analysis. The analysis output can be used as input to
the syntactic analyser, SAX, which can also be executed in parallel
logic languages.
This method cau process unknown morphemes which are not in
the dictionary in linear order. This paper also reports the evaluation
of this method in sequential languages.
 189

1 Introduction
This paper presents a morphological analysis method which can be executed
both in parallel and sequential languages, tn this method, a morphological
dictionary is transformed into sequential logic programming languages, such
as ESP or Prolog, or parallel logic programming languages, such as GHC [8].
These programs analyse a sequence of Japanese characters deterministically
[7].
Each word entry in the dictionary contains the appropriate morphological
information. The analysis procedure in LAX can be regarded as the break-
down of the input Japanese string into a concatenation of sub-strings that
correspond to words with some constraints. This mechanism can be repre-
sented as a non-deterministic finite state automaton (NFSA) [5]. The layered
stream technique [6] is applied to obtain multiple solutions. This method
gives high performance both in sequential and parallel logic languages. The
method uses local ambiguity packing [Tomita 85].
Section 2 describes the form of the lexicaI dictionary. Each lexical entry
has the sequence of surface characters, a category, and two kinds of special
features: special constraints on concatenation and state transition. This
section presents the model of concatenation in terms of a NFSA.
The compiled dictionary has a TRIE structure [3], in which lexical entries
are searched for at high speed with the indexed search mechanism provided
by logic programming languages.
Section 3 presents an elementary parallel algorithm. This algorithm can
be regarded as a parallel version of Cocke-Kasami-Younger algorithm or the
bottom up chart parsing algorithm method [4] used for regular grammars.
Section 4 presents a practical algorithm which is obtained by optimising
one of the elementary parallel algorithms. In this algorithm, the number of
processes required is reduced drastically, and data structures of morpholog-
ical constructions are packed. All results are given at the end of analysis
without backtracking, and unknown morphemes can be found in the algo-
rithm. The results can be used as input to the parser SAX [10].
Section 5 describes optimisation for sequential implementations and sec-
tion 6 evaluates the sequential implementation.
 190

2 Dictionary

2.1 Overview of the Lexical D i c t i o n a r y

Four kinds of information should be declared for each morpheme. The first
two are easy to describe: they are the surface of the morpheme and its
category. The third and fourth entries are described later.
For readers unfamiliar with the problem posed by Japanese orthography,
a few words of introduction are in order. A normal Japanese sentence is
written using both Kanji (Chinese characters) and Kana (Japanese phonetic
characters, each representing a syllable). The input is generally a mixed
sequence of both kinds of characters.
It should be noted that when the input is Kana only, heuristics must be
utilised to obtain a morphological analysis. However, we are not concerned
with this special case.
A recent study of Japanese morphology [5] reports that finite state au-
tomata have just enough power to analyse strings of this kind. We define
two more features for each morpheme. The first is the input category of a
morpheme for finite state automata and the second is the transition state.
These are the third and fourth kinds of information.

2.2 Rules for C o n c a t e n a t i o n

This section explains the rules for concatenation. In our approach, morpho-
logical analysis is performed by applying rules restricting the legal concate-
nation of adjacent morphemes. Before going further~ we define the notation.
We define a Japanese string (string,:In]) as a finite sequence, u : [n] -+ A,
where n is a natural number denoting the length of u, and A is a set of
Japanese characters. Thus, if u : [n] -+ A is a string and n > O, for every
i e In], u(i) is some element A.
A Japanese string as an entry of a lexical dictionary usually has lexieal
ambiguities; thus we define a function, get_entries, which obtains lexicat en-
tries corresponding to a string. As shown in (1) and (2), each morpheme
is a triple, (stringu:[n], A, B), where A and B are the left- and right-hand
restriction for string. We will use SLF to refer to the set of left-hand-side
features of morphemes. The symbol ~ denotes the left-hand-side feature
 191

of morphemes. Similarly, SRF is the set of right-hand-side features of mor-

phemes, and 4,.. denotes the right-hand-side feature of morphem%.

get_entries( siring. :[n])

d4 st~ing,:[n] ~ {mo~p~eme,~:[~]... morpheme~k:[~]} C So,~,~ (1)
where Sent.~= dot{(strmg.:[nj,
. A,B)]A E SLFA B ¢ SRF} (2)
We will use expressions of the form morphemes:In] to denote elements of
the set S~.trie~. It is sometimes convenient to refer to ~ and ~ by the
functional expressions lefff(morphem%:[n]) and righff(morpheme.:[n]).
Restrictions on legal concatenations are expressed by the predicate C~.1~.
The intention is that two morphemes, morpheme~,:[m] and morphem%:[n],
can be concatenated if the right context of morpheme~,:[m] is related to the
left context of morphem%:[n] as shown in figure t.

string, :[m] string~ :[n]

.'." [morp~eme~,:[m]l~ <~, lmorpheme~q:[n][~

string~:[m + u]
r ~i .... I vJ TM

¢==[morpheme~ :[m+n] j :"

string~,(i) = { string,(i) if 1 < i < m

%
string,(i-- m) i f m + l < i < m + n
where ~l~k .,~) = T

Figure 1: Concatenation of morphemes

In the traditional method [11], Cry. is defined as in line (3) and the
functions /e/tf and rightf are realized as in lines (5) and (6). Here, lefff
 192

and righff will return a natural number. Setconnec t in line (3) consists of a
finite number of pairs (Pre, Post) where Pre E SLF(SLF C N) and Post E
SRF(SRF C N) (N is a finite set of natural numbers). In [11], the cardinality
of N is less than 300.

e {T,F} (4)
where rightf(morpheme~:[m])e N (5)
lef~r(mo~pheme~:[~])e N (6)
In this traditional method, the definition of Setco,,~a is difficult for gram-
mar writers to handle, since the same number occurring in sets SLF and SRF
will have different meanings. An element in Set~o~,,~a does not have anything
that makes it easy to handle. However, in the LAX system, we define the
functions righff and leftf, and predicate C~.1~as follows.

= T (7)
,~ leftf(mo~pbe~e~:[m])
~ight~(~o~p~em~ :[~]) c Sro~o (8)
lefff(morpheme~:[n]) c &oo~o (9)
Sf,~,~ in (8) and (9) is the set of morphological features which represents
the relationships between morphemes. In other words, C~le holds exactly
when rightf(morpheme~:[m]) and lefff(morpheme.:[n]) have only one common
element, as shown in line (7).
If the rules are of this form, the process of finding legal concatenation
in LAX can be formalized as a non-deterministic finite state automaton as
follows.

M = (K,E,5, Qo,F) (10)

g d4 { x l x = ({right~,st~i.g~:[.])} (11)
where { rightf} C g f ~ t ~
 193

E a4 { x l x = smng,:[.]} (12)
6( ( q, st ingj : stri,g, :[,]) = U(righff(M),stringk:[mq-n]) (13)
M
where M e get_entries(string~:[n]) A Cr~¢(q, leftf(M))
6(q,¢) = (q, st J.go:[O]) (14)
6(q, z) = U 6(p,x) (is)
where q C K , a E a , x E a*
= @ 5(q,x) (16)
(q,input_string) es
whereSCK, xEa*,q6 F

The NFSA M is defined in (10). K is the set of states of M, and a is the

set of alphabets for M. 3 is the transition function. Qo is the initial state,
and E is the final state.
K is the set of right restrictions : ;- as defined in (11).
a is the set of Japanese strings that are in entries of the LAX lexical
dictionary. The string look-up process in the lexical dictionary is realized as
a T R I E structure, and takes a finite interval of less than 0.5 msec in the worst
case. Therefore, we start our discussion assuming that Japanese strings in
the dictionary have already been found.
The definition of the transition function, 6, is slightly complicated. If the
Japanese string string~:[n] has more than one ~=~=, then 6(q, a) is defined as
shown in (13).
The definition of 6 in (14 ), (15) and (t6) is the same as the usual definition
for an NFSA.
In LAX, plural transition tables ( T r T b l ) can be used as shown in figure
2. Therefore, the elements of S~e~tureare defined as follows.

Sr~ture de=_f {XIX = N a m e _ o f _ T r T b l ( S t a t e ) } (17)

where S t a t e is the name of a state in Trtbl

The multiple transition table is very convenient for representing rules of

derivation and rules of inflection in separate forms. See [12] and [13] for
details.
 194

Figure 2 shows the analysis of the Japanese word "N~5 ~ ~ " [itameruna]
(do not hurt: N~5 ~ = to hurt, ~ = do not). In this figure, the input category
Id,~ of the word stem, ~ (word stem of the verb "hurt"), is accepted in the
initial state, Sdi. Then, the next state is Sdi or Sdj. The next input character
is *5. leftf(*5) is accepted in the state Se~ and the next state becomes Sek or
F, that is, the final state. The next character is accepted in the state Sdk
and the next state becomes Syi or F. Sfz is the state in the inflection table.
The last input, ~, is accepted in the inflection table and the word ~ ~ is
accepted.

Inputs:
i~a me rH n&
\Input
"N~ateg°ry [
States \ v
// I

Idn Ido
Initial ~d~ Sdi V Sdj
!t
: I

Sdi . . . . 5'dk V F

Inflection
tabte
5dj ......... Sj~ v F

Derivation table

Figure 2: Multiple transition table

2.3 E x p a n s i o n of N F S A
The LAX system also provides a "handle" which allows the lexicon-writer to
build arbitrary data structures in parallel with the analysis.
The arbitrary data structures can be attached by assigning the function
shown in (19) to each element of Sf~atur+. This function enables us to attach
some data structures to < : and :::-->; using them, we can write programs
 195

to construct morphological semantic information, which will be described

further in section 2.4.

info( X ) -~ {Some data structures} (18)

where x = V X =

morphemes:In] E S~t,~e~ (19)

2.4 Format of the Dictionary

The syntax of the dictionary in the LAX system is outlined below. Each
entry has the following structures.

begin(Category of morpheme). (20)

SURF :: TrTbla([Stat, ...]) (2t)
&& [TrTblb([Stbl, ..]), ...] (22)
$$ [[in(Input),Common_Procedures], (23)
TrTblb(Stbl, (24)
[pro cedureb11, send(O)], (25)
[procedurebl2, call_sax(02)])... (26)
end(Category of morpheme). (27)
: (28)
Morphemes with the same category are defined between lines (20) and
(27). The category name of morphemes is written in lines (20) and (27),
In line (21), the surface character of a morpheme is written first. ::
is a delimiter. Then, leftf(Morpheme) is written, lefff(Morpheme) can be
regarded as input to the NFSA. If the current state is rightf(Morpheme~)
= { ~ . ~ (word stem for the verb "me"), 9 ~ 4 L ~ " (word stem for
the verb "ma') } and the input is Iefff(Morpheme~)= { ~ ~ } , then
string, :In]will be accepted according to the definition of C~le defined between
(7) and (8).
In (22), righff(Morpheme) is defined.
In (23) to (26), we can write programs which correspond to the next state
transitions.
 196

In (23), common programs can be written. Programs written in this part

are activated when the morpheme is accepted, in(I) is a function to obtain
information from the preceding morpheme, in(Z) is the realization of the
complex function info(lefff(morpheme)) as shown in (t9) and (9).

in( ) %f =info(]eft (Morphe e))

From (24) to (26), programs which correspond to the next state tran-
sitions can be written. The predicate caI1_s~x is used to call the syntactic
analyser SAX. The predicate send is used to attach some information to the
left-hand-side feature of the morpheme, send is info(righff(morpheme)).

out(O) ~f O=info(righff(Morpheme))
When Crule('-~. , ~ : ) is true, information on the right-hand-side and
left-hand-side features is unified.

3 Elementary Algorithm
The elementary algorithm is executed by multi-processes which are layered
in a well formed substring table, as shown in figure 3.
In figure 3, the input string is ~ = ~ [i-ta-me-ru-na] (do not fry). In
this figure, a string of Kana is used as input to the LAX for expository
reasons. The characters are numbered from 1 (~) to 6 (the punctuation
mark 'o ').
Pij represents a search process. F1 to F6 represent the raw numbers of
table 3.
In the LAX dictionary, a process, pnc, looks up morphemes correspond-
ing to the string which starts from the character at R and ends at C-1.
For example, process p23 looks up morphemes which have ~: as the surface.
Search processes p12 to p67 can be executed simultaneously.
Each process is connected through streams (data paths) which the searched
morphemes are passed through. Outputs from the processes at column C are
merged to be the input to processes at raw C. For example, outputs from
p14, p24 and p34 are merged to be the inputs to p45 and p46; p14, p24 and p34
 197

look up morphemes corresponding to the string which ends at ~6; and p45
and p46 look up morphemes corresponding to the string which starts from
~. Therefore, connecting these processes by a stream, we can concatenate
morphemes.
An outline of the elementary algorithm is as follows.

