Wright State University

CORE Scholar

Computer Science and Engineering Faculty Computer Science & Engineering

Publications

4-2013

Reasoning with Inconsistencies in Hybrid MKNF Knowledge

Bases
Shasha Huang

Qingguo Li

Pascal Hitzler
[email protected]

Follow this and additional works at: https://corescholar.libraries.wright.edu/cse

Part of the Computer Sciences Commons, and the Engineering Commons

Repository Citation
Huang, S., Li, Q., & Hitzler, P. (2013). Reasoning with Inconsistencies in Hybrid MKNF Knowledge Bases.
Logic Journal of the IGPL, 21 (2), 263-290.
https://corescholar.libraries.wright.edu/cse/160

This Article is brought to you for free and open access by Wright State University’s CORE Scholar. It has been
accepted for inclusion in Computer Science and Engineering Faculty Publications by an authorized administrator of
CORE Scholar. For more information, please contact [email protected].
 Reasoning with Inconsistencies in Hybrid MKNF
Knowledge Bases
Shasha Huanga , Qingguo Lia,∗, Pascal Hitzlerb
a
College of Mathematics and Econometrics, Hunan University, P. R. China
b
Kno.e.sis Center, Wright State University, Dayton, OH, USA

Abstract
This paper is concerned with the handling of inconsistencies occurring in
the combination of description logics and rules, especially in hybrid MKNF
knowledge bases. More precisely, we present a paraconsistent semantics for
hybrid MKNF knowledge bases (called para-MKNF knowledge bases) based
on four-valued logic as proposed by Belnap. We also reduce this paracon-
sistent semantics to the stable model semantics via a linear transformation
operator, which shows the relationship between the two semantics and in-
dicates that the data complexity in our paradigm is not higher than that
of classical reasoning. Moreover, we provide fixpoint operators to compute
paraconsistent MKNF models, each suitable to different kinds of rules. At
last we present the data complexity of instance checking in different para-
MKNF knowledge bases.
Keywords: Knowledge representation, Description logics and rules,
Non-monotonic reasoning, Paraconsistent reasoning, Data complexity

1. Introduction and Motivation

The Semantic Web [3, 15] extends the current World Wide Web by stan-
dards and techniques that help machines to understand the meaning of data
on the web to enable more powerful intelligent system applications. The
essence of the Semantic Web is to describe data on the web by metadata that

∗
corresponding author
Email addresses: [email protected] (Shasha Huang),
[email protected] (Qingguo Li), [email protected] (Pascal
Hitzler)

Preprint submitted to a journal October 22, 2012

conveys the meaning—the semantics—of the data, and that is expressed by
means of so-called ontologies, which are knowledge bases as studied in the
field of Knowledge Representation and Reasoning.
The Web Ontology Language OWL [14] has been recommended by the
World Wide Web Consortium1 for representing ontologies. However, OWL
is not as expressive as needed for modeling some real world problems. For
example, it cannot model integrity constraints or closed-world reasoning that
may be more suitable in some application scenarios. Consequently, how
to improve OWL has become a very important branch of research in the
Semantic Web field.
Knowledge representation approaches using rules in the sense of logic pro-
gramming (LP), which is complementary to modeling in description logics
(DLs, which underly OWL, see [15]) with respect to expressivity, have be-
come a mature reasoning mechanism in the past thirty years. Thus combining
rules and DLs is of continuous interest for the Semantic Web. However, sig-
nificant differences between DLs and rules make the development of merged
paradigms a hard problem. One of these differences is that the Open World
Assumption (OWA) is employed in DLs, while the Closed World Assumption
(CWA) is adhered to rules. Naive combinations of DLs and rules also lead
to undecidable languages which are often deemed undesirable and should be
avoided when developing DL-based paradigms.
A significant number of different approaches have been proposed for in-
tegrating DLs with rules. They can roughly be divided into two kinds: On
the one hand, there are homogeneous approaches that unify DLs and LP in
a special, unified, knowledge representation language. DLP [12], SWRL [16],
ELP [25], nominal schemas [4, 5, 23, 24] and Hybrid MKNF knowledge bases
[21, 33] are methods that belong to this kind of approach. On the other
hand, there are hybrid approaches that view DLs and rules as independent
parts, retaining their own reasoning mechanisms. AL-log [6], CARIN [26],
HEX-programs [8], DL-programs [7] and DL+log [37] are all examples of this
integration approach.
Among these approaches, hybrid MKNF knowledge bases, originally based
on the stable model semantics [11], is one of the most mature integration
methods. It has favorable properties of decidability, flexibility, faithfulness
and tightness. A well-founded semantics [40] for such knowledge bases has

1
W3C, http://www.w3.org/

2
been proposed subsequently for better efficiency of reasoning [20, 21]. How-
ever, the integration of different knowledge bases can also easily lead to
inconsistencies, even if both of the integrated knowledge bases are consis-
tent if taken alone. Accordingly, reasoning systems based on the previous
two semantics will break down. Therefore it is necessary to present a new
semantics for hybrid MKNF knowledge bases to handle inconsistencies.
Traditionally there are two kinds of approaches to handle inconsistencies,
one of which is recovering consistencies [39, 13] by repairing the knowledge
base. But this approach may cause new problems, such as different results
caused by different methods of recovering consistencies, inability of reusing
information that has been eliminated, and so on. The other method admits
inconsistencies and deals with them directly in a paraconsistent logic, and
usually a four-valued logic [2, 28, 29, 30, 35, 38] is chosen for such purpose.
In this paper, we adopt the four-valued logic from [28, 29, 30], and present
a paraconsistent semantics for hybrid MKNF knowledge bases. We will call
the obtained paradigm para-MKNF knowledge bases. Our contribution can
be summarized as follows:

• The paraconsistent MKNF model is faithful w.r.t. the four-valued

model of the description logic ALC from [28, 29, 30] and w.r.t. the
paraconsistent stable model of extended disjunctive logic programs
from [38].

• We present a transformation from para-MKNF knowledge bases to

hybrid MKNF knowledge bases, which shows that our paraconsistent
semantics is also faithful w.r.t. two-valued semantics for hybrid
MKNF knowledge bases, and indicates that the data complexity of
paraconsistent reasoning is not higher than that of standard MKNF
reasoning.

• We define a fixpoint operator to compute paraconsistent MKNF

models for para-MKNF knowledge bases in the positive case and
stratified case, and provide a transformation from general MKNF
rules to positive MKNF rules, such that the fixpoint operator can
evaluate the paraconsistent MKNF models of para-MKNF knowledge
bases in the general case.

• We discuss the data complexity of instance checking in different kinds

of para-MKNF knowledge bases.

3
 The remainder of the paper is organized as follows. In Section 2, we
recall preliminaries on the four-valued Description Logic ALC4 and on hybrid
MKNF knowledge bases. In Section 3, we propose paraconsistent semantics
for hybrid MKNF knowledge bases, and study its fundamental properties. In
Section 4, we present a transformation from paraconsistent semantics to the
stable model semantics of hybrid MKNF knowledge bases. In Section 5, we
characterize the paraconsistent MKNF models via fixpoint operators, design
procedures for computing the paraconsistent MKNF models and analyze the
data complexity of different types of rules. In Section 6, we discuss related
work. We conclude and discuss future work in Section 7.
This paper is a significantly extended and revised version of [17].
Throughout the paper, we make use of the following penguin example.
Example 1. Most bird species can fly, with some exceptions, such as pen-
guins.
A bird ontology can structure and maintain the database. But it is not
sufficient to correctly explain this statement by just building a bird ontol-
ogy. In fact, an ontology specifies concepts, such as bird and penguin, and
the relationships between them, such as every penguin is a bird. However,
exceptions can not be represented correctly in the bird ontology. Thus one
needs to employ other proper knowledge representation tools, such as non-
monotonic rules.

2. Preliminaries
In this section, we introduce notions and notations used in the sequel. In
detail, we present preliminaries for the four-valued description logic ALC4,
the logic of minimal knowledge and negation as failure (abbreviation MKNF)
and hybrid MKNF knowledge bases.

2.1. The Four-valued Description Logic ALC4

The basic idea of the four-valued description logic is
k ⊤
to substitute four truth values for the two truth values 6 @
@
used in classical logic: the four truth values are t, f , ⊤ f @ t
and ⊥, representing true, false, contradictory (both true @
and false) and unknown (neither true nor false) respec- @@
⊥ -t
tively. With two partial orders ≤k and ≤t , which stand
for a measure of the amount of information and a mea- Figure 1: FOUR
sure of truth, respectively, the set IV = {t, f , ⊤, ⊥}

4
 Syntax Semantics
A AI = ⟨P, N ⟩, where P, N ⊆ △I
o oI ∈ △I
¨
⊤ ⟨△I , ∅⟩
¨
⊥ ⟨∅, △I ⟩
¬C ⟨N1 , P1 ⟩, if C I = ⟨P1 , N1 ⟩
Concepts C1 ⊓ C2 ⟨P1 ∩ P2 , N1 ∪ N2 ⟩, if CiI = ⟨Pi , Ni ⟩, i = 1, 2
C1 ⊔ C2 ⟨P1 ∪ P2 , N1 ∩ N2 ⟩, if CiI = ⟨Pi , Ni ⟩, i = 1, 2
∃R.C ⟨{x| ∃y ∈ △I : (x, y) ∈ RI and y ∈ proj + (CI )},
{x| ∀y ∈ △I : (x, y) ∈ RI implies y ∈ proj − (CI )}⟩
∀R.C ⟨{x| ∀y ∈ △I : (x, y) ∈ RI implies y ∈ proj + (CI )},
{x| ∃y ∈ △I : (x, y) ∈ RI and y ∈ proj − (CI )}⟩
C 7→ D △I \ proj − (CI ) ⊆ proj + (DI )
C@D proj + (CI ) ⊆ proj + (DI )
Axioms CD proj + (CI ) ⊆ proj + (DI ),
and proj − (DI ) ⊆ proj − (CI )
C(a) aI ∈ proj + (CI )
R(a, b) (aI , bI ) ∈ RI