1. Activation of processes

(a) Activate (N(N + 1 ) / 2 ) + 1 processes corresponding to an input

string whi& has N Japanese characters and a Japanese punctua-
tion mark.
(N(N + 1)/2) processes correspond to N Japanese characters.
Activated processes are linked with streams which are indicated
by the arrows in figure 3.
(b) ~ indicating the initial state of the NFSA is put in the stream
to the processes at raw" t.
(c) Data from pgv+l)(N+2) is the result of analysis.

2. An activated process PNK then executes the following algorithm

(a) Wait for input Mi~.

(b) If the input data is nil, indicating the tail of the stream, output
nil and stop. Otherwise, go to the next step.
(c) Look up morphemes Mr~] which correspond to the string starting
from N and ending at K - 1. If M~] exists, go to the next step.
Otherwise, stop.
(d) Call Cr~ to check whether Mi~ and Mr~l can be concatenated or
not. If they can be concatenated, concatenate Mr~I and Mi~, then
output the concatenated morpheme. Go to 2a.

This method makes it possible to perform parallel lexical analysis, al-

though there are still some problems in practical use, as follows.

1. Although ((N + 1) × N/2) + 1 processes will be activated in la, few of

these processes will be able to look up morphemes Mr,l at 2c. There-
fore, the efficiency of analysis will drop because of this useless activa-
tion.
 t98

2. The elementary algorithm cannot handle unknown lexical entries.

3. The elementary algorithm produces each result separately so that the

worst complexity will be as follows.

,~-1 (n - 1)!
=
a ~+1 x (n - 1 - i)!i!
i=O
a(1 + a) ~-1
where a is the worst ambiguity of the morpheme and
n is the character length of the input sentence

Therefore, there should be a packing algorithm. As described in section

5, this reduces the complexity to O(n3) as follows.
n-1
= i x (n - i) x a 2
i=1
a 2 × (n 3- n)
6
where a is the worst ambiguity of the morpheme and
n is the character length of the input sentence
n
A %
r C
C1,C~,...,G, i+l,...,C,
i n-i

4 Practical Algorithm

In this section, a practical algorithm which activates minimal processes, can

handle the unknown lexical entries, and has complexity O(n 3) is represented.
First, efficient activation of the processes is represented. Second, the data
flow with packing is described.

4.1 Making Process Activation Demand-Driven

This sub-section describes an efficient way to minimize unnecessary activation
of processes. At the beginning of the analysis, a 'dict' process is activated
for each input character as shown in figure 4.
 199

tz
i ~a me ru n&
o

1 2 3 4 5 6

F2NN
F3 N N
F4 )R p~

Figure 3: Data flow in the basic algorithm

tz
i ~a me ru n&
o

1 2 3 4 5 6
Z dict p13 p14

F2 dict

F3 dict p35

F4 dict

, F5 dict

Figure 4: Process activation in the practical algorithm

 200

N + 1 'dict' will be activated when an input string has N characters

and a punctuation mark. (In figure 4, six 'dict' processes were activated.)
'dict' looks up morphemes which correspond to the string from the point at
which 'dict' was placed. (For notation, we use function position(dictproc¢s~)to
indicate this point.) For example, the 'dict' process at F1 in figure 4 looks up
morphemes corresponding to strings which start from w in the first column.
If there are morphemes corresponding to wit in the dictionary, the 'dict' at Ft
will activate sub-process P13. Similarly, if v, tc~ is in the dictionary, the 'dict'
at F1 will activate sub-process P14. If there is no morpheme corresponding to
~tc~6 ~, no process is activated. For convenience, the notation refmorph(dict)
or refmorph(PXY) will be used in this paper to represent a set of morphemes
corresponding to the substring to which 'dict' or PXY is attached. A 'dict'
process and its sub-process can be obtained as programs by compiling the
lexical dictionary of the LAX. For example, let us consider the tiny lexical
dictionary shown in figure 5. It can be represented as a TRIE structure, as
shown in figure 6, and then compiled into a 'dict' program, as shown in figure
7. To simplify the representation, figure 7 shows only the look-up processes.

Surface Category
Word stem of the noun ~ (stomach)
[i]
wtc Word stem of the adjective ~i (sore)
[ita]
wtc Word stem of the verb ~ (to hurt)
[ita]
wtcg) Word stem of the verb ~')g) (to fry)
[itame]
~tc26 Word stem of the noun ~ (wood grain)
[itame]
Word stem of the nominal verb ~ (trust, to entrust)
[itaku]

Figure 5: Example of a LAX lexical dictionary

In figure 7, the first argument of 'diet' is the character to which 'dict'

corresponds. The second argument is the next sequence of input characters.
 201

[l,'~, word-stem-of-noun(~/)
[i]
[to, word-stem-of-adjective(@),
Ira]
word- stem-of-verb(~),

[&, word-stem-of-verb(#.)&),
[me]
word-stem-of-noun( )],
[<, word-stem-of-nominal-verb(~)]
[ku]

Figure 6: The LAX dictionary as a TRIE structure

The last argument returns the result of the search.

Let us see what happens when we input the sequence [~, ~, ~, ~]. The
'dict' corresponding to the first character, ~, will accept this sequence.
Therefore, 'dict' will obtain the word-stem-of-noun(~). Next, 'dict' calls
v, ([~, ~, ~],RR,Ir), and this call does not bring any more results, so that no
redundant processes which search for sub-strings such as [~,, ~], [~, ~, &] or
[~, ~, &, ~] will be activated.
With regard to the efficiency of the search, the program in figure 7 is fast
because logic programming languages like Prolog search for clauses using the
first argument hashing. The TRIE structure enables us to utilise this hashing
mechanism easily, making the 'dict' process perform its search rapidly.
As stated above, compilation of the lexical dictionary enables us to min-
imise the number of activated processes, and to look up morphemes quickly.

4.2 Data and Its Flow in the Practical Algorithm

This section describes the data flow between 'dict' and its sub-processes,
then presents a data packing mechanism which makes the complexity of the
 202

dict ( ~ , R , I) :- R = [HIRR],
I = [ [word-stem-of-noun(~)] fir] ,
(H,RR,Ir).
w (~,R,I) :- R = [H]RR],
I = [ [word-stem-of-adjective(~)] ,
[word-stem-of-verb(~)] llr] ,
~ (H,R,Ir).
~( .... []).
v,,'= ( m , R , T ) :- I = [[word-stem-of-verb0:p~5)] ,
[word- st era- of- a-noun (N [~) ] ] .
~ (< ,R,I) :- I = [ [ w o r d - s t e m - o f - n o m i n a l - v e r b ( ~ ) ] ] .
~,~ (_,_, []).

Figure 7: Prolog program for the 'dict' process

practical algorithm O(n3).

4.2.1 D a t a flow
The 'dict' processes described in section 4.1 are connected with each other
through data streams, as shown in figure 8.
The data stream structure is shown in figure 9. This structure represents
the trace of morphologicaJ analysis, and can be regarded as an application
of the layered stream technique.
Unlike the SAX, which analyses inputs according to CFG with the layered
stream technique, LAX does not place any identifier between the rules.
In figure 9, (r-features)corresponds to rightf(morpheme). (array of results)
enables the packing of resultant data, described in the next section. In the
practical algorithm, a different type of data will flow into a 'dict', which
should distinguish the data to be processed in it. For example, in figure
8, the results from P14 will be passed through the 'dict' in lines F1, F2
and F3 before flowing into the 'dict' in F4, in which the results should be
processed. Therefore, in the practical algorithm, each item of data has a
tail position number representing the process by which it should be pro-
cessed. (We use the function position(datad~tj to mark this position.) Cor-
 203

i ta me ru na o

1 2 3 4 5 6
dict p13 p14
•5".~ 4
\
F2 ~dict
\,
F3 "dict p35
\
F4 'dict
k
\
F5 ~dict
N

Figure 8: Data flow in the Practical Algorithm

(data stream) ::= [(array of datal]

(array of data) ::= (data), (array of data 1
(array of data) ::= (data/
(data) ::= [(position),
<r-features),
(results/]
(position) ::= N(1 _< N _< length of input)
(r-features} ::= [(re},..., (re}]
(re} e Sf~,,ure
(results) ::: [(surface of a morpheme),
(morphological category),
(array of results)]
(array of results) ::= []
(array of r e s u l t s ) : : = (results),
(array of results)

Figure 9: Structure of Data stream

 204

[2,[rightf(N)], [ ,word-stem-of-noun,[]]],
[3,[rightf( )],[ ,word-stem-of-adjective,[]]],
[3, [rlgh tf(f word-stem-of- verb,D]] ,
[4,[right@') 5 )], [t' ,word- stem-of- verb,[]]],
[4,[rightf( N )],[ N,word-stem-of-noun,[]]]

Figure 10: Example of data in the stream

responding to position(datad=,=), ea& 'dict' is tagged with the position num-

ber position(dictpro=~ss) when it is activated. For example, in figure 8, the
'diet' in line F1 is tagged with 1 and the 'diet' at line F3 is tagged with 3 at
the beginning of the analysis.
Let us trace tile example computation in detail. First, the token [1,
[end(top)],[]] is given to the 'dict' in line F1. This can be regarded as initial-
izing the NFSA.
The 'diet' at F1 will output the token [2, rightf(N), [N, word-stem-of-
noun,Ill] after calling CruIe to check whether top and word-stem-of-noun(~)
can be concatenated.
The 'diet' in F1 activates P13, and P13 activates P14 in succession. Then,
the data stream coming out from the 'dict' in F1 into the 'dict' in F2 is now
as shown in figure 10.

4.2.2 C o n c a t e n a t i o n w i t h d a t a packing
On receiving the inputs, as shown in figure 10, the 'diet' in F2 picks up data
whose position(datad~) is 2, and outputs data whose position(datad~J is
greater than 2 to the 'diet' in line F3. The 'diet' in F2 will check whether
the data picked up can be concatenated with a morpheme corresponding to
character strings starting from point 2. K there is no such morpheme, the
'diet' in F2 will output [] and stop. Otherwise, the ~dict' in F2 will work like
the 'dict' in F3.
The 'diet' in F3 inputs data as follows.
 205

[3,[rightf(f~)l,[~,word-stem-of-adjective,[]]],
[3,[righff(~)],[~,word-stem-of-verb,[]]],
[4,[righff(~'P*5)],[~')*5,word-stem-of-verb,D]],
[4,[rightf(N H )],[~ ~,word-stem-of-a-noun,~]]

Let us suppose that, using the 'dict' in F3, a morpheme *5 is found as

follows.

ieftf(inttect(~)) f~ rigbtf(word-stem-of-adjective(~)) ¢ ¢
teftf(~) fl rightf(word-stem-of-verb(~)) ¢ ¢
Then, the output from the 'dict' in F3 will be as follows.

[4,[~iehtf((*5)),[*5,inflect,
[ [~,word-stem-of-adjective,[]],
[~,word-stem-of-verb,[]]]]],
[4,rightf(~,)*5),[~')~,word-stem-of-verb,[lt],
[4,rightf(~N )),[~N,word-stem-of-noun,[]]]

Figure 11: Packed data

The result shown in figure 11 is very important. In LAX, combinatorial

ambiguities which occur in concatenation are not propagated to any other
concatenation. If there are ambiguities, they are packed as shown in figure
11. Therefore, the complexity becomes O(n3). Details of complexity will be
discussed in section 5.
 206

4.3 Summary of the Practical Algorithm

This section summarizes the practical algorithm.

1. Activation of 'dict' processes

(a) Activate N + 1 'dict' processes corresponding to each input string

which has N Japanese characters and a Japanese punctuation
mark.
Each 'dict' should have a position(dictvroc~ss).
N 'dict' processes correspond to N Japanese characters.
Activated processes are linked with streams as indicated by the
arrows in figure 8.
(b) righff(top),indicating the initial state of the NFSA, is put in the
stream to the processes at raw t.
(c) Data from the 'dict' at N + 1 is the resutt of analysis.
2. Algorithm for the 'diet' processes
The 'dict' processes activated in line 1 continue as follows. A 'dict'
has many sub-processes, which are connected with each other through
communication streams, as shown in figure 12. Each sub-process can
be characterised by its input and process and its output to its connected
processes. From this line, the function input(Data) represents the data
input of each process, and the function output(Data, ProcessN) repre-
sents the data output to the Process whose number is N. Numbers in
figure 12 correspond to the following item numbers.

(a) Do input(Data).
If Data is the tail of a data stream, do output(tail, Process2b) and
stop.
Otherwise, do output(Data,Process2b).
Go to 2a.
(b) The fiIstr process
Do input(Data).
If Data is the tail of a data stream, do output(tail, Process2c),
output(tail, Process2d), ou~put(~ait,Process2~), and stop.
If position(Datad.,J is equal to position(dicQ.,J,
do output(Data,Process o) and ou*put(Data,P ocess, ).
If posi~ion(Datad.t.) is greater than position(dictda~J,
do output(Data, Process2J.
Go to 2b.
 207

(c) The filcon process

input(Data).
Check whether Cru~(Data, re[morph(dict)) is true or not. If it
is true, concatenate Data and refmorph(dict) to make new mor-
phemes and output(new morphemes, Processd~t:pTo:5). Go to 2c.
(d) The PXY sub-process
i. Do input(Data).
If Data is the tail of a data stream, output(tail,Processad)~ J
and stop.
Otherwise, do output(Data,Process2(a)~), recursively call PXY,
and do output(Data,Processdi¢t:pTo:4).
Go to 2(d)i.
ii. Do input(Data).
If Data is the taitof a data stream, do output(tait,Process2(d)~)
and stop.
Check whether Cru~,(Data,refmorph(sub_pro)) is true o r not.
If it is true, concatenate Data and refmorph(sub_pro)to make
new morphemes and do ou tput(new morphemes,Processd~a:p~o:5).
Go to 2(d)ii.
(e) Do input(Data).
If Data is the tM1 of a data stream, and it is the first input data,
then process the unknown word, else do output(Data,Processth~,td~t)
and stop.
Otherwise, do output(Data,Processthent~td~a).

"l Input j ~

IUnknown words
.......

filstr

I l 1--
. . . . . . . . . . . . . . . . I- ...

1 .... t
Figure 12: 'dict' procedure
 208

5 Complexity
This section discusses the complexity of the elementary and practical algo-
rithms in sequential and parallel execution.

5.1 Complexity in Sequential Execution

First, the elementary algorithm has the following complexity. In the following
expressions, n is the number of input characters and T is the total arrlount of
time required to obtain all solutions, so that complexity is that of T related
to n. We assume that any input substring corresponds to a morphemes.
All adjacent morphemes can be concatenated, d is the time required for
dictionary search, c is the time required to check and concatenate a pair
of morphemes, i is the number of delimiting positions, as shown in the
chart under the expressions, a ~+1 x ,-1Ci is the time required to check
and concatenate all the pairs of morphemes when there are i delimiters in
character strings.
Morpheme look-up is realised with the TRIE structure, and the search
time for a morpheme of m characters takes d x m where d is the time required
to trace the TRIE structure. Thus, the total look-up time for a sub-string
contains j input characters, d ~=1 i(n - i + 1). Thus, we have:

n--1 n

T = c ~ ( a i+~x , - l C i ) + E d j ( n - j + l ) (29)
~=0 j=l

= c x a ( 1 + ~ ) ~ - 1 + d ~ ( ~, + 1 ) ( ~ + 2)
6
wherea, n > _ l , c,d>O
o(r) = O(a n) (30)

and thus,

morphemel morpheme2 morphemei+ l

C l , C2, . . . , 0 o , C o + l , . . . , Cb , ..., C , ~ + l , . .. , C ~
1 2 i
 209

There should be a limitation on the morpheme length so that the max-

imum character length of a morpheme can be set to I. If l is greater than
n, the complexity becomes O(a'O. Otherwise, it becomes as follows. It is
impossible to obtain one expression for 31 because we cannot obtain any
expressions for polynomials whose maximum degree is ~n'.

T = cA. (31)
wheFc
I
A~ = EA~-, (k > 0 (32)
i=1
Ak = cak(k < t) (33)
where c > O, k > l

Then the complexity of the practical algorithm is as follows:

n--1 n
T = c~(ix(n-i) xa2)+~dj(n-j+l)
i=l j=l

= a 2 x (n 3 n) + dn(n + 1)(n + 2)
- -

6 6
0 = O(n 3) where a is the worst ambiguity of the morpheme and
n is the character length of the input sentence
1¢

i n-i

If the m a x i m u m length of the morpheme is I, the complexity is given by:

n t
T = c~a2+d~j(n-j+l) (34)
i=1 j=l
= C X a2 X n + d(1 + I)l/2 × n + K (35)
1 l
where K = - d ~ j ' +~ (36)
j=l i=l

0 = 0(~) (37)
 210

The conclusion to be drawn is clearly that packing is critical to the algo-

rithm.

5.2 Complexity in Parallel Execution

We define the time complexity of a parallel execution in terms of the time
from the beginning to the end of parallel execution. In parallel execution,
the dictionary search process can run simultaneously, so that the search time
is proportional to the maximum length of morphemes. However, the con-
catenation processes cannot run simultaneously, because a process suspends
itself until the preceding process outputs data to it. Therefore, in parallel
execution, the time complexity will be proportional to the input length.

6 Algorithm for Sequential Logic Languages

This section describes an implementation of the LAX algorithm in terms of
the sequential logic programming language, ESP. ESP, which can be executed
on the PSI or PSLII machine, is an object oriented Prolog and has an array
structure called the stack vector. Using the stack vector, we can access each
logical variable with an index.
Due to the execution time and the memory efficiency consideration, it is
not a good idea to simulate the layered stream mechanism sequentially for the
implementation of the LAX algorithm. We have therefore adopted another
mechanism, the tableau method, in place of the layered stream mechanism.
In this method, a tableau stores the analysis, and analyses breadth-first.
This method is also a sort of parallel algorithm, which our implementation
simulates sequentially.
First, the sequential version of the LAX algorithm is explained. Next,
the result of efficiency evaluation of this algorithm is reported.

6.1 Sequential Algorithm

This algorithm consists of two phases. In the first phase, it looks up all the
morphemes that occur in the input sentence and stores them in a stack vector
(tableau). These morphemes are the only candidates for the morphemes that
belong to the correct interpretation of the input sentence. This phase is
 211

implemented as a clause indexing search on the programs compiled from the

TRtE structure. Fundamentally, this part of the algorithm is the same as in
the parallel version.
In the second phase, the algorithm attempts to connect two adjoining
morphemes. If analysis fails in some position, namely, if there are no two
morphemes connected with each other at that point, it calls the error recovery
process that connects the last two morphemes. We can therefore obtain an
almost correct analysis even when there are some u n k n o w n words, or words
that are not entered in the morphological dictionary, in the input sentence.
This enables us to process sentences flexibly that contain unknown words.

O The sequential LAX Mgorithm:

• I n p u t : a string of n characters
• Output: the result of morphological analysis (a packed structure)

• First phase:
Make a stack vector of size n + 1, and set the right-hand feature of the
start of the sentence and left-hand feature of the end of the sentence to
the first and last positions of the stack vector, respectively. For each
number i (1 < i < n), do the following.

o Look up all the morphemes that start from the ith character
and register them in the stack vector at t h e / t h position. Each
morpheme has two kinds of features, the left-hand feature
and right-hand feature, and is represented as a list structure
with its tail unbound.

• Second phase:
For each number i (1 < i < n), do the following.

o For each morpheme of size l stored at the ith element of

the stack vector, pick up the morphemes at the (i + 1 - l)th
element of the stack vector. If two adjoining morphemes can
be concatenated, unify the left-hand morpheme with the tail
of the right-hand morpheme.
 212

K no morpheme can be connected to any adjoining mor-

pheme, and if no morpheme stored at less than the ith posi-
tion is connected to the morpheme stored at greater than the
ith position, then select one morpheme which is connected
to the morpheme stored at less than the ith position and
connect it to one of the adjoining morphemes.

If there is no sequence of morphemes from the start of the sentence

to the end of the sentence, that is, anMysis fails, search the missing
morphemes backwards and connect them so that there is at least one
sequence of morphemes.

6.2 Evaluation of Sequential Algorithms

This section reports the evaluation of the sequential LAX Mgorithm. We
adopt two texts as the test sentences. One is a biography F ~ - ~ ) - ~ ) k J
("Madame Curie"), ~nd the other is an article about nature conservation
entitled r N ~ q ~ j ("Nature Conservation"). These are in the textbooks
for the fifth and sixth grades of elementary school, respectively. Appendix A
shows an extract of the text FQ~,k.q'~j ("Nature Conservation") for eval-
uation. Table 1 shows the size of the texts and of morphological dictionary
entries. We use the same affix dictionary to analyse each text, and provide
a word stem dictionary for each text.
The analysis of these texts was performed on a PSI-II machine with the
SIMPOS 4.1 operating system.
Appendix B is part of the result of the analysis for the second text r N ~ :
~ j ("Nature Conservation"), and shows the length (number of characters),
number of ambiguities and analysis time for each sentence. LAX can analyse
each sentence in less than 100 msec even when there are many ambiguities.

Text F4~:~ ~ -~K)kJ F~ , ~ ~j

"Madame Curie" "Nature Conservation"
Number of sentences 126 82
Word stems 780 450
Derivational and
inflectional affixes 520 520

Table 1: Contents of the test texts and their dictionaries

 213

This result shows that the LAX sequential algorithm can analyse Japanese
sentences in linear order.

7" C o n c l u s i o n
This paper described the LAX algorithm, its time complexity and its imple-
mentation on a sequential logic based machine. It also reported the evalua-
tion of the result of analysis for two types of Japanese sentences. We believe
that there are two areas for further research. One is to refine the algorithm to
allow a more precise retrieval of unknown words. The second is to implement
the parallel LAX algorithm in terms of GHC.

A Sample Sentences
Extracts of the text F~ * ~ ~ J ("Nature Conservation")
 214

B Result of Analysis
S a m p l e text: FR~'t'~J ("Nature Conservation")

Sentence Number of Ambiguities Analysis time

number characters (reset)
46 57
g0 56
47 43
24 23
5 61 58
6' 64 62
31 2 34
59 2' 67
9 56 58
10' 65 12 85
11 64 89''
12 57 70
13 16 2 24
14 18 19
15 20 15
16 26 49
17 21 17
18 45 1 46
19 26 1
20 21 1 14
21 45 3 42
22 27 1 32
23 24 32
24 43 3 34
25 39 1 46
26 33 4 44
27 38 4 37
28 63 6 81
29 66 1 L 88
30 60 74
 215

References
[1] Akasaka, K., Sugimura, R. and Matsumoto, Y., "Chikuji
Shorikei-j8 deno LAX no Mondai-ten to sono Kaiketsu-h6"
("Problems in LAX in sequential implimentation, and their So-
lutions" (in Japanese)), Proceedings of the 35th Conference of
Information Processing Society of Japan, pp.1355-1356, 1987
[2] Aizawa, T. and Ebara. T., "Mechanical Translation System of
Kana Representations to Kanji-kana Mixed Representations"
(in Japanese), NHK Gijutsu-Kenky~t, Vol.25 No.5, pp.23-60,
1973
[3] Kamiwaki, T. and Tanaka, H., "Jisho no TRIE KSz6-ka to
Jukugo Shori" ("Idiom Handling Using the TRIE Structure
Dictionary" (in Japanese)), Proceedings of the Logic Program-
ming Conference '85, pp.329-340, 1988
[4] Kay, M., "Algorithm Schemata and Data Structures in Syntac-
tic Processing", TechnicalReport, CSL-80-12, XEROX PARC.,
Oct. 1980
[5] Mizutani, S., "Kokubunp5 ShibyS" ("Sketch of Japanese Gram-
mars" (in Japanese)), Bunp~ to Imi I, Asakura Nihongo K6za
No.3, Asakura Shoten, 1983
[8] Okumura, K. and Matsumoto, Y., "Layered Stream wo Mochi-
ita Heiretsu Pogramming" ("Parallel Programming using Lay-
ered Stream" (in Japanese)), Proceedings of the Logic Program-
ming Conference '87, pp223-232, 1987
[7] Sugimura, R., Akasaka, K. and Matsumoto, Y., "Heiretsu
Keitaiso Kaiseki System LAX no Jitsugen" ("Imprementation
of Parallel Morphological Anatyser LAX" (in Japanese)), Pro-
ceedings of the 35th Conference of Information Processing So-
ciety of Japan, pp.!323-1324, 1987
[8] Ueda K., Guarded Horn Clauses, ICOT Technical Report
No.103, 1985
 2t6

[9] Earley, J., "An Efficient Context-free Parsing Algorithm",

CACM, Voll3, No.2, pp.94-i02, 1970
[i0] Matsumoto,Y and Sugimura, R., "Ronri-gata Gengo ni Mo-
tozuku K&bun Kaiseki System SAX" ("Logic Programming
Based Parser SAX" (in Japanese)), Computer Software,Vol3,
No.4, 1986
[tl] Yoshimura,K., Hidaka, I. and Yoshida, S., "Nihon-go Bun no
Keitaisokaiseki ni okeru Saich6-itchi-h6 to Bunsetsu-sfl Saish6-
h6 nitsuite" ("SaichS-itchi and Bunsetsu-sfi Saish6-h6 in Mor-
phological analysis " (in Japanese)), IPSJ-NLU 30-7, 1982
[12] Morioka, K., "Goi no Keis.ei" ("Formation of a Vocabulary" (in
Japanese)) Gendai-go Kenky~t Series 1, Meiji-Shoin, 1987
[13] Sano, H., Akasaka, K., Kubo, Y., and Sugimura, R., "Go-KSsei
ni Motozuku Keitaiso Kaiseki" ("Morphological Analysis with
Derivation and Inflection" (in Japanese)), Proceedings of the
36th Conference of Information Processing Society of Japan,
1988

[14] Mukai K., A System of Logic Programming for Linguistic Anal-

ysis Based on Situation Semantics, Proceedings of the Work-
shop on Semantic Issues in Human and Computer Languages.
Half Moon Bay. CSLI 1987
 Extraction of Characteristic Facts and Abstract Generation

Naoyoshi T A M U R A , Tomomi K A M I O , Masato K U R O S A K I

Yokohama National University

156 Tokiwadai Hodogayaku Yokohama 240, J a p a n

Abstract

This paper presents a formalization of the summary generation process, which takes a set of observed
facts as input and generates its summary in the form of natural language text. Our assumption is that
human summarization process consists of the following four phrases : (1) reconstructing the original
semantic structure in his/her mind from observed facts, (2) extracting the characteristic facts from the
reconstructed structure, (3) constructing the surface structure of text in the macroscopic view and finally
(4) generating a natural language text from the surface structure. Our system generates newspaper-like
texts in the domain of baseball games. The implementation is also briefly presented.