Table 1: Four-valued Semantics for ALC4

forms a bilattice FOUR [2] as shown in Figure 1. We now recall the syntax
and semantics of the four-valued description logic ALC4 [28] that is based
on FOUR.
Syntactically, the elementary ingredients of ALC4 are similar to the crisp
ALC, except for three kinds of class inclusions, which are called internal
inclusion @, material inclusion 7→ and strong inclusion , corresponding to
the three implication connectives in the four-valued logic case [2]. An ALC4
knowledge base O consists of axioms of the forms presented in Table 1, where
C(a) is called a concept assertion and R(a, b) is called a role assertion.
Semantically, a four-valued interpretation I = (·I , △I ), where △I is a
nonempty set (i.e., the domain of the interpretation) and ·I is a function
that assigns a distinct element aI ∈ △I to each individual a, a pair ⟨P, N ⟩
of (not necessarily disjoint) subsets of △I to each concept A and a binary
relation RI ⊆ △I × △I to each role R. Intuitively, P is the set of elements
that are known to belong to the extension of a concept C, while N is the

5
set of elements known to be not contained in the extension of concept C.
Complex concepts are interpreted inductively as shown in Table 1, where
proj + (⟨P, N ⟩) = P and proj − (⟨P, N ⟩) = N . The semantics of the axioms
in an ontology is formally defined in Table 1.
A four-valued interpretation I is a paraconsistent model of an ontology O
iff it satisfies each axiom as shown in the bottom part of Table 1. An ontology
O is paraconsistently satisfiable iff there exists such a model. Furthermore,
O paraconsistently entails an axiom α, written O |=4 α, iff I satisfies α, for
every paraconsistent model I of O.

Example 2. Consider the penguin scenario of Example 1.The following ax-

ioms can be built as a part of the bird ontology.

Penguin @ Bird (1)

Bird @ Fly (2)

Penguin(Tweety) (3)

¬Fly(Tweety) (4)

Axiom (1) states that every penguin is a bird, axiom (2) expresses that
every bird can fly. Axioms (3) and (4) refer to individual Tweety. Axiom
(3) states that Tweety is a penguin, and axiom (4) expresses that Tweety
cannot fly.

Let C be a concept in ALC4 and a ∈ ∆I an instance. For the four-valued

semantics defined above, the following indicates the correspondence between
truth values from FOU R and concept assertions:

- C I (a) = t, iff a ∈ proj + (CI ) and a ̸∈ proj − (CI );

- C I (a) = f , iff a ̸∈ proj + (CI ) and a ∈ proj − (CI );

- C I (a) = ⊤, iff a ∈ proj + (CI ) and a ∈ proj − (CI );

- C I (a) = ⊥, iff a ̸∈ proj + (CI ) and a ̸∈ proj − (CI ).

6
 In [28], a transformation from ALC4 to ALC was defined as follows (where
A is an atomic concept, C, D and E are concepts and R is a role):

- If C = A , then λ(C) = A+ , where A+ is a new concept;

- If C = ¬A, then λ(C) = A− , where A− is a new concept;

¨ then λ(C) = ⊤,
- If C = ⊤, ¨ where ⊤
¨ is the top concept;

¨ then λ(C) = ⊥,
- If C = ⊥, ¨ where ⊥
¨ is the bottom concept;

- If C = E ⊓ D, then λ(C) = λ(E) ⊓ λ(D);

- If C = E ⊔ D, then λ(C) = λ(E) ⊔ λ(D);

- If C = ∀R.D, then λ(C) = ∀R.λ(D);

- If C = ∃R.D, then λ(C) = ∃R.λ(D);

- If C = ¬¬D, then λ(C) = λ(D);

- If C = ¬(E ⊓ D), then λ(C) = λ(¬E) ⊔ λ(¬D);

- If C = ¬(E ⊔ D), then λ(C) = λ(¬E) ⊓ λ(¬D);

- If C = ¬(∀R.D), then λ(C) = ∃R.λ(¬D);

- If C = ¬(∃R.D), then λ(C) = ∀R.λ(¬D).

An ALC4 knowledge base O is transformed to an ALC knowledge base

O based on the above transformation λ:
λ(C 7→ D) = ¬λ(¬C) ⊑ λ(D);

λ(C @ D) = λ(C) ⊑ λ(D);

λ(C  D) = {λ(C) ⊑ λ(D), λ(¬D) ⊑ λ(¬C)};

λ(C(a)) = λ(C)(a); λ(R(a)) = λ(R)(a),

where C and D are concepts, a is an individual, and R is a role.

Theorem 1 ([28]). O |=4 α if and only if O |=2 λ(α) for an ontology O,

where |=2 is the entailment in ALC.

7
2.2. The Logic of Minimal Knowledge and Negation as Failure
The Logic of Minimal Knowledge and Negation as Failure (MKNF) [27]
has been proposed as a unifying framework for different nonmonotonic for-
malisms, such as default logic, autoepistemic logic, and logic programming
[33]. It is a variant of first-order modal logic with two modal operators: K
and not.
Let Σ be a signature that consists of constants and function symbols,
first-order predicates, and the binary equality predicate ≈. A first-order atom
P(t1 , . . . , tl ) is an MKNF formula, where P is a first-order predicate and ti are
first-order terms. Other MKNF formulae are built over Σ by using standard
connectives in first-order logic and two modal operators as follows: true, ¬φ,
φ1 ∧φ2 , ∃x : φ, Kφ, notφ. Moreover, formulae false, φ1 ∨φ2 , φ1 ⊃ φ2 , ∀x : φ,
φ1 ≡ φ2 , t1 ≈ t2 and t1 ̸≈ t2 are abbreviations, respectively, for formulae
¬true, ¬(¬φ1 ∧¬φ2 ), ¬φ1 ∨φ2 , ¬(∃x : ¬φ), (φ1 ⊃ φ2 ) ∧ (φ2 ⊃ φ1 ), ≈ (t1 , t2 ),
¬(t1 ≈ t2 ). First-order atoms of the form t1 ≈ t2 and t1 ̸≈ t2 are called
equalities and inequalities, respectively, and have a predefined interpretation
[10]. Formulae of the form Kφ (notφ) are called modal K-atoms (not-
atoms); modal K-atoms and not-atoms are called modal atoms. An MKNF
formula φ is called closed if it contains no free variables, and called ground if
it is without any variables. An MKNF formula φ is called modally closed if it
is closed and all modal operators are applied to closed subformulae. Formula
φ[t/x] is obtained from φ by substituting the term t for the variable x.
Let △ be a universe, which contains an infinite supply of constants, be-
sides constants occurring in the formulae. Just like in first-order logic, a
first-order interpretation I over Σ and △ assigns an object aI ∈ △ to each
constant a ∈ Σ, a function f I : △n → △ to each function f ∈ Σ and a
relation P I ⊆ △n to each predicate P ∈ Σ. Moreover, it interprets ≈ as
equality predicate, that is to say, ≈ (t1 , t2 ) iff t1 = t2 . Unlike in first-order
logic, for each element α ∈ △, the signature Σ is required to contain a special
constant nα , called a name, such that nIα = α.
An MKNF structure is a triple (I, M, N ), where I is a first-order inter-
pretation over Σ and △, while M and N are nonempty sets of first-order
interpretations. Satisfaction of a closed MKNF formula φ is defined induc-
tively as follows:
(I, M, N) |= true for each structure (I, M, N)
(I, M, N) |= P(t1 , . . . , tl ) iff P(t1 , . . . , tl ) ∈ I
(I, M, N) |= ¬φ iff (I, M, N) ̸|= φ
(I, M, N) |= φ1 ∧ φ2 iff (I, M, N) |= φ1 and (I, M, N) |= φ2

8
(I, M, N) |= ∃x : φ iff (I, M, N) |= φ[nα /x] for some α ∈ △
(I, M, N) |= Kφ iff (J, M, N) |= φ for all J ∈ M
(I, M, N) |= notφ iff (J, M, N) ̸|= φ for some J ∈ N
An MKNF interpretation M over a universe △ is a nonempty set of first-
order interpretations. M satisfies a closed MKNF formula φ, written M |= φ,
iff (I, M, M) |= φ for every I ∈ M . M is an MKNF model of φ if: (i) (I, M,
M) |= φ for each I ∈ M ; (ii) for each set of first-order interpretations M′
such that M′ ) M, we have (I ′ , M′ , M) ̸|= φ for some I ′ ∈ M ′ . An MKNF
formula φ is MKNF satisfiable if an MKNF model of φ exists; otherwise φ
is MKNF unsatisfiable. Furthermore, φ |=M KN F ψ if and only if M |= ψ for
each MKNF model M of φ.
Note that the definition of MKNF model indicates the preference of the
maximal set M that satisfies φ. The bigger the MKNF model is, the less
knowledge we get from the knowledge base. In fact, if M1 ⊆ M2 , then M1
satisfies φ if M2 satisfies φ.

Example 3. Consider the MKNF formula φ = not q ⊃ Kp, where p and q

are two propositional atoms. M = {{p}, {p, q}} is the MKNF model of φ.

As argued in [33], the MKNF semantics has two undesirable properties.

One is counterintuitive semantics caused by an arbitrary universe, and the
other one is different constants in different interpretations, which leads to
the counterintuitive semantics of existential quantification. Therefore, the
standard name assumption was adopted in [33].

Definition 1. (Standard Name Assumption [33]) A first-order interpreta-

tion I over a signature Σ employs the standard name assumption if

(1) the universe △ of I contains all constants of Σ and a countably infinite

number of additional constants called parameters;

(2) tI = t for each ground term t constructed using the function symbols
from Σ and the constants from △; and

(3) the predicate ≈ is interpreted in I as a congruence relation – that is,

≈ is reflexive, symmetric, transitive, and allows for the replacement of
equals by equals [10].