1 Introduction

In this paper we present a formalization of the summary generation process, which takes a sequence of
observed facts as input and generates a summary in natural language text from the facts. We also present
a part of its implementation which is written in C-Prolog.
Among the wide current studies of text or story understanding, text summarization is one of the
difficult themes which require highly advanced processes of semantic analysis. In summary generation, the
mechanism to decide whether a fact is important or not is primarily significant. The following approaches
seem to be typical: (1) an approach based on the frequency of key words, (2) an approach based on scripts
[7], (3) an approach based on the connective relation between sentences [2,8], (4) an approach based on
the story grammars and rules for summarization [6]. We assume that the human process of extracting
characteristic facts can be described only with semantic information manipulation. Therefore we focus on
semantic information processing and consider the input information, which is a series of facts (we will ca~
this facts observation), to have no surface structure. And we restrict the domain of summarization to be
so small that sufficient understanding and deep inferences can be performed for as many cases as possible.
Our system basically uses summarization rules in the extraction process. However, there are some text
generation systems which decide the degree of refinement of generated texts with the importance decision
mechanism based on semantic analysis [1,3]. In our system the semantic information is represented in
networks. We aim at construction of the human model of interest~ the importance decision mechanism and
the summarization model by arranging all of the networks.
The world to which we apply our summarization system is that of games with wins and defeats,
especially the world of baseball game. The reasons why we chose this are the overall goals of a group
 218

(to win the game, to hit a ball at a chance, to defend at a disadvantage) is obvious, it is possible to
appropriately formalize the game by simplifying rules of baseball and suificient inferences axe possible in
the small domain.
Our aim is to generate sentences in aewspaper-hke style which we think it is easy to deal with outwardly,
not only because we want to lighten a burden on the generation process but aJso because our final object
is to build a sports journalist robot who can write news from the scorebook of a game.
In chapter 2, we present a formalization of the extraction process in the most abstract model, and we
extensively study the model of human interests, the existence of a reader and the hierarchy of semantic
representation. In chapter 3, we concretely reconsider the extraction process again for an implementation.
Especially, we study the similarities of our importance decision mechanism with the connectioulst model.
In the final chapter, chapter 4, we present an implementation of our system and some inner representations
of knowledge or information. We also explain each sentence of generated texts.

2 A model of summarization process

In this chapter, we present a model of the human summarization process which comprehends a sequence
of facts and generates a text as a summary.

2.1 A m o d e l of e x t r a c t i o n

First, we study the process of summary generation and present a model of extraction.
Given information about a series of facts~ we divide the summary generation process into three phases.

1. Deletion of uninteresting information and extraction of inteIesting infoImation: a deletion phase of

uninteresting information and information from which any other interesting information cannot be
derived.

2. Transformation to "the macroscopic structure " (the surface structure) and the text planning: We
will call a structure stating assertions of a mass (members participate in events) in the macroscopic
view "the macroscopic structure '~. The macroscopic structure is reconstructed from facts of each
element in order to generate more natural sentences. In our objective world, the world of baseball
games, structures collectively describing a movement of a team or describing win or loss of a team
are macroscopic ones, whereas descriptions of each member's movement are microscopic ones. Topic
of paragraphs or documents correspond to such structure.

3. Generation of surface sentences: We consider the text generation phase as an isomorphism except for
a rhetorical transformation.

We define the process of the extraction of characteristic facts as in Definition 1. In the definition,
each domain of discourse is a set of information, more concretely speaking, which corresponds to a set
of assertions and rules about facts. And we think the deletion and the extraction of information are
 219

influenced "persona] interests'~ so we deal with human interests as a parameter in the process of the
summary generation.

Definition 1 (Extraction process)

The extraction process is defined as followings.

1. Let F be a set of information about series of original events to be considered, the act of observation
is an apphcation of the function o to F. We define the result O to be an observation.

o = 0(F)

2. The reconstruction R of o~igina] events is constructed by an application of the function r to the

observation O and the knowledge K of an objective world. By this application some obvious facts
which were deleted at the observation are suppfied.

R = r(O, K)

3. The set S of characteristic facts is constzucted by an application of the function s to the reconstruction
/ / a n d paxameter P of human interests.

S = s(R, P)

In our apphcation of this definition, the knowledge K of an objective world consists of rules of baseball,
some information about each player and so on. The outline of the summarization process is shown in Fig.1.
Though we wi~ mention in more details in section 3.1, the input of the system is currently the observation
O.

Let us consider "the quantity of knowledge" as the number of assertions in a domain, where each
element is an assertion for a knowledge. Fig.2 shows the domain and the region of each function (i.e. set of
origiital events F, reconstruction R, observation O and characteristic facts S). Next, we study the relation
of each domain.

1. The observation O is included in the set F of original events, because O has less information than F.

OCF

2. The reconstruction R has more information than the observation O because R is made from O by a
supplementation of some obvious facts.

OCR

3. The set S of characteristic facts is a subset of the reconstruction R.

From Fig.2 we find the followings. (O, R stand for the complement set of O, R, respectively.)
S N O is the important information from the observation.
R N 6 is the important information supplemented by the inference and is obvious to the observer.
 220

function o

(
Fig.2 Inclusion relation }n the domain

Fig.1 Model ol extraction process

S n k is the information of human interests supplied at the extraction process.

Definition 2 represents a transformation from a description of each event to a description of re-structured
in macroscopic view.

Definition 2 (Transformation to the macroscopic structure)

A macroscopic structure (surface structure) M is generated by function m from a set of char-

acteristic facts S and a parameter P describing human interests.

M = m(S, P)

The parameter P of human interests indicates what standpoints the macroscopic structure is viewed
from, for example in a baseball game, P is used to indicate whether one scene is favorable to one team or
not favorable to the other team.

2.2 Model of human interests

The parameter of human interests, which appeared in Definition 1, is referred at the characteristic facts'
extraction and the transformation to the surface structure. This parameter was introduced to model the
human interests (or view points) which a man referes at writing text from a series of events. Hereafter, this
we will call this model She model of(human) interests, or simply interests. The model of human interests is
 221

represented by a set of items about what the man is interested in. The model of human interests depends
on the objective world of summarization. Also, there are some common interests among people and some
kinds of personal interests of a man who generate summaries or surface structures, although both view
points are not strictly distinguishable. We show an example model of human interests, which is represented
in natural language. The items are selected for the summarization of baseball games, and ace partially
implemented as rules of a production system.

[The m o d e l of h u m a n interests]

1. Common view points among almost people

(a) the issue of the game

(b) score (tie, reverse, home run, etc.)

(c) activity of favorite players

(d) evolution of severely changing game

(e) unexpected scene

(f) points-expected scene

2. Specific view points of summary writer

(a) the issue of the favorite/notfavorite team

(b) score or chance of the favorite/notfavorite team

(c) activity of favorite/notfavorite players

(d) standpoint (either offense of one team or defense of the other)

In this model, (1) seems to be common interests among almost people. (l-d) and (l-e) relate to a
method in which events not matching a script iemain in the summary. We thought, items in (2) represent
a personahties of the writer who 5kes/dislikes a specific team. On the other hand we can say that items in
(2) also represent common interests in a sense the he has favorite/notfavodte particular one team. In this
sense we can think of models of specific interests focused on runnings or, extremely speaking, the second
base. In actual newspaper articles of professional baseball league, victory in a season and the record of
players usually become a major subjects of discussion. To generate newspaper-like texts, more precise
models of interests are necessary, in which time transitions and the relation to other games have to be
taken into account. However we left these p~oblems to be future works.

2.3 Existence of a reader

The process told in Definition 1 proceeds from the inference of original events to the extraction of character-
istic facts. We think such a process of summarization corresponds to one in which the summary writer has
 222

his own standards of extraction and he doesn't recognize the existence of a reader. On the other hand there
may be another style of summary generation, i.e. the summary writer has a reader model and through the
model the writer proceeds with the summarization while he is considering reader ~s ability of Ieconstructing
the points of original events from the summarization. In this section we present the summarization process
with areader model.
At first, we abstractly redefine Definition 1. Let i be a function which has the domain and the region
both on the world of knowledge K(R, O, S E K) and infers a set of information as much as it can infer.
Conversely let s be a function to get summaries from a set of information.

i,s:K~K

Then the process of Definition 1 can also be represented by the following way:

R = i(O) {reconstruction from the observation O}

S = s(R) {summarization from ~he reconstruction R}

where the writer generates summaries on the base of his interests only.
Now
R' = i( S)

represents one more reconstruction from the summary by inference i~ and

S'= s(R')
represents one more summarization from the re-reconstruction.
We next dei~ne a function s~ which repeats reconstructions and summarization n times.

s0(x) - s(i(x))

~÷1(*) -~ s(i(s,(*)))

And moreover let S~ be equal to s~(O) i.e.,

s~ - ~ . ( o )

then the sequence S h S.% S~,... corresponds to the polishing of summaries.

Fig,3 represents the model of summarization process considering the existence of readers. In this
summarization model the summary writer knows the existence of readers, ltowever, the writer thinks
readers are interested in the same points that the writer has and readers can infer in the same way that
the writer does.
We can study more precise models of summarization. For example, distinguishing interests of a reader
(or the knowledge of a reader's world) from that of the writer will make a model with more ~espect for
readers. In these summarization models deletion is the only operation, however, summaries are to be
polished not only by deletion of information but also by restoration of deleted information. To formalize
the restoration, extension of inference mechanism, such as ~eference to the reconstruction R which is the
information with no deletion performed, is necessary.
 223

inference i

summariaztion s

Fig.3 Model of summarization process considering

the existence of a reader.

2.4 T r a n s f o r m a t i o n to m a c r o s c o p i c s t r u c t u r e

Each event in the reconstruction of original events were described separately. However, the information
represented by several sentences usually can be processed together. Moreover, when the summarization
process has finished, the whole text is often gathered into one sentence. Thus there is an hierarchy in
representation level. We call this hierarchy the hierarchy of representation. We divide this hierarchy into
three levels.

1. fine description level: each sentence in the text summarization and each player's movement in our
application.

2. middle description level: a summary fo~ one paragraph in text summarization and topics in a game
in our application.

3. inclusively description level: a theme in the text summarization and win or loss in our application.

3 Extraction of characteristic facts

In this chapter, we formalize more concretely ~he obsezvation which is the input of the system, the inference
process of original events and the extraction process of characteristic facts.

3.1 Definition of the observation

At first, let us define what is the input to be summarized. A series of events to be summarized is a sequence
of facts in which a group cooperates to accomplish a goal, for instance to win a baseball game. The input
is defined as follows. This is called set of events observation.
 224

Definition 3 (Observation)

1. An evenl is a triple of a time point, an agen$ and an action. Therefore the domain of event is

Evt = Tp x Agt x Ac~

where Tp stands for an ordered set repiesenting time points (this relation is denoted by e.g. t l < t2),
Ag¢ stands for a finite set of agents and Act stands for a finite set of actions.

2. An observation is a set of events.

O = {evtl,evt2,...}

In our system, which deals with the world of baseball games, Agt corresponds to a set of players of
both teams, and Act to a set of simplified action.