Therefore, Herbrand first-order interpretations, in which each constant

is interpreted by itself, are used to replace the first-order interpretations in

9
MKNF interpretations. Moreover, it has been proved in [10] that each first-
order formula is satisfiable iff it is satisfiable in a model that employs the
standard name assumption.

2.3. Hybrid MKNF Knowledge Bases

Hybrid MKNF, based on the MKNF logic, is a mature approach for
integrating Description Logics and Logic Programming, proposed by Boris
Motik and Riccardo Rosati [32, 33]. Hybrid MKNF knowledge bases consist
of a finite number of MKNF rules and a decidable description logic knowledge
base which can be translated to first-order logic.
More concretely, the approach based on Hybrid MKNF knowledge bases
is applicable to any first-order fragment DL that satisfies these conditions:
(i) each knowledge base O ∈ DL can be translated to a formula π(O) of
function-free first-order logic with equality (see [1] for standard translation
for Description Logic axioms), (ii) it supports ABox-assertions of the form
P (t1 , . . . , tl ), where P is a predicate and each ti is a constant of DL, and (iii)
satisfiability checking and instance checking (i.e., checking entailments of the
form O |= P (t1 , . . . , tl )) are decidable. Note that description logics around
OWL satisfy theses conditions. The restriction to function-free first-order
logic guarantees the decidability of the language. Therefore, in the rest of
the paper, we will not allow function symbols in hybrid MKNF knowledge
bases.

Definition 2 ([33]). A formula ξG is a grounding of a nonground formula

ξ if ξG is obtained from ξ by replacing its free variables with constants. A set
of generalized atoms GA is a set of first-order formulae such that, if ξ ∈ GA,
then ξG ∈ GA for each grounding ξG of ξ. A generalized atom is ground if
it does not contain free variables.

Definition 3. Let O be a DL knowledge base. An MKNF rule r has the

following form, where αi , βi , γi are function-free generalized atoms:

Kα1 ∨ . . . ∨ Kαn ← Kβn+1 ∧ . . . ∧ Kβm ∧ notγm+1 ∧ . . . ∧ notγk (5)

The sets {Kαi }, {Kβi }, {notγi } are called the rule head, the positive body
and the negative body, respectively. An MKNF rule r is nondisjunctive if
n = 1; r is positive if m = k; r is a fact if m = k = 0. A program P is
a finite set of MKNF rules. A hybrid MKNF knowledge base K is a pair
(O,P).

10
 Mapping Concepts to FOL
¨ X) = ⊤
πy (⊤, ¨ , πy (⊥,¨ X) = ⊤ ¨
πy (A, X) = A(X), πy (¬ C, X) = ¬ πy (C, X)
πy (C1 ⊓ C2 , X) = πy (C1 , X) ∧ πy (C2 , X)
πy (C1 ⊔ C2 , X) = πy (C1 , X) ∨ πy (C2 , X)
πy (∃R.C, X) = ∃y : R(X, y) ∧ πx (C, y)
πy (∀R.C, X) = ∀y : R(X, y) → πx (C, y)
Mapping Axioms to FOL
π(C(a)) = πy (C, a), π(R(a, b)) = R(a, b)
π(C ⊑ D) = ∀x : πy (C, x) → πy (D, x), π(a ≈ b) = a ≈ b
where X is a meta variable and is substituted by actual variable, and
πx is defined as πy by substituting x and xi for y and yi , respectively.

Table 2: Translation of ALC into FOL

To obtain a practically useful formalism, the language GA should at least

include the standard function-free first-order atoms of the form P (t1 , . . . , tl ),
negative literals of the form ¬P (t1 , . . . , tl ), and conjunctive queries over DL.
In this section, we refer to these three types of generalized atoms.
Hybrid MKNF knowledge bases, as defined above, are not of the form of
MKNF knowledge bases. In order to semantically interpret hybrid MKNF
knowledge bases, the transformation π (see Table 2) that translates descrip-
tion logic expressions to first-order formulae is extended to MKNF rules and
hybrid MKNF knowledge bases as follows:
Definition 4. Let K = (O,P) be a hybrid knowledge base. We extend π to
r of form (5), P, and K as follows, where x is the vector of the free variables
of MKNF rule r:

π(r) = ∀x : (Kα1 ∨ . . . ∨ Kαn ⊂ Kβn+1 ∧ . . . ∧ Kβm ∧

notγm+1 ∧ . . . ∧ notγk )
∧
π(P) = r∈P π(r) π(K) = Kπ(O) ∧ π(P)
K is satisfiable if and only if an MKNF model of π(K) exists, and K entails
a closed MKNF formula φ, written K |= φ if and only if π(K)|=M KN F φ.
To ensure that the MKNF logic is decidable, DL-safety is introduced as
a restriction to MKNF rules.

11
Definition 5 ([33]). We assume that the signature Σ contains a subset
ΣDL ⊆ Σ such that ≈ ∈ ΣDL . We call the predicates in ΣDL DL-predicates
and assume that DL refers only to such predicates; furthermore, we call the
predicates in Σ \ ΣDL non-DL-predicates.
A generalized atom ξ is a DL-atom if it contains only predicates from
ΣDL ; furthermore, ξ is a non-DL-atom if it is of the form (¬)P (t1 , . . . , tn )
for P a non-DL-predicate. A modal atom Kξ or notξ is a DL-atom or
non-DL-atom if ξ has the respective property.
An MKNF rule is DL-safe if each modal atom in it is either a DL- or
a non-DL-atom, and if every variable in r occurs in the body of r in some
non-DL-atom of the form Kβ. A hybrid MKNF knowledge base K is DL-safe
if each rule r ∈ P is DL-safe.
In the rest of this paper, without explicitly stating it, hybrid MKNF
knowledge bases are considered to be DL-safe.

Definition 6. Given a hybrid MKNF knowledge base K = (O,P). The

ground instantiation of K is the knowledge base KG = (O,PG ), where PG
is obtained from P by replacing each rule r of P with a set of rules substi-
tuting each variable in r with constants from K in all possible ways.

Grounding the knowledge base K ensures that rules in P apply only to

objects that occur in K. And it was shown in [33] that the MKNF models
of K and KG coincide.

Example 4. Consider a DL knowledge base O consisting of axioms (1) and

(4) from Example 2, and the set P consisting of the rules (6) and (7) below.
Let K = (O, P). Predicates from O start with an uppercase letter and others
with a lowercase letter.

KF ly(x) ← Kanimal(x), KBird(x), notP enguin(x) (6)

KP enguin(T weety) ← (7)

The knowledge base K describes correctly the statement in Example 1 – that

is, most birds can fly, with some exceptions, such as penguins. The excep-
tion is expressed by the not operator, and closed world reasoning is used for
Penguin in (6). Note that the non-DL-atom animal ensures the DL-safety of
rule (6).

12
 Hybrid MKNF knowledge bases provide a paradigm for representing data
sources on the web using rules and description logics simultaneously. Local
closed world reasoning in the knowledge bases bridges the rules and DLs,
and accordingly overcomes the expressive limitation of rules and DLs, and
enhances the expressivity.
However, real knowledge bases will be distributed and multi-authored. It
is unreasonable to require every knowledge base to be logically consistent.
Inconsistencies may arise when rules and DLs are reconciled in hybrid MKNF
knowledge bases, even if the rule-part and DL-part are consistent if taken
alone.

Example 5. Consider a hybrid MKNF knowledge base K = (O,P). The DL

part O consists of axioms (1) to (3) from Example 2, and P = {K¬F ly(x) ←
KP enguin(x)}. Note that O and P are consistent knowledge bases, respec-
tively. However, the combination causes the inconsistency of F ly(T weety).
Therefore K has no MKNF model.

In Example 5, classical reasoning broke down due to the inconsistency of

F ly(T weety). Some useful information will be lost, e.g., P enguin(T weety).
In order to handle these problems, we will present a paraconsistent semantics
for hybrid MKNF knowledge bases.

3. Para-MKNF Knowledge Bases with Paraconsistent Semantics

For distinguishing the two hybrid MKNF knowledge bases with stable
model semantics and paraconsistent semantics, we call the latter para-MKNF
knowledge base. Now we describe its syntax and semantics.
Syntactically, para-MKNF knowledge bases hardly differ from hybrid
MKNF KBs, and similarly a para-MKNF KB K has two components: O
and P, where O is an ALC KB and P is a nonempty set of MKNF rules that
differ only slightly from the ones in Definition 3. In the previous section, we
have mentioned that there are three kinds of implication connectives in four-
valued logic. However, in our paper, we will employ inclusion implication [2],
as recent research [28, 30] has shown that it has desirable properties which
the other two lack.2

2
In particular, as shown in [30], inclusion implication scales much better in implemen-
tations, and it also preserves polynomial time complexity for important DLs such as those
underlying the Web Ontology Language OWL [31].

13
Definition 7. Let O be a DL knowledge base. L is a literal if it is an atom
P, or of the form ¬P , where P is an atom. An MKNF rule r has the follow-
ing form, where Hi , Ai , Bi are first-order function-free literals:

KH1 ∨ . . . ∨ KHn ← KAn+1 ∧ . . . ∧ KAm ∧ notBm+1 ∧ . . . ∧ notBk (8)

The sets {KHi }, {KAi }, {notBi } are called the rule head, the positive body
and the negative body, respectively. An MKNF rule r is nondisjunctive if
n = 1; r is positive if m = k; r is a fact if m = k = 0. A program P is a
finite set of MKNF rules. A para-MKNF knowledge base K = (O, P) is a
pair of a DL knowledge base O and a program P.