Act = {hit(l), hit(2), hit(3), homerun, ou~, gethome . . . . }

Note that all facts that occurred in the objective world axe not entered into the system. The sequence
of events is entered into the system through some kind of means such as an observer. In this style of
input we must recognize the existence of a filter on the objective world. But viewing the observation as
abstractly represented original events, we take the existence of a filter as a future work.

3.2 Semantic network and inference

In this section we define the form of semantic network and the inference process for original events as a

function on the network.

Definition 4 (Semantic network)

1. A set of nodes N consists of a set of names W and a set of vaJaes D.

N=WUD

2. A set of link names Lbl is given as follows.

Lbl = {case labels} U {attribute names} U {next, isa, state A, state B , e f f e c t , . . . }

3. A set of link L is given as follows.

L = {(nl,l, n2) l ul,n2 e N,t e Lbl}

4. A knowledge K is represented in the network as follows.

K = (N', L')

wheIeN tCN,L ~CL.

Given an observation O, the initial semantic network is built as following definition.

 225

Definition 5 (Initial semantic network)

For each event in t h e observation O, i.e.

(tp, agt, act) • O,

there exists an event node n a m e d e, and the initial semantic network/to = (No, L0) h a s the following links.

1. (e, isa, act) • L0,

which represent t h a t event e belongs to a class act.

2. (e, agent, agt) • Lo,

which represent t h a t an agent of event e is agt.

3. (e, stateB, si) • Lo,

which represent t h a t the s t a t e node before event e occurs is st.

4. (e, stateA, si+l) • L0,

which represent t h a t the s t a t e node after event e occurs is s~+l.

5. ( s . next, s~+l) • L0,

which represent t h a t there is a relation next (relation of temporal ordering) between s t a t e node si
and Si+l, provided t h a t

tp~ < tp < tp/+~

for the event (tp,agt, act) where time points of state nodes sl and Si+l correspond to tpi and tp~+~,
respectively.

6. (e, next, e') • Lo,

which represent t h a t there exists an event e ~ after the event e (relation of temporal ordering).

W h e n s o m e events occu~ in t h e same t i m e point, we treat t h e m as a group on t h e network a n d consider

these events as one event. Trivial facts and inferable facts from other facts are usually o m i t t e d from
t h e observation when given as an input. Bu*, we a s s u m e t h a t facts represented as a group include all

exceptional related fa~ts. This is because exceptional related facts are difficult to reconstruct. Therefore,
there is no inference for g~oup facts (Definition 6, 3-a).

T h e inference on t h e initial semantic network is performed as follows.

Definition 6 (Inference process)

For a initial semantic network Ko = (No, Lo), we define the process o£infexence as a construction of
semantic network

K = (N', L')
according to 2 a n d 3 given as follows.

1. For a s t a t e node s, we call a set of h n k

s* - {(s, l, v) ] 1 E {attribute names}, v e N}

the state ors.

 226

2. Attach a set of link

S~* = fs(e, s~-l)

to each state node in the initial semantic network K0, where function S~* = fs(e, s~-l) maps occurring
event e and state s~-l*, which is the state before event e occurs~ to next state. And

(e, sta~eB, si-1) e Lo,

(e, stateA, sl) E Lo,

(si-l,uex~,s~) e Lo.

3. Replace event node e in the initial semantic network Ko by a set of node re(e, si), where sl stands for
a state before event e occurs (which is indicated by a link named s~a~eB), and function fe is defined
as follows:

(a) When e is a group event node,

fe(~, s~) = ~,

i.e., inference is not performed.

(b) Otherwise
re(e, s~) = {events derived from even~ ¢ a~ s~a~e s~}

Fig.4 show an example semantic network in our application. Some of the node names and link names
which are specia~y used in the net are explained as follows.

1. nodes
event~N represents an event node.
state~N represents a state node related to an event.
team~N represents an object at middle description level.
game~N represents an object at inclusively description level.

other atoms

2. links
effect links an effected node.
cause links a cause node.
other case labels

The situation presented by Fig.4 is as follows: at second half of the first inning, when Shinozuka and
Matsumoto are on the second and the third base, respectively, batter tiara makes a timely two base hit,
then Shinozul~ and Matsumoto come back to home base and turn the table to score 1-2.
Note that this scene is extracted as a cha~acterlstic facts because turning the table is in the model of

interests told at section 2.2.

 227

1-0
• [2:shinozuka,3:matsumoto]

l : S e ~ r ning ~ e \ ~' l'~,~seore runners [1:l'~ra]

stateA stateB stateA stateB

Fig.4 An exampleof semanticnetwork.

3.3 E x t r a c t i o n process of characteristic facts

In this section, we formalize the extraction process of characteristic facts. The principle of extraction is
based on weighting on the network: ]) Select significant items according to the model of human interests
mentioned in section 2.2 and give them initi~, vaaues of importance, 2) Compute new value of importance
for each node by summing up the multiplication between values of importance of those nodes which connect
to the node and the predifined weights of links between them.

Definition 7 ( I m p o r t a n c e decision function on semantic network)

T h e i m p o r t a n c e decision function p on a semantic network

K = (N', L')

is defined as follows.

1. Define the first importance decision function

po:N~R

where R is the domain of real number.

2. Define weight for each rink n a m e

w : Lbl-~, R.
 228
1-0
[2:shinozuka,3:matsumoto]
1:second ~
score
t 1-2 [1:hara]
nnmg score runner
next~
stateB

agent~is a

the giants

( " rnm~home
ag~sa~
' ~ ~ shin~oVzuka
2nV~a gethome

Fig,5Anexampleofsemantnet
ic work
inmoremacroscopivicew,
For each link (u{, l, nj), let the weight of the ]ink be

3. the m + 1st importance decision function pm+l : N --* R is defined by

4. when
tp~+~(~) - p~(~) i2<

for some threshold number 5, we define pro(u) to be the importance decision function p(n) for semantic
network K = (N', L').

We ca~ think our extraction mechanism of characteristic facts on the connectionist model. Making
importance factors correspond with activity levels on neural net, the computation of importance decision
function p is very similar to that of activity level of connectionist model. On the other hand, we use the
model of human interests to initialize the importance factors, however, assignment to the weight of hnks
is also thought to be a kind of the model of human interests. These problems are left to be further works.
 229

3.4 Transformation to macroscopic structure (planning)

In order to generate natural and readable texts, transformation flom microscopic description to macroscopic
structure and how to make that structure into text (planning) is very important. So far, we considered the
extraction process and the transformation process separately. However, in the implementation, it is more
effective to use the information obtained at the extraction of characteristic facts for the transformation
to the macroscopic structure or planning. Moleover, theoretically, we think that such transformation to
macroscopic structure has its meaning in the process of extraction. Because of these reasons, formalization
is not sufficient. In chapter 4 we will show several examples of the transformation using experimental
results.
Fig.5 shows a transformed macroscopic structure obtained from that of Fig.4. In Fig.4 events are found
on the focus of each player, but in Fig.5 the view point is moved to higher level, i.e., the team level, where
~he agent of the representative event (event#1000) is a team (team#]0). Fig.5 corresponds to the sentence
"The Giants turned the table by ttara's two-base hit ..."

4 I m p l e m e n t a t i o n and e x p e r i m e n t

In this chapter we present our summary generation system based on the formalism shown in chapter 2, 3,
and we also present texts generated by the system from observations of (imaginary) baseball games.

4.1 Implementation

Our system is built on a Prolog system and the semantic network is represented in DCKR [5]. An example
of network in DCKR is as follows, which corresponds to event#6 in Fig.4. ~#' and ':' found in the followings
are infix operators (xfx) with precedence 200 and 250, respectively.

sem (event #6, s t a t e B : s t a t e#30).

sem(event#6, s t a r e a : s t a t e#31).
sem (event#6 , n e x t : e v e n t # 9 ) .
s em (event #6, agent : h a r a ) .
s e m ( e v e n t # 6 , 2 s a : h i t (2)).
sern (event#6 , e f f e c t : e v e n t # 7 ) .

Searching is quite simple by DCKR representation, such as omission of inverse links. Note that our
DCKR representation is simple because searching doesn't need any complex operations on the isa hierarchy.
Rules of inference for the reconstruction (Definition 4,5,6) are written in production rule format and
we implemented a very simple production system for them. The format of the rule is

pattern on the semantic network = = > t r a n s f o r m a t i o n on the network

pattern on the semantic network is written in a conjunction (or disjunction) of prolog clauses in

DCKR format, and t r a n s f o r m a t i o n on the network is mainly to assert a new relation of objects. Even
 230

though for simplified baseba~ rules mentioned in section 3.1, production rules are so comphcated that we
will not mention them any more.
We defined the importance decision function in Definition 7 and mentioned the relation to the connec-
tionist model, however, the function has not yet implemented.
Grammar for the generation of surface sentences consists of about 40 DCG rules, whose nontermi-
nal symbols have difference lists and a pointer to a macroscopic structure on the semantic network as
parameters.

4.2 Parameters of the system

We have parameterized several concepts in the summary generation process as follows.

1. Common view points of human interests (eg. score or the issue of the game).

2, Personality of summary w~iter (eg. favorite player/team).

3. Existence of a summary reader (i.e. whether to make another inferences or not).

4. Importance decision function (i.e. weights of links and initial values of nodes).

Here we have implemented two models of interests as common view points, that is 1 in the above list.
Other parameters are not implemented so the system generate summaries at the neutral standpoint and
pays no attention to the summary readers.

[ I n t e r e s t s 1]

1. turning the table

2. same points

3. getting ahead

4. "if hit ..."

5. three consecutive strikeouts in a pitcher's game

6. home run

[Interests 2]

1. turning the table

2. winning hit

3. three consecutive strike-outs in a pitcher's game

4. home run
 231

Each item represents following situations:

"same points" is a situation where both teams get to the same points.
'~getting ahead" is a situation where a team gets point(s) after %ame points" situation.
~turning the table" is a situation where losing team defeats another.
"if hit ..2 is a situation where one team would get some points if a hit were made.
~%hzee consecutive strikeouts in a pitcher*s game ~ is a situation where a pitcher gets three consecutive
strikeouts in a pitcher's game. Here, we define a pitcher's game is the game in which the total score of
both team is less than two. This is based on the assumption that strike-out is a characteristic fact in a
pitching game though it is not in hitting games.
"winning hit" is a situation where that hit causes the victory.

4.3 Input and generated texts

Next we show a part of input (the observation). In the input

event (28, [ ( o u t , c r o m a t i e ) , (out (2) ,yoshimura) ] )

represents that %romatie ~ was put out and ~'yoshimura" was also put out on the second base (out(2)) at
time point 28.
Prolog clauses given below are two parts of the observation, the former represents a defense of "Hanshin"
team at first half of the first inning, the latter represents that of ~Giants ~ team at second half of the third,
respectively.
The input to our system is the observation of the following form only, and information such as "first
half of the first inning" is inferred automatically by the system. We made experiments for three more
imaginary games. (cf. Appendix A)

event (1, [ ( h i t (2) ,mayumi) ] ) .

event (2, [ ( h i t (1) , t a o ) ] ).
e v e n t (3, [ (gethor, e ,mayumi), ( o u t , k a k e f u ) , (out (2) , t a o ) ] ) .
event (4, [ ( s t r i k e o u t , b a s s ) ] ).
. , .

e v e n t (24, [ ( h i t (1) ,mat sumoto)] ) .

e v e n t (25, [ ( h i t (1) , s h i n o z u k a ) ] ) .
event (26, [ (walk, yoshimura) ] ).
event (27, [(strikeout ,hara)] ).
event (28, [(out, cromat ie), (out (2) ,yoshimura) ] ).

4.4 Results and discussion

In this chapter we present two results by two models of interests (shown in section 4.2) f~om one (input)
observation and explain some expression generations. Because texts generated are in Japanese, we give
 232

word-to-word translations of the outputs. The originals in Japanese texts are given at Appendix B. The
time taken to generate texts for one model is about four minutes (in elapsed time, Suu-3/140, C-Prolog
Interpreter). Results from other three inputs are also given in the Appendix A.

[Game 1, Interests 1: word-to-word translation]

(1)Giants won the severely hitting game. (2)At first half of the first inning, Hanshin gained the first one
point by Mayumi's two-base hit, Tao's hit and Mayumi~s get-home during Kakefu's out. (3)At second half
of the first inning with no out, losing one point, Giants tied by Matsumoto~s hit, Matsumoto's advance to
the second base during Shinozuka's out and Yoshimura's timely hit. (4)Also at second half of the second,
Giants added one point by Nakahata's home run. (5)At first half of the third inning with no oat, losing
one point, ttanshin turned the table by Kid&s walk, Kid&s advance to the second base during Ikeda's out,
Mayumi~s timely hit, Kakefu's two runners home run and Bass's home run. (6)At second half of the third
inning with score 5-2, Giants got a full base by Matsumoto's hit, Shinozuka's hit and Yoshimnra's walk,
but lost the chance by Cromatie's out, (7)however at the last inning, ended the game by Shinozuka's hit,
]%shimura's two-base hit and Itara's clinching home run.