Note that, for simplicity, we substitute literals for first-order formulae

in MKNF rules. In our paradigm, negative literals have the same status
as positive literals, and we consider a modified version of Herbrand first-
order interpretations, namely the set of ground literals occurring in K, and
call them paraconsistent Herbrand first-order interpretations. Negative liter-
als are allowed in paraconsistent Herbrand first-order interpretations, while
Herbrand first-order interpretations are restricted to first-order atoms.

Definition 8. A four-valued (paraconsistent) MKNF structure (I, M, N )

consists of a paraconsistent Herbrand first-order interpretation I and two
nonempty sets of paraconsistent Herbrand first-order interpretations M and
N . A nonempty set of paraconsistent Herbrand first-order interpretations M
is called a paraconsistent MKNF interpretation.

I is supposed to interpret first-order formulae, while M and N are used to

evaluate modal K-atoms and modal not-atoms, respectively. Every MKNF
formula is assigned to an element in the bilattice FOUR.

Definition 9. Let (I, M, N ) be a paraconsistent MKNF structure, and {t, f ,

⊥, ⊤} with partial order ≤k be the set of truth values. We evaluate MKNF
formulae inductively as follows:


 t iff L(t1 , . . . , tl ) ∈ I and ¬L(t1 , . . . , tl ) ∈
/I

f iff L(t1 , . . . , tl ) ∈
/ I and ¬L(t1 , . . . , tl ) ∈ I
(I, M, N )(L(t1 , . . . , tl )) =

 ⊤ iff L(t1 , . . . , tl ) ∈ I and ¬L(t1 , . . . , tl ) ∈ I


⊥ iff L(t1 , . . . , tl ) ∈/ I and ¬L(t1 , . . . , tl ) ∈
/I

14
 

t iff (I, M, N )(φ) = f

f iff (I, M, N )(φ) = t
(I, M, N )(¬φ) =
⊤
 iff (I, M, N )(φ) = ⊤


⊥ iff (I, M, N )(φ) =⊥
(I, M, N )(true) = ⊤, for each paraconsistent structure (I, M, N )
(I, M, N )(false) =⊥, for each paraconsistent structure (I, M, N )
(I, M, N )(φ1 ∧ φ2 ) = (I, M, N )(φ1 ) ∧ (I, M, N )(φ2 )
(I, M, N )(φ1 ∨ φ2 ) = (I, M, N )(φ1 ) ∨ (I, M, N )(φ2 )
∨
(I, M, N )(∃x : φ) = (I, M, N )(φ[nα /x])
α∈△
∧
(I, M, N )(∀x : φ) = (I, M, N )(φ[nα /x])
α∈△
∧
(I, M, N )(Kφ) = (J , M, N )(φ)
J ∈M
{
t iff (I, M, N )(φ1 ) ∈ {f , ⊥}
(I, M, N )(φ1 ⊃ φ2 ) =
(I, M, N )(φ2 ) otherwise


 t iff ∃ J ∈ N s. t. (J , M, N )(φ) = f



 and no other J ∈ N , s. t. (J , M, N )(φ) =⊥





f iff (J , M, N )(φ) = t for all J ∈ N ;
(I, M, N )(notφ) = or ∃J ∈ N such that (J , M, N )(φ) = t,



 and for other J ∈ N , (J , M, N )(φ) = ⊤



 ⊤ iff (J , M, N )(φ) =⊥ for some J ∈ N



⊥ iff (J , M, N )(φ) = ⊤ for all J ∈ N
Moreover, as defined in classical semantics, formulae φ1 ≡ φ2 , t1 ≈ t2 and
t1 ̸≈ t2 are abbreviations, respectively, for formulae (φ1 ⊃ φ2 ) ∧ (φ2 ⊃ φ1 ),
≈ (t1 , t2 ), ¬(t1 ≈ t2 ).

As in two-valued semantics, the evaluation of notφ basically follows the

“mirror” idea of Kφ, except for the case for φ being evaluated in some to
true and in some to false.

15
Definition 10. Let (I, M, N ) be a paraconsistent MKNF structure. Para-
consistent satisfaction of closed MKNF formulae is defined inductively as
follows:
(I, M, N ) |=4 true for each paraconsistent structure (I, M, N )
(I, M, N ) |=4 P(t1 , . . . , tl ) iff PI (t1 , . . . , tl ) ∈ {t, ⊤}
(I, M, N ) |=4 ¬φ iff (I, M, N )(φ) ∈ {f , ⊤}
(I, M, N ) |=4 φ1 ∧ φ2 iff (I, M, N ) |=4 φi , i = 1, 2
(I, M, N ) |=4 φ1 ∨ φ2 iff (I, M, N ) |=4 φ1 or (I, M, N ) |=4 φ2
(I, M, N ) |=4 ∃x : φ iff (I, M, N ) |=4 φ[nα /x] for some α ∈ △
(I, M, N ) |=4 ∀x : φ iff (I, M, N ) |=4 φ[nα /x] every α ∈ △
(I, M, N ) |=4 φ1 ⊃ φ2 iff (I, M, N ) ̸|=4 φ1 or (I, M, N ) |=4 φ2
(I, M, N ) |=4 Kφ iff (J , M, N ) |=4 φ for all J ∈ M
(I, M, N ) |=4 notφ iff (J , M, N ) ̸|=4 φ for some J ∈ N

It can be easily verified that Definition 9 of paraconsistent semantics

is compatible with Definition 10 of paraconsistent satisfaction, and we will
work mostly with the latter. For a closed MKNF formula φ, a paraconsistent
MKNF interpretation M paraconsistently satisfies φ, written M |=4 φ, iff
(I, M, M) |=4 φ for each I ∈ M.

Definition 11. A paraconsistent MKNF interpretation M is a paraconsis-

tent MKNF model of a given closed MKNF formula φ, written M|=4M KN F φ
if and only if the following two conditions are satisfied :

(1) M paraconsistently satisfies φ;

(2) for each paraconsistent MKNF interpretation M′ , such that
M′ ) M, we have (I ′ , M′ , M) ̸|=4 φ, for some I ′ ∈ M′ .

For a para-MKNF knowledge base K = (O, P), K is paraconsistently

MKNF satisfiable iff a paraconsistent MKNF model of π(K) exists, where
π(K) is defined as in Section 2.3. φ para-MKNF entails ϕ, written φ |=4M KN F
ϕ, iff M |=4 ϕ for each paraconsistent MKNF model M of φ.
When a para-MKNF knowledge base K has consistent MKNF models,
we distinguish these consistent MKNF models as preferred MKNF models.
A consistent MKNF model M means that for all Kξ such that M |=4 Kξ,
no Kξ and K¬ξ hold simultaneously. A para-MKNF knowledge base K is
consistent if there exists a preferred MKNF model. The preferred MKNF
model coincides with the MKNF model under the two-valued semantics.

16
Example 6. Consider the para-MKNF knowledge base K in Example 5. K
has a paraconsistent MKNF model M = {I | I |=4 {P enguin(T weety),
F ly(T weety), ¬F ly(T weety), Bird(T weety)}}.

Paraconsistent semantics in our paradigm is faithful. That is to say, the

semantics yields the paraconsistent semantics for DLs according to [28, 29, 30]
when no rules are present, and the p-stable model of LP from [38] when the
DL-component is empty.
In order to show this property, we need to recall the notion of p-stable model
of an answer set program.
An extended disjunctive program P is a finite set of clauses of the form

L1 ∨ . . . ∨ Ll ← Ll+1 ∧ . . . ∧ Lm ∧ notLm+1 ∧ . . . ∧ notLn (9)

where the Li are literals, 1 ≤ i ≤ n. P is called a positive extended disjunctive

program if m = n.

Definition 12. (Paraconsistent Semantics of ASP [38]) Let P be an ex-

tended disjunctive program and I be a subset of the Herbrand base of P. The
reduct of P w.r.t. I is the positive extended disjunctive program P I such that
a clause
L1 ∨ . . . ∨ Ll ← Ll+1 ∧ . . . ∧ Lm
is in P I iff there is a ground clause of form (9) from P such that {Lm+1
, . . . , Ln } ∩ I = ∅. For a positive extended disjunctive program P , an inter-
pretation I is a model of P if I satisfies every ground clause from P , and a
p-minimal model if there exists no model J of P such that J ⊂ I.
Then I is called a paraconsistent stable model (shortly, p-stable model)
of P if I is a p-minimal model of P I .

Proposition 1. Let K = (O,P) be a para-MKNF knowledge base, φ a closed

first-order formula, and A a ground literal.

• If P = ∅, then K |=4M KN F φ iff O |=4 φ.

• If O = ∅, then K |=4M KN F A iff P |=4 A, where P |=4 A means for all

the p-stable models I of P, that I |=4 A.