Sentence (1) is a macroscopic expression for the whole game.

Sentence (2) has subject "Hanshin" instead of some player by the transformation to macroscopic struc-
ture. This sentence is extracted from "getting ahead" view point, however, because the score at this time
is 0-0 the system generate "gained the first one point" as a surface phrase. Such a mechanism is built into
the grammar as constraints.
Sentence (3) represents a situation from "same points" view points.
Sentence (4) represents a situation from "home run" view points.
In sentence (5),"advance to the second base during Ikeda's out, Mayumi's timely hit ..?'( r ...2 ~ r .
;~. ~ 0~ ~, 4 ~, 9 --.J ) represents a situation from "same points" view point and "turned the table
by Kakefu's two runners home run and Bass's home run" ( [ ~ 0 ~ 2 ~ :J, J~--:~ 0 ) . t . - ~ ~ >~~ j . ~ j )
represents a situation from '~turning the table ~' and "home run" view points.
Sentence (6) represents a situation from "if hit ..." view point.
Sentence (7) represents a situation from "turning the table" view point but the inning is last one so
the system generates "end the game" ( [- ~ J: ~'£ ~ ~ J ).

[Game 1, Interests 2: word-to-word translation]

(1)Giants won the severelly hitting game. (2)At second half of the second inning with score t-1, Giants
added one point by Nakahata's home run. (3)At first half of the third inning with one ou L score 1-2,
ttanshin turns the table by Mayumi's timely hit, Kakefu's two runners home run and Bass's home run.
(4)At second half of the last inning with no out, score 5-3, Giants settled the game by Shinozuka's hit,
Yoshimura's two-base hit and Hara's three runners home run.
 233

Note that the generated text by Interest 2 is shorter than by Interest 1, because there is less items in
Interests 2.
Sentence (1) is a macroscopic expression for the whole game.
Sentence (2) represents a situation from "home run" view point. Because the system has no interests in
"same points" and "getting ahead" view points the situation occurred before this situation is not extracted.
Sentence (3) represents a situation f~om "turning the table" view point.
In sentence (4) "settled the game by ... Hara's home run" ( r ~ ¢ ~ 3 ~ > " ~ ' ~ . j ) represents a
situation f~om "winning hit" view point. Note that in Interests I this situation was extracted from "turning
the table" view points so the surface sentences in both Interests are different.

5 Discussion and further research

In this paper we have formalized the summarization process, that is reconstruction of the original events
f~om an observation and the extraction mechanism of characteristic facts. And we have also presented our
summary generation system and explained the transformation to the surface structure through generated
texts. In those texts, considerably advanced expressions are found. The reason seems to be that by
limitation of the objective world to baseball game, goals of each event in the world are clear, that decision
of connectives between sentences are simple because, for instance, advantage of one team is disadvantage
of the other, and that the system can cover almost all of transformation to surface structures.
Some processes axe not fully formalized and implementation has not completed. Finally we collect these
problems which are still left to be further research.

1. Investigation and formalization of the transformation to macroscopic structures and the planning.

2. More realistic model of human interests. In actual newspaper scripts of base ball games, contents are
selected not only by events in one game but also by events in the past games. How does the system
use such an information ?

3. Investigation of the existence of readers. We must study the convergence of ~polishment ~model when
the system distinguishes the world of summary writer and readers.

4. Relation to the connectionist model in the extraction process. How does the system assign initial
values to nodes and weights to links in the importance decision function ? Convergence and so called
local minimum must be investigated.

5. More natural sentence generation. Because of poor grammar, only stereotyped sentences are gener-
ated in our system. Not only newspaper-like style but also any styles of sentences would be generated.
Pronominalization, omission of subjects and varying sentence styles are also important.

6. Investigation of the relation between the parameter of human interests and texts generated.
 234

Acknowledgement

We are grateful to Prof. Hozumi Tanaka of the department of computer science, Tokyo Institute of
Technology and members of artificial intelligence research group AIUEO fo~ their valuable suggestive
opinions.

Bibliography

[1] K. t~asida, S. Ishizaki, and It. Isahara. A connectionist approach to the generation of abstracts. In G.
Kempen, editor, Natural Language Generation, pages 149-156, Martinns Nijhoff Publishers, 1987.

[2] J.R. Hobbs. Coherence and coreference. Cognitive Science, (3):69-90, 1979.

[3] Kakiuchi, Makimoto, Uehara, and Toyoda. Explanation generation on manualless system. ICOT NLS
WG, 1986.

[4] K. Kita, E. Komatsu, and H. Yasuhara. Summarization support system, cogito. Preprints Work.Gr.for
Natural Language Processing, 1986. IPSJ, 86-NL-58-7.

[5] It. Koyama and It. Tanaka. Definite clause knowledge representation. In Logic Programming Confer-
ence.'85, pages 95-106, ICOT, 1985. (in Japanese).

[6] D.E. Rumelhart. Notes on a schema for stories. Academic Press, 1975.

[7] R. Schank and R. Abelson. Script Plan8 Goals and Understanding. Lawrence Eflbaum Associates,
1977.

[8] K. Uchiumi and M. Shigenaga. Abstract generation of english text. Preprints Work.Gr.for Natural
Language Processing, 1986. IPS3, 86-NL-54-8.

Appendix A. Results of the generation (English translation)

[Game 2, Interests 1]

Hanshin won by one side game. At ~rst half of the first inning with no out~ Hanshin gained the tlrst
one point by Mayumi's hit, Mayumi's advance to the second base during Tao's out and Kakefu~s timely
two-base hit. Also at first half of the third inning, Hanshin added three points by Tao's walk, Kakefu's
two runners home run and Bass's home run. At second half of the third inning with no out, score 5-1,
the Giants loaded bases by Egawa's hit, Shinozuka's hit and Yoshimura's walk, but lost t]~e chance by
Cromatie's strikeout. At the last inning with no out, score 5-1, Hanshin added one point by Bass's home
run.

[Game 2, Interests 2]
Hanshin won by one side game. At first half of the first inning with one out, Hanshin settled the game by
Kakefu's timely two-base hit and Bass's timely hit. Also at first half of the third inning, Hanshin added
 235

three points by Tao's walk, Kakefu's two runners home run and Bass's home ~un. At the last inning,
ttanshin added one point by Bass's home run.

[Game 3, I n t e r e s t s 1]

tIaushin shut out the Giants by pitcher's game, such as Ikeda's three strikeouts at the third inning and
Egawa's three strikeouts at the fifth inning. At first half of the first inning with no out, tIanshin got the
first point by Mayumi's hit, Maynmi's advance to the second base during Tao's out and Kakefn's timely
two-base hit. At ssecond half of the first inning with no out, score 1-0, the Giants lost the chance by Itara's
strikeout in spite of Matsumoto's hit, Shinozuka's hit and advance to the third base during Yoshimara's
out.

[Game 3, Interests 2]

Hanshin shut out the Giants by pitcher's game, such as Ikeda's three strikeouts at the third inning and
Egawa~s three strikeouts at the fifth inning. At first half of the first inning with no out, ttanshin settled
the game by Maynmi's hit, Mayumi's advance to the second base daring Tao's out and Kakefu's timely
two-base hit.

[Game 4, Interests 1]

The Giants won by one side game. At second half of the first inning with no out, the Giants added
four points by Matsnmoto's hit, Shinozuka's hit, Matsumoto's advance to the third base and Shinozuka's
advance to the second base during Yoshimnra's out, Itara's two points timely hit and Cromatie's two
runners home rnn. At first half of the second inning with no out, score 0-4, Hanshin added one point by
Bass's home run.

[Game 4, Interests 2]

The Giants won by one side game. At second half of the first inning with no oat, the Giants settled the
game by Matsumoto's hit, Shinozuka's hit, Matsumoto's advance to the third base and Shinozuka's advance
to the second base during Yoshimura's out, Hara's two points timely hit and Cromatie's two runners home
run. At first half of the second inning with no out, score 0-4, ttanshin added one point by Bass's home
run.

Appendix B. Results of the generation

[Game 1, Interests 1]

o N + ~ * - a 4 :, b - c I ~ , ~ J . (3) 1 ~ - c ~ _ t c - - N ~ , ~_At;t~¢) ~ .~ ~, ~ ¢ ) 7 ~
 236

~ a m e 1, I~erests2]

~. (3)1-2~Z~, ~Ii~4A~--, ~29~, ;~--X~--A~.

[Game 2, Interests 1]

[Game 2, Interests 2]

[Game S, Interests 1]

[Game 3, Interests 2]

[Game 4, Interests 1]
 237

[Game 4, Interests 2]
 Knowledge Representation and Reasoning for Discourse Understanding*

Satoshi KINOSHITA, Hiroshi SANO, Teruhiko UKITA,

Kazuo SUMITA, and Shin'ya AMANO
Toshiba Corp. R&D center
1, Komukai Toshiba-cho, Saiwai-ku,
Kawasaki 210, Japan

ABSTRACT
Extra-linguistic knowledge is necessary for discourse understanding. In
this paper, we classify the knowledge, and present a framework to describe it
using frames and rules. With this framework, it is easy to represent an IS-A
hierarchy, which is based on a classification by different viewpoints, and to
describe functions of objects as declarative knowledge. Furthermore, to treat
ambiguities in discourse understanding and to process utterances based on
assumptions, the system has a world mechanism for inference. Lastly, we
report the evaluation of this framework through the knowledge representation
of a VCR and the conversation experiment by the dialogue system.

1. Introduction
Context processing technology is most important in building a question-answering
system which reads sentences and answers the user's questions, or a machine
translation system which translates sentences correctly, based on their contexts.
Determining anaphoric reference and complementing ellipses are two of the most
fundamental problems in context processing, and they are especially difficult in
understanding conversations.
For context understanding, it is obvious that the extra-linguistic knowledge and
reasoning based on it are necessary as well as the linguistic one. However, how the
extra-linguistic knowledge should be represented and used has not been fully discussed.
In this paper, we propose a framework for representing the extra-linguistic knowledge,
which is a multi-paradigm knowledge representation system based on frames and rules.
Knowledge representation based only on frames is insufficient in the following points.
(a) In classifying the concepts in a task domain, we can classify them in a different way
by changing the view point, but frames cannot express the taxonomy based on such
classification simply.
(b) Negative expression and inference concerning negation are insufficient.
(c) We cannot describe the behavior of objects and logical relations between attributes
as declarative knowledge.

*This work is supported by ICOT (Institute for New Generation Computer Technology).
 239

The framework we propose here solves these problems. Moreover, the inference
system is based on a world mechanism, which offers multiple hypothetical worlds for
discourse analysis and problem solving.
We have been developing a consulting system which helps the user to use electrical
equipment such as a VCR (video cassette recorder) through conversation. We then
report the appraisal of the framework, which is obtained through the description of
knowledge for this task and experiment of the consulting system.

2. Knowledge and inference for context understanding

2.1 Knowledge for context understanding
Let's discuss "context understanding" before thinking about "knowledge" for it. We use
the term "context understanding" in two meanings. One is understanding context itself,
that is, understanding the contents of whole sentences or understanding the topic or the
structure of the conversation. The other is understanding each sentence correctly in the
flow of the context, that is, determining the meaning of a sentence by resolving anaphoric
reference and ellipsis, using the information of the context. But these two meanings are
complementary. If we can't understand the meaning of each constituent sentence, we
can't understand the meaning of the whole piece of text, and in some cases, understanding
one sentence helps understanding the whole. Furthermore, if an ambiguity exists in a
sentence that can't be resolved in semantic analysis, the context may be able to give the
answer. Since an utterance in conversations depends on its context very much, the second
meaning -understnding a sentence in the context flow- is more important. In this paper,
we will use the term mainly in this meaning.
Needless to say, there are two kinds of knowledge for understanding a sentence. One
is linguistic knowledge such as a grammar, and the other is extra-linguistic knowledge,
knowledge of the world around us. In this paper, we focus on the latter and just call it
"knowledge".
As knowledg e necessary for context understanding, we classify the knowledge into 4
categories as listed below.
(a) Knowledge of objects
(b) Knowledge of events
(c) Relationships between events
(d) Concept hierarchy (thesaurus)

(a) Knowledge of objects

We use the term "objects" to distinguish entities or things, with, or without, physical
existence. Intuitively they correspond to what is denoted by nominals. Knowledge of
objects in a task domain is essential for tmderstanding a sentence.
When we read a sentence, we sometimes can't understand, nor even imagine, the
meaning of the whole sentence even though we can understand the meaning of each word.
This sort of situation may be caused by lack of knowledge of what properties the object in
the sentence has by nature. We think this is the same reason that Minsky proposed
frames to recognize objects[Minsky 75]. The most systematic knowledge of an object is
a model. Having a model of an object helps the understanding of sentences about it.