17
Proof. Case 1: P = ∅. (Necessity) K consists only of DL axioms. Thus
K |=4M KN F φ equals to Kπ(O) |=4M KN F φ, which means M |=4 φ for each
paraconsistent MKNF model M of Kπ(O). Therefore, I ∈ M iff I is a
paraconsistent model of π(O). From M |=4 φ we can infer π(O) |=4 φ,
which is the first-order form of O |=4 φ.
(Sufficiency) Similarly, the sufficiency can easily be proved by contradiction.
Case 2: O = ∅. K consists only of MKNF rules. K |=4M KN F φ if and only
if π(P) |=4M KN F φ. We denote the program of the form (9) corresponding
to P by P . First we have a property (†): M is a paraconsistent ∩ MKNF
model of K iff I is a p-stable model of program P such that I = I∈M M.
Then π(P) |=4M KN F φ infers that for every paraconsistent ∩ MKNF model M
of π(P), M |=4 φ. Then we have I |=4 φ, where I = I∈M M. From (†), we
can infer that the p-stable model I of program P is one-to-one corresponding
to the paraconsistent MKNF model M of P. Then we obtain that P |=4 φ.
The converse can be proved similarly.
Proof∩of (†): (Necessity) Let M be a paraconsistent MKNF model of K
and I = I∈M M. For every rule r in P I of form (9) such that ∩ {Lm+1 , . . . ,
Ln } ∩ I = ∅, if {Ll+1 , . . . , Lm } ⊆ I, then {Ll+1 , . . . , Lm } ⊆ I∈M M, and
then M |=4 KLi , for every i, l + 1 ≤ i ≤ m. Since M is the paraconsistent
MKNF model of P, we can infer that there exists Lj , 1 ≤ j ≤ l, such that
M |=4 KLj , then we obtain that Lj ∈ I and I satisfies every rule in P I .
Next we prove the minimality of I. Suppose there exists an interpretation J,
such that J ⊂ I and J satisfies P I . Let M′ = {J ′ | J ′ ⊇ J}, then M ⊂ M′ .
For every MKNF rule r of form (8), if M′ |=4 notLj , m + 1 ≤ j ≤ k, and
M′ |=4 KLi , l + 1 ≤ i ≤ m, then Lj ∈ / J, for every m + 1 ≤ j ≤ k, and
Li ∈ J, for every l + 1 ≤ i ≤ m. Since J satisfies P I , we obtain that there
exists Li , 1 ≤ i ≤ l, such that Li ∈ J. Thus M′ |=4 KLi , and M′ |=4 P,
which contradicts the fact that M is the paraconsistent MKNF model of K.
Hence I is a p-stable model of P . (Sufficiency) Sufficiency can be proved
similarly. We omit the details. 

For a para-MKNF knowledge base K = (O,P), let PG be the set of rules

obtained from P by replacing in each rule all variables with all constants
from K in all possible ways; the knowledge base KG = (O,PG ) is called the
ground instantiation of K.
As argued in [33], for a DL-safe hybrid knowledge base K, the two-valued
MKNF models of KG and K coincide. In our paradigm, the same conclusion
holds for paraconsistent MKNF models.

18
4. Transformation from Para-MKNF Knowledge Bases to Hybrid
MKNF Knowledge Bases
Y. Ma et al. [28] have proposed a transformation λ from ALC4 to ALC.
In this section, we extend λ to MKNF rules and then transform para-MKNF
knowledge bases to hybrid MKNF knowledge bases based on the transforma-
tion operator.
Given a para-MKNF knowledge base K = (O,P), the transformation
operator λ transforms every axiom in O as presented in Section 2.1. In this
paper, MKNF rules are made up from KAi , notBj and some connectors,
where Ai and Bj are literals. We first define the transformation on modal
atoms.

- If φ = Kψ, then λ(φ) = Kλ(ψ), where ψ is a literal;

- If φ = notψ, then λ(φ) = notλ(ψ), where ψ is a literal.

Based on this, MKNF rules of form (8) are transformed to λ(KH1 )∨ . . .∨

λ(KHn ) ← λ(KAn+1 ) ∧ . . . ∧ λ(KAm ) ∧ λ(notBm+1 ) ∧ . . . ∧ λ(notBk ). Then
the para-MKNF knowledge base K is transformed inductively to a hybrid
MKNF knowledge base, denoted by K. We say K is classically induced by
the para-MKNF knowledge base K, if all the axioms and rules in K are
exactly the transformations of the axioms and rules in K. We also extend λ
to MKNF formulae such that modal operators K and not are restricted to
literals.

- If φ = P (t1 , . . . , tl ), then λ(φ) = P+ (t1 , . . . , tl ), where P (t1 , . . . , tl ) is a

first-order atom occurring in K and P+ (t1 , . . . , tl ) is a new first-order
atom;

- If φ = ¬P (t1 , . . . , tl ), then λ(φ) = P− (t1 , . . . , tl ), where P− (t1 , . . . , tl )

is a new first-order atom;

- If φ = φ1 ∧ φ2 , then λ(φ) = λ(φ1 ) ∧ λ(φ2 ), where φ1 and φ1 are two

MKNF formulae;

- If φ = φ1 ∨ φ2 , then λ(φ) = λ(φ1 ) ∨ λ(φ2 );

- If φ = ∃x : ψ, then λ(φ) = ∃x : λ(ψ);

- If φ = ∀x : ψ, then λ(φ) = ∀x : λ(ψ);

19
 - If φ = ¬¬ψ, then λ(φ) = λ(ψ);
- If φ = ¬(φ1 ∧ φ2 ), then λ(φ) = λ(¬φ1 ) ∨ λ(¬φ2 );
- If φ = ¬(φ1 ∨ φ2 ), then λ(φ) = λ(¬φ1 ) ∧ λ(¬φ2 );
- If φ = ¬(∃x : ψ), then λ(φ) = ∀x : λ(¬ψ);
- If φ = ¬(∀x : ψ), then λ(φ) = ∃x : λ(¬ψ);
- If φ = ¬Kψ, then λ(φ) = Kλ(¬ψ), where ψ is a literal;
- If φ = ¬notψ, then λ(φ) = notλ(¬ψ), where ψ is a literal.
An MKNF interpretation of K can be induced by a paraconsistent inter-
pretation of K. First of all, we define the interpretation structure of K.

Definition 13 (Classical Induced MKNF Structure). Let (I, M, N ) be

a paraconsistent MKNF structure of a para-MKNF knowledge base K, and K
be the classical induced knowledge base of K. The classical induced MKNF
structure of (I, M, N ), written (I, M, N ), is defined as follows:
1. △ = △, where △ represents the universe of the classical induced
interpretation I;
2. for a first-order atom P (t1 , . . . , tl ),
{
I t iff P I (t1 , . . . , tl ) ∈ {t, ⊤}
P+ (t1 , . . . , tl ) = (10)
f iff P I (t1 , . . . , tl ) ∈ {f , ⊥}
{
t iff ¬P I (t1 , . . . , tl ) ∈ {t, ⊤}
P− I (t1 , . . . , tl ) = (11)
f iff ¬P I (t1 , . . . , tl ) ∈ {f , ⊥}
and we call I a classical induced interpretation;
3. M is a nonempty set corresponding to M and consists of classical
induced interpretations I such that I ∈ M. N is a nonempty set
corresponding to N and consists of classical induced interpretations
I such that I ∈ N .

Conversely, given an MKNF structure of a hybrid MKNF knowledge base

K, we can define the four-valued induced MKNF structure of K easily. Par-
ticularly, when K is consistent, the four-valued induced MKNF structure
coincides with the original MKNF structure.

20
Definition 14. Given a hybrid MKNF knowledge base K, let (I, M, N ) be
a MKNF structure of a hybrid MKNF knowledge base K. The four-valued
induced MKNF structure of (I, M, N ), written (I, M, N ), is defined as
follows:

1. △ = △, where △ represents the universe of the four-valued induced

interpretation I;
2. for a first-order atom P (t1 , . . . , tl ),



 t iff P+ (t1 , . . . , tl ) ∈ I and P− (t1 , . . . , tl ) ∈
/I

f iff P+ (t1 , . . . , tl ) ∈
/I and P− (t1 , . . . , tl ) ∈ I
P I (t1 , . . . , tl ) =

 ⊤ iff P+ (t1 , . . . , tl ) ∈ I and P− (t1 , . . . , tl ) ∈ I

⊥ iff P+ (t1 , . . . , tl ) ∈
/I and P− (t1 , . . . , tl ) ∈
/I
and we call I a four-valued induced interpretation;
3. M is a nonempty set corresponding to M and consists of four-valued
induced interpretations I such that I ∈ M. N is a nonempty set
corresponding to N and consists of four-valued induced interpretations
I such that I ∈ N .

Lemma 1. For a paraconsistent MKNF structure (I, M, N ) of a para-MKNF

knowledge base K and any MKNF formula ϕ such that modal operators K
and not are restricted to literals, we have
{
t iff (I, M, N )ϕ ∈ {t, ⊤}
(I, M, N )λ(ϕ) = (12)
f iff (I, M, N )ϕ ∈ {f , ⊥}
{
t iff (I, M, N )¬ϕ ∈ {t, ⊤}
(I, M, N )λ(¬ϕ) = (13)
f iff (I, M, N )¬ϕ ∈ {f , ⊥}

Proof. Let φ be a literal. From Definition 13, it is easy to verify that φ

satisfies the equations (12) and (13). First of all, we prove the equations (12)
and (13) for Kφ and notφ inductively.
Case 1: ϕ = Kφ.
Equation (12): (I, M, N ) λ(Kφ) = t iff (J , M, N )λ(φ) = t for every
J ∈ M. From the assumption, (J , M, N )λ(φ) = t for every J ∈ M, iff
(J , M, N )φ ∈ {t, ⊤} for every J ∈ M, which means that (I, M, N )Kφ ∈
{t, ⊤}. Then we obtain that (I, M, N )λ(Kφ) = t iff (I, M, N )Kφ ∈ {t, ⊤}.

21
 Equation (13): (I, M, N ) (¬Kφ) = t or ⊤ iff (I, M, N ) (Kφ) = f or
⊤. If (I, M, N ) (Kφ) = f , then there exists an interpretation J ∈ M such
′
that φJ = f and no J ′ ∈ M such that φJ = ⊥ or t. Then ¬φJ = t or
⊤ for each J ∈ M and hence λ(¬φ)J = t for each J ∈ M. Therefore, (I,
M, N ) λ(¬Kφ) = t. If (I, M, N ) (Kφ) = ⊤, then for each J ∈ M such
that φJ = ⊤, and hence for each J ∈ M, ¬φJ = ⊤. Thus for each J ∈ M,
λ(¬φ)J = t. Therefore, (I, M, N ) λ(¬Kφ) = t. The other side can be
proved similarly. We omit the details here.
Case 2: ϕ = notφ.
Equation (12): (I, M, N ) λ(notφ) = t iff (J , M, N )λ(φ) = f for some
J ∈ M. From the assumption, (J , M, N )φ ∈ {f , ⊥} for some J ∈ M,
which means that (I, M, N )notφ ∈ {t, ⊤} from Definition 9. Therefore we
obtain that (I, M, N )λ(notφ) = t iff (I, M, N )notφ ∈ {t, ⊤}.
Equation (13): (I, M, N ) (¬notφ) = t or ⊤ iff (I, M, N ) (notφ) = f
or ⊤. If (I, M, N ) (notφ) = f , then there exists at least one J ∈ N such
′
that φJ = t and for other J ′ ∈ M, φJ = ⊤. Then there exists J ∈ N
such that ¬φJ = f and hence there exists J ∈ N such that λ(¬φ)J = f .
Therefore (I, M, N ) notλ(¬φ) = t. If (I, M, N ) (notφ) = ⊤, then there
exists a J ∈ N such that φJ = ⊥. Then there exists J ∈ N such that ¬φJ
= ⊥, and hence there exists J ∈ N such that λ(¬φ)J = f . Therefore, (I,
M, N ) notλ(¬φ) = t. Similarly, the other side can be proved. We omit the
details.
It can be easily verified that equations (12) and (13) hold for MKNF
formulae of the other forms. We omit the proofs here. 

From Lemma 1, we can get an important conclusion as follows:

Theorem 2. For a para-MKNF knowledge base K and a closed MKNF for-

mula φ, we have K |=4M KN F φ iff K |=M KN F λ(φ).

Proof. (Necessity) Suppose M is a paraconsistent MKNF model of K,

we only need to prove that the classical induced interpretation M of M is
an MKNF model of K. M |=4M KN F π(K) implies M |=4M KN F π(P) and
M |=4M KN F Kπ(O).
Let α be an axiom in O. M |=4M KN F Kπ(O) implies I |=4 π(α) for each
I ∈ M. Then I |=2 π(λ(α)) by Theorem 1, and hence M |= Kπ(λ(α)).
Therefore, M |= O.
For MKNF rules of form (8), (I, M, M) |=4 r iff (I, M, M) |=4 KHi ,
for some 1 ≤ i ≤ n, or (I, M, M) ̸|=4 KAj , for some n + 1 ≤ j ≤ m, or

22
(I, M, M) |=4 notBt , for some m + 1 ≤ t ≤ k, each case corresponding
to the case (I, M, M) |= λ(KHi ), (I, M, M) ̸|= λ(KAj ), or (I, M, M) ̸|=
λ(notBt ), respectively. Each of the cases means (I, M, M) |= λ(r).
If there exists an MKNF interpretation M′ ) M such that (I ′ , M′ , M)
π(K) = t, then this MKNF interpretation structure must satisfy all the
induced axioms and rules, which contradicts the maximality of M.
(Sufficiency) For any MKNF interpretation M of K , let M be the four-
valued semantics of M. Similarly, we can prove that M is an MKNF model
of K if M is an MKNF model of K. We omit the details here. 

Note that the transformation operator is linear. Thus from Theorem 2,

we can conclude that the data complexity of our paradigm is not higher than
that of classical reasoning3 .

5. Characterization of Paraconsistent MKNF Models

In this section we present several fixpoint characterizations of paracon-
sistent MKNF models, each suitable to different kinds of rules. According to
the discussion in Section 3, a para-MKNF knowledge base K = (O,P) has
exactly the same paraconsistent MKNF models as KG . Therefore in the rest
of the paper, we only consider grounded knowledge bases KG .

Definition 15. Let KG be a ground para-MKNF knowledge base. The set

of K-atoms of KG , written KA(KG ), is the smallest set that contains (1) all
ground K-atoms occurring in PG , and (2) a modal atom Kξ for each ground
modal atom notξ occurring in PG . Furthermore, HA(KG ) is the subset of
KA(KG ) that contains all K-atoms occurring in the head of some rule in PG .

As argued in [33], MKNF models of KG are determined by subsets of

HA(KG ). The same holds for paraconsistent MKNF models.

Example 7. We consider the following para-MKNF knowledge base KG ex

=
(O ,OG ), where O consists of axiom (14) and PG consists of MKNF rules
ex ex ex ex

(15)-(17):

q⊃r (14)

3
Data complexity of reasoning in MKNF KBs estimates the performance of reasoning
algorithms measured in terms of ABox assertions and MKNF rules.

23
Kq ← notp (15)

Ks ← notr (16)

K¬q ← (17)

G = {I | I |=4 {q ∧ ¬q ∧ (q ⊃ r)}}. It can be verified that MG is a

Let Mex ex

paraconsistent MKNF model of KG ex

.

Definition 16. Let KG be a ground para-MKNF knowledge base, and Ph a

subset of HA(KG ). The objective knowledge of Ph w.r.t. HA(KG ) is the
first-order theory OBO,Ph defined by
OBO,Ph = {π(O)} ∪ { ξ | K ξ ∈ Ph }.

In Example 7, let Phex = {Kq, K¬q}. The paraconsistent MKNF model

Mex
G equals {I | I |=4 OBOex ,Phex }.

Definition 17. For a paraconsistent MKNF interpretation M and a set of

ground K-atoms S, the subset of S paraconsistently induced by M is the set
{K ξ ∈ S | M |=4 ξ}.

Lemma 2. Let KG = (O,PG ) be a ground para-MKNF knowledge base, M

be a paraconsistent MKNF model of KG , and Ph be the subset of HA(KG )
paraconsistently induced by M. Then M coincides with the set of paracon-
sistent MKNF interpretations M′ = {I | I |=4 OBO,Ph }.

Proof. Given I ∈ M, since M is a paraconsistent MKNF model of KG ,

we clearly have I |=4 O. For ξ such that Kξ ∈ Ph , M |=4 ξ from the fact
that Ph is paraconsistently induced by M. Then I |=4 ξ, thus I ∈ M′ , and
M ⊆ M′ .
Suppose I ′ ∈ M′ /M. From the definition of M′ , (I ′ , M′ , M) |=4 O.
For every rG ∈ PG of form (8) in Section 3, if M ̸|=4 KHj , for every
1 ≤ j ≤ n, then either M ̸|=4 KAj , for some n + 1 ≤ j ≤ m; or
M ̸|=4 notBj , for some m + 1 ≤ j ≤ k. Since M ⊆ M′ , M ̸|=4 KAj
implies (I ′ , M′ , M)̸|=4 KAj ; and (I, M, M) ̸|=4 notBj implies (I ′ , M′ , M)
̸|=4 notBj . Then (I ′ , M′ , M) |=4 rG . If M |=4 KHj , for some 1 ≤ j ≤ n,
then KHj ∈ Ph , from the definition of M′ , (I ′ , M′ , M) |=4 KHj . Therefore
(I ′ , M′ , M) |=4 rG . This contradicts the assumption that M is a paracon-
sistent MKNF model of KG . 

24
 Accordingly, every paraconsistent MKNF model is represented by a subset
of HA(KG ), and conversely each subset of HA(KG ) corresponds to a paracon-
sistent MKNF interpretation of KG . We will get the paraconsistent MKNF
models of KG by finding proper subsets of HA(KG ).

5.1. Positive Rules

A positive MKNF rule has the form

KH1 ∨ . . . ∨ KHn ← KA1 ∧ . . . ∧ KAm (18)

where Hi and Ai are literals occurring in KG .

To search for the appropriate Ph , we define a fixpoint operator.

Definition 18. Let KG = (O,P) be a ground positive para-MKNF knowledge

HA(K ) HA(K ) HA(K )
base and S ∈ 22 G . A mapping TKG : 22 G → 22 G is defined as
∪
TKG (S) = TKG (S),
S∈S
HA(KG )
where the mapping TKG : 2HA(KG ) → 22 is defined as follows.
• If OBO,S |=4 Ai , 1 ≤ i ≤ m for some ground integrity constraint ←
KA1 ∧ . . . ∧ KAm in PG , then TKG (S) = ∅.

• Otherwise, TKG (S) = {Q ⊆ HA(KG ) | Q = S ∪ R ∪ H}, where R = {KHi |

for each ground MKNF rule Cj of form (18) in PG such that OBO,S |=4
At for each 1 ≤ t ≤ m}, and H = {Kξ ∈ HA(KG ) | OBO,S |=4 ξ}.

Given a set S ⊆ HA(KG ), for each ground rule Cj in PG such that

OBO,S |=4 At for each 1 ≤ t ≤ m, we choose one rule head in every such
MKNF rule Cj , this collection is the set R. Since the MKNF rule is disjunc-
tive, there may be various choices when we choose the rule head, and hence
TKG (S) is the set collecting all the choices.
The definition of operator TKG differs from the fixpoint operator in ASP.
In fact, we substitute the set S for OBO,S , which means that the DL-part
and the rule part in the knowledge base KG can directly affect each other
when reasoning with KG . This is the original intension of combining DLs
HA(K )
and rules. Moreover, TKG is defined on 22 G , not on 2HA(KG ) , indirectly
leading to the fact that TKG is not monotonic. The following example gives
evidence of the non-monotonicity of TKG .

25
Example 8. Consider a para-MKNF knowledge base KG = (O, PG ), where
O = {p, p ⊃ f } and P = {K¬c ∨ K¬f ← Kp; Ke ← K¬f }.
Let S1 , S2 be subsets of 2HA(KG ) , such that S1 = ∅, and S2 = {{K¬f },
{K¬c}}. Clearly, S1 ⊂ S2 . However, we cannot conclude TKG (S1 ) ⊂ TKG (S2 ).
In fact, TKG (S1 ) = TKG (∅) = {{K¬f }, {K¬c}}. TKG (S2 ) = TKG ({K¬f })
∪ TKG ({K¬c}), in which TKG ({K¬f }) = {{K¬f, Ke}, {K¬c, K¬f, Ke}},
and TKG ({K¬c}) = {{K¬c}, {K¬f, K¬c}}. Then TKG (S2 ) = {{K¬c, K¬f,
Ke}, {K¬f, Ke}, {K¬c}, {K¬f, K¬c}}.