(b) Knowledge of events

"Events" distinguish some kind of action, state or movement of an object. If we talk
about something, we can't describe it without a predicate and other objects, i.e. in a form
 240

of an event. So we must have knowledge of events. Although this seems to be rather

linguistic, knowledge that what kind of object may be an event agent, source or goal is
strongly connected to the world.

(c) Relationships between events

Determining anaphoric referents and complementing elliptic information are the most
fundamental problems of context understanding. For this, we have to recognize the
relationships between the sentences being processed and the context. For example,
Schank proposed a method of using knowledge about typical sequences of events, called
scripts [Schank 77], but the method cannot analyze sentences which cannot be matched
to the scripts, this means it lacks flexibility. In order to solve this problem, Wilensky tried
to recognize the relationship between events as a goal and a plan for it [Wilensky 83].
We have this sort of knowledge and utilize it to understand sentences. In our current task,
we recognize knowledge of procedures to use a VCR.

(d) Concept hierarchy

Concept hierarchy is necessary in context understanding because we use various
expressions to refer the same object in a dialogue. Let's consider the following sentences.
(1) Did you push the play-back switch?
(2) Yes, I pushed the switch.
Except when using pronouns, we use different expressions that mean different concepts
when we refer to the same entry, as in the example above: a 'switch' is a super class of a
'play-back switch'.

2.2 Inference for context understanding

We can see various kinds of inference in the process of context understanding (for
example [Rieger 75 and Ishizaki 86]). Here we will discuss an inference framework for
treating ambiguities of interpretation.
The most important requirement for an inference system, for context understanding, is a
framework to describe the situations according to the various interpretations of an
ambiguous sentence simultaneously. As an example, let's consider the utterances below.
(3) I put a cassette tape into the VCR, and pushed the play-back switch.
(4) But it didn't work.
Assume that the candidates for 'it' are 'cassette tape 001', 'VCR 00t', and 'playback
switch 001' ('001' denotes instances). Then the hearer (system) can assume three
situations as interpretations shown in Figure 1. We call each situation described in a box
a 'hypothetical context world' or simply a 'context world'.

F ,,---]

Figure [ Exampleof context worlds

 241

The process of context understanding is to select the most appropriate context world.
To do so, we must evaluate every possible interpretation. But in some cases, we can't
evaluate them unless we make some inferences from the interpretation fin'st, which may
change the contents of the working memory, and for that, we must be able to create
context worlds simultaneously.
The interpretation which is decided by the context analysis is not always correct. If the
system finds an inconsistency, it must resolve it. In the field of expert systems, Doyle
proposed an inference system called TMS [Doyle 79]. A context understanding system
should have this sort of inference system, and the world mechanism is a key to its
realization.
Considering human language activity, we can find that we talk not only about objects
and events in a real world but also about those in an imaginary world. So the system
should process utterances based on conditionals. We think this is possible by utilizing
context worlds. For processing such sentences, the system has only to create a new
context world and process the following sentences in this world. The context world which
corresponds to the imaginary world per se differs from the one which is created for context
understanding, but can be processed in the same framework.

3. Design and realization of the knowledge representation system

As stated earlier, we are developing a consulting system which guides the usage of
electrical equipment, such as a VCR, through a conversation with the user. In this chapter
we show the concrete framework for knowledge representation of the task domain and the
inference mechanism for understanding the user's utterance.
There are various knowledge representation paradigms, and the features of each have
been discussed. We use 'frames' and 'rules' because of the ease in describing the
concept hierarchy and the logical relation and in managing the described knowledge.
Figure 2 illustrates the composition of a knowledge base in the knowledge
representation system. Knowledge of objects and events is described in a framework
called a 'schema' that is composed of a frame part and a rule part. Knowledge of the
relations between events, such as the causal relations, is described as a set of rules.

3.1 Knowledge representation

3.3.1 Knowledge representation using schema
Figure 3 shows a schema example (a part of the full description of a VCR). The
schema is composed of a frame part and a rule part. Static knowledge is described in the
frame part: such as super/sub class relations which compose an IS-A hierarchy, whole-
part relations, and attributes of the object. In the rule part, actions or the behavior of the
object is described.
There are two types of schemata in the knowledge base. One is an instance schema
and the other is a class schema. Roughly speaking, an instance schema corresponds to a
real object, and the information of the object is described in it. A class schema
corresponds to a class of objects, and the restrictions of attributes or possible behavior of
the objects are described.

(A) Expression of knowledge by frames

As stated above, static knowledge such as an IS-A hierarchy is described in the frame
part. As KRL [Bobrow 77], a frame is defined as a set of slots, and each slot is a set of
facets. The features of the frame system in this paper are described below.
 242

Figure 2 Co~position of Knowledge Base

schema(cls, vcr_model_750,
[(superC, type_of_.soundrecording, [(value,[hi_fi_ver])]),
(superC, type of easette, [(value,[vhs_vcr])]),
(haspart, [(value,[power_switch, play_back_switch,...])]),
(status, [(car_num,1 ),
(enumeration,[stand_.by,play_back, record, ...])]),
(power, [(car_num,1 ), (enumeration,Ion,off ])]),
(channel [(car_hum, 1),
(value__class, integer),
(value_.condition, (_,V,(V>=I,V=<12))) ])
],
[s_rule(power, Power__on,VCR,
( kr_schema(VCR, [(power, [off] )],true),
event(push, 1,[(object,[SW])] ) ),
kr_schema(VCR, [(power, [on] )],true),
[('$var.constraint',(
SW#powerswitch :- parLof(SW,VCR) ))]),
s_rule(power,power_lampon,VCR,
kr_schema(VCR,[(power, [on] )],true),
kr_schema(LP, [(lamps tams, [on] )] ,true),
[('$var_constraint',(
LP#power_.lamp :- part_of(LP,VCR))])
]).
Figure 3 Schema example for a VCR (a part).
Kr_schema is a predicate that extracts or verifies slot values.
Words beginning with a capital letter denot~ variables.
 243

Representation of the concept hierarchy

We can classify one concept in different ways by changing the viewpoint, that is the
feature we consider for discrimination. Figure 4 is a part of an IS-A hierarchy of our
current task domain. If we classify a VCR, it can be classified in two different ways.
Considering the type of cassette tape, we can classify it into 'VHS-type VCR' and 'Beta-
type VCR', but by the type of sound recording, it is classified into 'hi-fi VCR' and
'monaural VCR'. So far, it has been difficult to describe such a hierarchy naturally, using
conventional frame systems. In our system, it is described quite easily and naturally by
attaching the viewpoint of the classification as additional information for a slot (See
superC slot of Figure 3).

classification by recording of picture

Beta-type VCR - - - ~ / " ~ vcr440x

Hi-fi VCR --- vcrhfrT0

moaaurat VCR- - = ~ - - - ~ vcrvs8

classification by recording of sound

Figure 4 Exampleof classification fro~ different viewpoints

This attachment of viewpoints is used in judging anaphoric referents. In a conversation,

we use different expressions to refer to the same object, and sometimes they mean
different concepts. In such a case, we can use the hierarchical relation between the
concepts for the judgement.
Let's consider checking of identification of two objects; one is in the context and the
other is in the sentence being processed. Assume that B is a concept of the former, and A
is of the latter. If A and B are different, there are four possible relations between them.
(a) A is a sub-class of B.
Example: A:VHS VCR and B:VCR
(b) A and B are both sub-classes of a common concept C in the same viewpoint.
Example: A:television, B:VCR, and C:electfical equipment
(c) A and B are both sub-classes of a common concept C in a different viewpoint.
Example: A:VHS VCR, B:hi-fi VCR and C:VCR
(d) B is a sub-class of A (opposite case of a).
Example: A:VCR and B:VHS VCR
Figure 5 illustrates the above classification. An arrow between concepts is a subC
(sub-class) link, and a dotted line around arrows is the viewpoint for the classification.
Case (a) is the most general case. If two objects don't have any inconsistent property,
we can judge that they are the same. In case (b), since these two objects essentially
have inconsistent properties, we cannot recognize that they are the same object, but in
the case of (c), it is possible to assume they are the same. Case (d) is the opposite case
of (a). In this case, by judging that they are same, the information about the preceding
object will increase.
The above example illustrates the case when a sub-class relation between two classes
is direct. The same procedure should be applied for indirect relations among concepts.
 244

q) ®
B O

A A B
(saMe viewpoint)
® ®
C A

A B B
(different
viewpoint )

Figure 5 Relations between concepts

Negative expression and the open/closed assumption

For a system which performs problem solving through a dialogue with a user, most data
necessary for problem solving are unknown at the begining of inference. Furthermore an
input sentence by the user may be a negative sentence. Although many systems based on
Prolog treat negation on a closed world assumption, this is not always appropriate.
To solve this problem, we must have a framework to describe negation explicitly and
also to describe whether the data used for reasoning am 'open' or 'closed'. 'Open' and
'closed' indicate the assumption for inference. If the system holds all the possible data
about an item, we say the data are closed for the item, and if there is a possibility of
another piece of data other than the current ones, we say the data are open.
In our knowledge representation, we can set a negative value as a slot value and add
the information that the data for the slot are 'open' or 'closed'. With this framework, we
can use the open/closed world assumption properly in the comparison of slot values and
the reasoning based on them.

Knowledge of events
We use case representation to describe an event. Using a schema, a case
representation is described quite naturally by representing each case as a slot. Figure 6
shows an example description for a verb "rokuga-suru (to record something on a video
tape)", where slots (agent, object and tool) have a special facet describing constraints on
fillers of these cases. In this example, the constraint is that the agent, object and tool

schema(cls, record,
[(super(], [(value, [action])])]),
(agent, [(value_class, human_being)]),
(object, [(value_class, tv program)]),
(tool, [(value class, vtr)]),
(obligatory_case,
[(value, [object,tool])] ) ], []).
Figure 6 Eventexpressionby schema.
 245

must be a human, a TV program and a VCR respectively. The requirement for slot
fulfillment, in implementing a task, is described in an 'obligatory-case' slot. These
constraints are used to reduce search space in interpreting an input sentence.

(B) Knowledge representation using schema rule

In addition to static knowledge, there is knowledge of functions of an object or logical
relations between attributes. We use a rule in a schema called a 'schema rule' to
describe this sort of knowledge.
In the case of a VCR, we have to represent how the internal state of a VCR changes by
its operation. Figure 3 shows two schema rule examples. The first one describes that, if
the power-switch of a VCR is pushed when the power is 'off', then the power will be 'on'.
To describe such knowledge using only a frame, we must describe it using attached
procedures like if-added demons. Since there are many functions for usual devices, the
clarity of the knowledge will be lost. By describing such knowledge using a schema rule,
we can describe it as declarative knowledge, and the relation between the elements of
knowledge wiU be kept clear.
Describing such knowledge of an object bit by bit, we get its model, and by forward
reasoning with the knowledge, we can simulate its behavior. Lastly, these schema rules
are inherited by lower classes like the slots of a frame and, therefore, economical
description of functions is possible.

3.1.2 Knowledge representation using rule

In addition to knowledge described in a schema, we have to represent the relations
between events such as the causal relation. In our framework, we use 'if-then' rules to
describes knowledge of this sort.
For the current task, rules are used to represent procedures to operate a VCR and are
prepared for every operational procedure, such as play-back and recording. Figure 7 is an
example of a rule that describes how to play back a video cassette. The conclusion part
holds the goal (to play back), and the conditional part holds the actions to accomplish it.
Furthermore, every rule has some constraints on the variables. In the example, the
variable 'SW' is constrained so that its content must be an instance of 'play-back switch'
and be part of a VCR.

rule( operation, playback,

seq( event( puton, 1, [(object,[PW])]),
event( insert, t, [( object, [K]), ( goal, [VTR])]),
event( push, I, [( object, [SW] )]) ),
event( play_back, 1, [( object, [K]), ( tool,[VTR])] ),
[('$var_constraint',(
VTR # vtr;
K # cassette;
PW # power :- attribute_of( PW, VTR) ;
SW # play_back switch :- part_of ( SW, VTR) ) )]).
Figure 7 Operationprocedure exampleby rule.
 246

3.1.3 Representation of time

In a task which only has to treat the present state of the apparatus, such as trouble
diagnosis, there is little necessity to represent time. However, in guiding the usage of an
apparatus, the order of the user's operations is quite important. Since we have a
mechanism to describe context worlds, it is possible to represent the states of objects at
various points in time by creating worlds for every point in time under consideration, but
this representation makes it difficult to manage the context worlds. For this reason, we
extend the data structure of a slot value of our frame system to be the pairing of its value
and a time description. So every slot may have various values according to time.