Since TKG is not monotonic, then the Knaster-Tarski theorem does not
apply. Therefore we employ a procedure that is also used in [38] to compute
the fixpoint of TKG .

TK G ↑ 0 = ∅
TKG ↑ n + 1 = TKG (TKG ↑ n)
∪ ∩
TKG ↑ ϖ = TKG ↑ n
α<ϖ α≤n<ϖ

where n is a successor ordinal and ϖ is a limit ordinal.

Lemma 3. TKG ↑ ϖ is a fixpoint.

Proof. Let I ∈ TKG ↑ ϖ, suppose I ̸∈ TKG (TKG ↑ ϖ), then there is no set
J ∈ TKG ↑ ϖ such that I ∈ TKG ({J}). That is to say, for a set J such that
I ∈ TKG ({J}), we have J ̸∈ TKG ↑ ϖ. Then for any α there exists α ≤ n < ϖ
such that J ̸∈ TKG ↑ n, which implies I ̸∈ TKG ↑ n + 1. This contradicts
the fact that I ∈ TKG ↑ ϖ. Conversely, let I ∈ TKG (TKG ↑ ϖ), then there
exists J ∈ TKG ↑ ϖ such that I ∈ TKG {J}. J ∈ TKG ↑ ϖ implies there exists
α < ϖ such that J ∈ TKG ↑ n for each α ≤ n < ϖ. Then I ∈ TKG ↑ n, for
each α + 1 ≤ n < ϖ. Therefore I ∈ TKG ↑ ϖ. 

Example 9. Continue to consider the knowledge base K from Example 5.

TKG ↑ 0 = ∅, TKG ↑ 1 = {{K¬F ly(T weety)}}, TKG ↑ 2 = TKG ↑ 1. Then
TKG ↑ 1 is a fixpoint of the operator TKG .

Lemma 4. Let KG = (O,PG ) be a para-MKNF knowledge base, and Ph be

the subset of HA(KG ) paraconsistently induced by a paraconsistent MKNF
interpretation M of KG . Then M |=4 KG if and only if Ph ∈ TKG ({Ph }).

26
Proof. (Sufficiency) If Ph ∈ TKG ({Ph }), then {KA1 , . . . , KAm } ⊆ Ph im-
plies that there exists a rule head KHi such that in KHi ∈ Ph . Since Ph
is the subset of HA(KG ) paraconsistently induced by M, we infer M |=4
KAj , 1 ≤ j ≤ m, then M |=4 KHi . Therefore, M |=4 KG .
(Necessity) If M |=4 KG , then M paraconsistently satisfies all the MKNF
rules in the program P. If M |=4 KAj , 1 ≤ j ≤ m, which means {KA1 , . . . ,
KAm } ⊆ Ph , then there exist i, 1 ≤ i ≤ n, M |=4 KHi , which means that
KHi ∈ Ph . Therefore, Ph ∈ TKG ({Ph }). 

Ph ∈ TKG ({Ph }) indicates that implication ”←” is closed w.r.t. Ph . That

is to say, if {KA1 , . . . KAm } ⊆ Ph , then there must exists a rule head KHi in
Ph , which is equivalent to M |=4 PG . Lemma 4 means that Ph corresponds
to the paraconsistent MKNF model of KG .
Let γ(TKG ↑ ϖ) = {S | S ∈ TKG ↑ ϖ, and S ∈ TKG ({S})}, and min(S) =
{S | there exists no Q ∈ S such that Q ⊂ S}.

Theorem 3. Let KG = (O,PG ) be a ground para-MKNF knowledge base and

M be a paraconsistent MKNF model of KG . Then M coincides with M′ =
{I | I |=4 OBO,Ph }, where Ph is an element of the set Q = min(γ(TKG ↑ ϖ)).

Proof. If Ph ∈ Q, then Ph ∈ TKG ({Ph }). From Lemma 4, M′ |=4 KG .

Suppose there exists an MKNF interpretation M′′ such that M′′ ! M′ and
(I, M′′ , M′ ) |=4 KG for each I ∈ M′′ . Let Ph′ be the subset of HA(KG ) that
is paraconsistently induced by M′′ , and M1 = {I | I |=4 OBO,Ph′ }. Then
(I, M1 , M′ ) |=4 KG for each I ∈ M1 . From Lemma 4, (I, M1 , M′ ) |=4 KG
implies Ph′ ∈ TKG ({Ph′ }). Then Ph′ ∈ γ(TKG ↑ ϖ), and M′ ! M implies
Ph′ Ph , which contradicts the minimality of Ph . Therefore, the claim holds.
If M is a paraconsistent MKNF model of KG , let Ph ⊆ HA(KG ) be the
subset of HA(KG ) that is paraconsistently induced by M. From Lemma 2,
M = {I | I |=4 OBO,Ph }, and from Lemma 4, Ph ∈ TKG ({Ph }). Suppose Ph
is not in Q, then there exists Ph′ , such that Ph′ Ph . Let M′ = {I | I |=4
OBO,Ph′ }, M′ ) M. From the definition of M′ , M′ |=4 O. For a rule rG , if
OBO,Ph′ |=4 Ai , 1 ≤ i ≤ m, then there exists a rule head KHj ∈ Ph′ , 1 ≤ j ≤ n,
thus M′ |=4 rG and then M′ |=4 KG , which contradicts the fact that M is a
paraconsistent MKNF model of KG . 

When PG is nondisjunctive, in which MKNF rules are of the following

form
KH ← KA1 ∧ . . . ∧ KAm (19)

27
where Hi and Ai are literals occurring in KG , the rule head in each rule in PG
has at most one element. If OBO,Ph paraconsistently satisfies the rule body,
then there is at most one candidate to be added to the set R. In this case,
the set Q has only one element.

Proposition 2. Let KG = (O,PG ) be a ground para-MKNF knowledge base,

where PG is a positive nondisjunctive program. Then KG has a unique para-
consistent MKNF model.

In fact, the fixpoint operator can be simplified when PG is nondisjunctive.

Definition 19. Let KG = (O,PG ) be a ground nondisjunctive positive para-

MKNF knowledge base. The fixpoint operator TKG : 2HA(KG ) → 2HA(KG ) is
defined as follows:

TKG (S) = {KH | f or each rG ∈ PG of the f orm (19) such that

OBO,S |=4 At f or each 1 ≤ t ≤ m} ∪
{Kξ ∈ HA(KG ) | OBO,S |=4 ξ}

It can be easily verified that TKG is monotonic. Then it follows from the
Knaster-Tarski’s theorem that TKG has a unique least fixpoint T∞ ∞
KG . TKG can
be computed as in ASP by setting S0 = ∅ and Si =TKG (Si−1 ). If PG is a
finite set, Sn = Sn+1 = . . . = T∞
KG .

Proposition 3. For a ground nondisjunctive positive para-MKNF knowl-

edge base KG , TKG ↑ ϖ = {T∞
KG }. That is to say, the paraconsistent MKNF
model of KG is M = {I | I |=4 OBO,T∞K
}.
G

The constructing idea of the operator TKG is the same as the one of the
operator TKG . Then Proposition 3 follows.

5.2. General Rules

For general rules, we cannot apply the fixpoint operator directly to the
rules, since reasoning with the modal not operator is nonmonotonic. There-
fore, we transform each general program to a positive program, and then
compute the fixpoint of the transformed program.

28
Definition 20. Let KG = (O,PG ) be a ground para-MKNF knowledge base.
∗
Then its transformation is defined as KG obtained by replacing each general
rule in PG with the following positive MKNF rule.

Kµ1 ∨ . . . ∨ Kµn ∨ KBm+1 ∨ . . . ∨ KBk ← KAn+1 ∧ . . . ∧ KAm , (20)

KHi ← Kµi f or 1 ≤ i ≤ n, (21)

← Kµi ∧ KBj f or 1 ≤ i ≤ n, m + 1 ≤ j ≤ k, (22)
Kµi ← KHi ∧ Kµj f or 1 ≤ i, j ≤ n. (23)

This transformation, following Chiaki Sakama and Katsumi Inoue [38],

is originally introduced in [19] in a different form, in order to compute
answer sets for any class of function-free logic programs, including the ex-
tended disjunctive program of form (9).
The modal operator K can be understood as ”believe”. The intuition of
the transformed clauses is that if An+1 , . . . , Am are believed to be true, then
there exists a literal Hi (1 ≤ i ≤ n), such that Hi is believed to be true via
µi when Bm+1 , . . . , Bk are not believed to be true; otherwise, there exists a
literal Bj (m + 1 ≤ j ≤ k), such that Bj is believed to be true.
∗
Given a set S∗ that is a subset of 2HA(KG ) , Φ(S∗ ) = {S∗ ∩ KA(KG ) | S∗ ∈
S∗ }.

Theorem 4. Let KG = (O,PG ) be a ground para-MKNF knowledge base,

then each paraconsistent MKNF model of KG equals M = {I | I |=4 OBO,Ph },
where Ph is an element of the set Q = Φ(min(γ(TKG∗ ↑ ϖ))).