3.2 Inference function

As a basic function, the system has an inference mechanism for frames to inherit slots
or to check the restrictions of a slot, and a mechanism of forward/backward reasoning
using rules. As we described, we have two kinds of rules; a schema rule in a schema and
a rule in the rule base. In consultation about an apparatus, it is important that the system
holds the state of the machine and predicts a possible state based on a simulation. In our
system, this is performed by forward reasoning using schema rules.
Moreover, to realize the context world, we designed the system to create a working
memory for reasoning according to each context world. Hereafter we call each working
memory simply the 'world'. Using this mechanism, we can create context worlds for each
possible interpretation in discourse analysis, and evaluate their properties
simultaneously. Also in problem solving, we can process a sentence based on a
conditional using this framework. This will be described in 4.2.

3.3 Implementation
We implemented the knowledge representation system on a PSI (Personal Sequential
Inference machine) developed by ICOT. The programming language is ESP (Extended
Self-contained Prolog). The size of the program is about 7000 lines including comment
lines.
Since ESP is a kind of object oriented language, it is possible to translate all knowledge
written in a schema to an object of ESP. However, in the current system, for ease of
debugging and loading speed, we transform a set of schemata in one file to an object of
ESP, and load it into the main memory.
Execution speed is not satisfactory because the system processes the frame data by
interpretation. By compiling frame data to a Prolog program like DCKR (Definite Clause
Knowledge Representation)[Tanaka 86], we can make the system faster.

4. Application and appraisal

4.1 Application in a dialogue system
We are developing an experimental question-answering system called ISAC
(Information Service System by Analyzing Conversational Context) whose current task is
consultation for operating a VCR, and we use the knowledge representation system as
the knowledge module of the system.
Starting with SHRDLU[Winograd 72], many dialogue systems have been developed.
The biggest difference between our system and SHRDLU is, that in SHRDLU, the world
for conversation is closed and the system holds all information for inference, but in our
system, the system can get information about the world only through the user's input.
 247

This corresponds to a situation where an expert could consult via telephone conversation.

Discourse analysis
Here we briefly introduce a procedure of discourse analysis for interpreting an input
sentence[Ukita 88]. The procedure consists of four major steps; anaphora detection,
referent candidate extraction, non-contradiction testing, and optimal candidate selection.
First the procedure detects anaphoric indicators in a sentence, which are pronouns or
nouns, with or without definite articles. Furthermore, the omitted obligatory cases of the
events (predicates) are detected as ellipses.
Next, for every anaphoric indicator, the procedure searches for referent candidates that
have already emerged in the context, tn the system, they are usually defined as instance
schemata. If the procedure can't find an appropriate candidate for an indicator, it creates a
new instance schema and treats it as the candidate. The inter-object relationships such
as whole-part or attribute-value are also checked here.
When more than two possible interpretations remain after the non-contradiction test,
the procedure tries to find causal links between the interpretations and the context, and
decides on the most informative interpretation. This is the step of optimal candidate
selection. In inspecting causal connections, the procedure searches for rules that describe
the usage of a VCR. Let's consider an example sequence of sentences, whose preceding
context contains a cassette tape and a practical image on a television set.
(5) I pushed the eject button. It didn't come out.
Assume that the referent candidates of "if' are the cassette tape and the image. To
decide on the appropriate referent, the procedure finds a rule such as follows:
If (a cassette is inserted in a VCR),
and (the eject button is pushed),
then (the cassette comes out).
Using this rule, the procedure prefers the interpretation that the cassette didn't come out.
Currently we don't use schema rules which are described in a schema for this step. In
referent candidate extraction, if an anaphoric indicator is a noun, the procedure searches
for an object that is an instance of a concept meant by the noun. This is just the case in
(a) and (c) described in Figure 5. We have to improve the discourse analyzing program to
treat other cases.

Problem solving
The current task of the consulting system is categorized into 3 sub-tasks as follows.
(a) Answering an inquiry about an operating procedure
(example: Would you tell me how to play back this tape?)
(b) Solution of a fault which originates in an operation mistake
(example: I pushed the play-back switch, but it didn't work.)
(c) Prediction of an action of the VCR
(example: What happens if I push the play-back switch now?)
An input sentence is classified into one of these types, and the problem solving module
for this type is activated. Finally the result of the task is output as a question or an
answer to the user.

Dialogue example
Figure 8 shows a conversation example using the experimental system. Before this
 248

user: ~*~ ~ ~- - 7" :~:~7~[~L. tc ~ (I want to playback the video-tape)*

system: ~'~-'~©~/f~.~)k:h,-C~ ~:q-t'C;~-e ~, b~---7*-~.Tkah~7~
N ~ > ' ~ L."c< tE~
(Please turn on the VTR power, insert the cassette into the VTR, and push
the play-back button.)
user: ~-'~]/4z ~ b Y---7"~r.Tk']rb~ :~t~_~,N~r: / ~ L T ~
(I inserted the cassette in the VTR and pushed the play-back button)*
system: ~ , ~z~ 9 ~ L ~ (OK)
user: ~JJZ)~;~~ ("It doesn't operate" or "It doesn't move")**
system: k : ~ © ~ :-'~*'9~ (Is the VTR power 'ON'?),
user: ~ (Yes)
system: :N~[~7 :/7" ~,~q" I., -(7~ ~ 3t"~ (Is the play-back lamp lit?)

user: .~,'~-IL~w~-'~ (It's lit)*

system: V H F ~ N : ~ Y"1,- ~"t ~ , ~ ~ ;]'1~-C ~ ~ -~rz~
(Is the VHF output terrrhnal connected to the TV serf)
user: WWL (No)
system: VHF~S¢J~:/:"~ ~" tt ~ ~ t ~ I.."(7< 7"~~
(Please connect the VHF output terminal to the TV set.)
Figure 8 Dialogueexample using the experimental system
* In Japanese, the subject is omitted.
** Japanese verb 'g~J< (ugoku)' has several meanings.

conversation, the system was given the situation that there is a VCR and a television set,
and that the initial state of the devices is such that the power is off. Unless such an initial
state is given, the system asks the user about it and about his/her actions, and continues
the consultation.

4.2 Evaluation of the framework of knowledge representation

We report the evaluation of our framework of the knowledge representation, which was
gotton by representing knowledge about a kind of home hi-fi VCR. Using its operation
manual, we represented the knowledge of the VCR, but didn't represent either knowledge
concerning spatial relations, such as positions of switches, or knowledge which requires
adjectival or adverbial expression for its representation.
4.2.1 Representation using schema
The number of concepts which are described in schemata is about 300. 120 schemata
are concepts of events (verbs relevant to operations of the VCR), and 80 schemata are
concepts of the VCR itself and its parts, such as switches and lamps. The rest are
concepts of the attributes, such as channel, or of immaterial objects, such as a television
program or a picture.
 249

Representation of the concept hierarchy

In describing an IS-A hierarchy, we can represent a classification based on different
viewpoints easily and naturally. In the current hierarchy, there are only two sub-
hierarchical structures: one is shown in Figure 4, and the other is a sub-hierarchy whose
root concept is a 'terminal'. However we found the framework is quite valid.

Negative expression and the open/closed world assumption

Since, in the current task, each slot can only have one value, we are unable to use the
open/closed assumption to its full advantage. As for negation, when the knowledge
representation system gets a command to fill a slot with a negative value, it will try, first,
by reductio ad absurdum to infer a positive one. If this fails, the negative one will be used.
For example, since the slot 'power' in Figure 3 may only have the value 'on' or 'off' at
one time, the slot will be set to 'off' if the negation of 'on' is provided. In a consultation,
the user may answer the system's question with a negative sentence. So this reasoning
is very useful to process such a sentence.

Representation using schema rules

The functions of the VCR are described in about 170 schema rules: how it behaves
according to an operation and how the state of the lamps is affected. Since these rules
describe its model, we can simulate its behaivior, and this reasoning is quite useful for
problem solving.
Currently we can't describe the events which happen continuously or iteratively under a
certain state. For example, "If a VCR is in playback state, the cassette tape in it goes
around" or "the display of the VCR's clock changes every minute". To describe this kind of
knowledge, we have to extend our framework.

4.2.2 Representation using rules

The operation procedures of the VCR are described in about 30 rules. In describing the
knowledge, we used the expressions in the operation manual as much as possible.
However, we couldn't represent the following types of expressions as they are.
(a) The case of the purpose or the condition for an operation being described.
(example: "Please push the video button so as to put on the video lamp.")
In such a case, we neglected the purpose or the condition and wrote only the operation.
(b) The case of the goal of an operation being expressed as two events.
(example: "To watch a TV program while you record another program .... ")
In this example, we used a word that means 'a program on a different channel', but this is
not always possible.

4.2.3 Inference in the world mechanism

In processing a conditional sentence (we will call it a 'hypothetical sentence'), we
found inference within the world mechanism to be useful. To process a hypothetical
sentence, the system first creates a new world, evaluates its potential, and continues the
dialogue. Through the experiment, we found that our choice of world is important for
processing the next sentence. As an example, let's consider a sequence of the user's
input during a consultation.
(6) What happens if I push the recording switch now?
(7) What happens if I push the play-back switch instead?
(8) What happens if I push the fast forward switch after that?
 250

w°- 1

~
(7)

~ (8)

Figure 9 Exampleof world co~posit~on

Figure 9 illustrates an example of world composition in processing the sentences

above. The world W0 is a world at which sentence (6) was input, and W1 is a world
which is created in order to process (6). In processing (7), the supposition (pushing the
piny-back switch) must be evaluated instead of the supposition of (6), that is, we have to
create a new world W2 whose parent is also world W0. Furthermore, in processing (8),
we must evaluate it in the light of the supposition of (7), therefore we have to create W3
as a child of W2. For this, it is very important to manage relations between the context
and the worlds.

5. Conclusions
In this paper, we discussed a classification of knowledge for context understanding, and
a framework for its representation and reasoning based on this knowledge. We first
classified the knowledge into four categories; knowledge of objects, knowledge of events,
concept hierarchy, and relationships between events. The first three are described in a
framework called a schema, and the last one is described by rules. Using schemata, we
can naturally represent an IS-A hierarchy based on different viewpoints, and also describe
a negative slot value, and use the open/closed assumption properly. Furthermore, by
using schema rules, we can describe functions of an object and logical relations between
attributes declarativety. As for inference, we pointed out the need for a context world for
resolving an ambiguity in context understanding, and its validity in problem solving. Then
we discussed our knowledge representation system which is implemented as a part of an
experimental question-answering system for a VCR usage application. By representing
the knowledge of the task domain, we found the above points to be advantages in our
system. At present, there is some knowledge that can't be written in the current
framework, but the framework reported here is thought basically applicable to context
understanding in other task domains.

References
[Bobrow 77] Bobrow,D.G. and Winograd,T., "An Overview of KRL, a Knowledge
Representation Language", Cognitive Science, Vol. 1, 1977, pp.3-46.
[Doyle 79] Doyle,L, "A Truth Maintenance System", Artificial Intelligence 12, 1979,
pp.231-272.
 251

[Ishizaki 86] Ishizaki,S. and Isahara, H., "Contextual Processing Technology", Vol.27,
No.8, 1986, pp.897-905(in Japanese).
[Minsky 75] Minsky,M., "A framework for representing knowledge", in Winston,P. (ed.),
The psychology of computer vision, McGraw- Hill, 1975, pp.211-277.
~Rieger 75] Reiger,C.J., "CONCEPTUAL MEMORY AND INFERENCE", ha Schank,R.
(ed.), Conceptual Infomation Processing, NORTH-HOLLAND, 1975, pp. 157-288.
[Schank 77] Schank,R. and Abelson,R., "Scripts,Plans,Goals and Understanding",
Lawrence Erlbaum Associates, 1977.
[Tanaka 86] Tanaka,H., "DCKR - Knowledge Represetation in Prolog and Its Application
to Natural Language Understanding", in Nagao,M.(ed.), Language and Artificial
Inteligence, Else Ver Science Publishers, B.V. (North Holland), 1987, pp.353-366.
[Uldta 88] Ukita,T., Sumita,K., Kinoshita,S., Sano,H., and Amano,S., "Preference
Judgement in Comprehending Conversational Sentences Using Multi-Paradigm World
Knowledge", Proc. of FGCS' 88, 1988(to appear).
[Wilensky 83] Wilensky,R., "Planning and Understanding", Addison-Wesley, 1983.
[Winograd 72] Winograd,T., "Understanding Natural Language", Academic Press, 1972.