Proof. First of all, we prove every element in Q corresponds to a para-

consistent MKNF model of KG . Suppose that Ph ∈ Q, then there ex-
ists Ph∗ ∈ min(γ(TKG∗ ↑ ϖ)), such that Ph = Ph∗ ∩ KA(KG ). Let M∗ =
{I | I |=4 OBO,Ph∗ }. From Theorem 3, M∗ is a paraconsistent MKNF model
∗
of KG . M |=4 O by definition of M. For each MKNF rule rG of form (8) in
PG , if M |=4 KAi for n + 1 ≤ i ≤ m and M |=4 notBj , for m + 1 ≤ j ≤ k,
then from Ph ⊆ Ph∗ , we have M∗ |=4 Ai for n + 1 ≤ i ≤ m and KBj ̸∈ Ph
for m + 1 ≤ j ≤ k. KBj ∈ KA(KG ) implies KBj ̸∈ Ph∗ . Since M∗ para-
consistently satisfies each rule in PG∗ , then there exists µl , 1 ≤ l ≤ n, such
that M∗ |=4 Kµl , and by rule (21) M∗ |=4 KHl . KHl ∈ KA(KG ) implies
KHl ∈ Ph . Therefore, M |=4 KHl , for some 1 ≤ l ≤ n. Then M |=4 KG .

29
 Suppose there exists a paraconsistent MKNF interpretation M′′ ) M
such that (I ′′ , M′′ , M) |=4 KG . Let Ph′ be the subset of HA(KG ) paraconsis-
tently induced by M′′ , and M′ = {I | I |=4 OBO,Ph′ }. Then we can infer that
(I ′ , M′ , M) |=4 KG . Clearly, Ph′ ( Ph . For simplicity let Ph \Ph′ = {Kξ},
′ ′
Ph∗ = Ph∗ \ {Kξ, Kµi }, in which Kξ ← Kµi , and Ph′ = Ph∗ ∩ KA(KG ). Let
M′∗ = {I | I |=4 OBO,P ′ ∗ }. Obviously, (I, M′∗ , M∗ ) |=4 O. For a rule rG ∗
of
h
′∗ ∗
form (20), it corresponds to a rule rG of form (8) in PG . If (I, M , M ) |=4
′
KAi for each n + 1 ≤ i ≤ m, then KAi ∈ Ph∗ . KAi ∈ KA(KG ) implies
KAi ∈ Ph′ . Then (I ′ , M′ , M) |=4 KAi . From (I ′ , M′ , M) |=4 KG , we
know (I ′ , M′ , M) |=4 rG . In case of (i), if (I ′ , M′ , M) |=4 notBj , for
each m + 1 ≤ j ≤ k, which implies KBj ̸∈ Ph′ and M |=4 notBj , for each
m+1 ≤ j ≤ k, then there exist Hl , 1 ≤ l ≤ n such that (I ′ , M′ , M) |=4 KHl ,
′
which means KHl ∈ Ph′ ⊆ Ph∗ . On the other hand, Ph′ ⊆ Ph ⊆ Ph∗ implies
KAi ∈ Ph ⊆ Ph∗ for each n + 1 ≤ i ≤ m, since M∗ is a paraconsistent
MKNF model of KG , there exists a µv , 1 ≤ v ≤ n, such that M∗ |=4 Kµv ,
since there exists KHl such that KHl ∈ Ph′ ⊆ Ph ⊆ Ph∗ , then from rule (23),
′
Kµl ∈ Ph∗ . KHl ∈ Ph′ means KHl is not Kξ, then Kµl ∈ Ph∗ . Therefore
(I, M′∗ , M∗ ) |=4 rG ∗
. In case of (ii), if (I ′ , M′ , M) ̸|=4 notBj , for some j,
′
m + 1 ≤ j ≤ k, which implies KBj ∈ Ph′ ⊆ Ph∗ . Then (I, M′∗ , M∗ ) |=4 rG ∗
.
∗ ∗
For other rules rG of form (21), (22), and (23) in PG , it can be easily inferred
′
(I, M′∗ , M∗ ) |=4 rG ∗
by the fact that Ph∗ ( Ph∗ . Then it contradicts that M∗
∗
is a paraconsistent MKNF model of KG . Therefore every element in Q is a
paraconsistent MKNF model of KG .
Conversely, suppose M1 is the paraconsistent MKNF model of KG . From
Lemma 2 we imply that M1 = {I | I |=4 OBO,Ph }, where Ph is the subset
of HA(KG ) paraconsistently induced by M1 . For each MKNF rule, let Sµ =
∪
r {Kµl | M1 |=4 Ai , f or each n + 1 ≤ i ≤ m, M1 ̸|=4 Bj , f or each m + 1 ≤
j ≤ k, and M1 |=4 Hl }. Let S = Ph ∪Sµ . Then, S satisfies all the MKNF rules
in PG∗ and S ∈ γ(TKG∗ ↑ ϖ) by the construction. Now we define S∗ = Ph ∪ S′ ,
in which S′ is the minimal subset of Sµ such that each Kµl is chosen in a
way that satisfies the MKNF rules in PG∗ . Now we need to prove S∗ to be
minimal in γ(TKG∗ ↑ ϖ). Assume that S∗1 ∈ γ(TKG∗ ↑ ϖ) such that S∗1 ⊂ S∗ .
Since S∗ is defined to be the minimal set with respect to elements in Sµ ,
then S∗1 ∩ KA(KG ) ⊂ S∗ ∩ KA(KG ), and thus there exists Kξ ∈ S∗ \ S∗1 , and
Kξ ∈ Ph . Let Ph′ = Ph \ Kξ = S∗1 ∩ KA(KG ) and M′1 = {I | I |=4 OBO,Ph′ }.
Then from S∗1 ∈ γ(TKG∗ ↑ ϖ), we can easily infer that (I, M′1 , M1 ) |=4 KG ,
which contradicts the fact that M1 is the paraconsistent MKNF model of

30
KG . Then S∗ ∈ min(γ(TKG∗ ↑ ϖ)). That is to say, for each paraconsistent
MKNF model M of KG , there exists an element S in Q such that M =
{I | I |=4 OBO,S }. 

Example 10. Consider a ground knowledge base KG = (O,PG ), where O =

{p, p ⊃ b, b ⊃ f }, PG = {K¬f ∨ K¬c ← Kp; Ka ← K¬c; Kf ← Kb ∧ notp},
and p, b, f, c, a are atoms. By Definition 20, PG is transformed to PG∗ :

K¬f ∨ K¬c ← Kp

Ka ← K¬c
Kµ ∨ Kp ← Kb
← Kµ ∧ Kp
Kf ← Kµ
We compute the fixpoint by applying the procedure presented in section 5.1 to
∗
the knowledge base KG = (O,PG∗ ). By evaluating TKG ↑ n + 1 = TKG TKG ↑ n
recursively, min(TKG∗ ↑ n) = {{K¬c, Ka, Kp, Kf }, {K¬f, Kp, Kf }}. Since
the computation process is tedious, we omit it here. Also, it can be easily veri-
fied that M1 = {I | I |=4 {p, b, f, ¬c, a}}, and M2 = {I | I |=4 {p, b, f, ¬f }}
are two paraconsistent MKNF models of KG . Note that M1 is the preferred
MKNF model of KG , and KG is a consistent knowledge base.

5.3. Stratified Rules

In this section, we discuss the class of stratified program, which are
nondisjunctive, but can contain not-atoms. Stratified rules are of form (24),
and they can be separated into strata, each of which can be evaluated sepa-
rately.
KH ← KA1 ∧ . . . ∧ KAm ∧ notBm+1 ∧ . . . ∧ notBk (24)

Definition 21 ([33]). Let KG = (O, PG ) be a ground nondisjunctive MKNF

knowledge base, and let σ: KA(KG ) → N+ be a function assigning an integer
to each K-atom. For a modal atom Kξ, the sets of K-atoms [Kξ]↑ and [Kξ]↓
are defined as follows:

[Kξ]↑ = {Kφ | Kφ ∈ HA(KG ) such that σ(Kφ) ≤ σ(Kξ)}

[Kξ]↓ = HA(KG )\[Kξ]↑

31
 We say that σ is a stratification of PG if the following conditions hold:
(1) For each rule rG ∈ PG of the form (24), σ(KH) ≥ σ(KBi ) for each
1 ≤ i ≤ m, and σ(KH) > σ(KBi ) for each m + 1 ≤ i ≤ k;
(2) For each atom Kξ ∈ KA(KG ) and each subset Sh ⊆ [Kξ]↓ , if
OBO,Sh ̸|=4 ξ, then OBO,Sh ∪[Kξ]↑ ̸|=4 ξ.
The program PG is stratified if a stratification σ exists. A stratification
σ partitions PG into strata PG1 , . . ., PGζ as follows:

PGi = {rG ∈ PG | σ(KH) = i, where KH is the head atom of rG }

This sequence is often identified with σ and is called a stratification.

A nonground MKNF knowledge base K = (O, P) is stratified if KG =
(O, PG ) is stratified.
Let P and N be disjoint subsets of KA(KG ). The program PG [not, P, N ]
is obtained from PG by replacing each notξ with ⊤ if Kξ ∈ N and K¬ξ ∈
N ; with t if Kξ ∈ N and K¬ξ ̸∈ N ; with f if Kξ ∈ P and K¬ξ ̸∈ N ; with
⊥ if Kξ ∈ P and K¬ξ ∈ P .
A paraconsistent MKNF model of a stratified MKNF knowledge base can
be computed by processing strata sequentially.

Definition 22. Let KG = (O, PG ) be a ground stratified para-MKNF knowl-

edge base and PG1 , . . ., PGζ a stratification of PG . The sequence of subsets
U0 , . . . , Uζ of HA(KG ) is inductively defined as follows:

U0 = ∅; (25)

Pi = {Kξ ∈ HA(KG ) | σ(ξ) < i and OBO,Ui−1 |=4 ξ}; (26)

Ni = {Kξ ∈ HA(KG ) | σ(ξ) < i and OBO,Ui−1 ̸|=4 ξ}; (27)

χi = {Kξ ← | Kξ ∈ Ui−1 } ∪ PGi [not, Pi , Ni ]; (28)

Ui = T∞
Gi , (29)

where Gi = (O, χi ), and 1 ≤ i ≤ ζ for equations (26) - (29). We define

UK∞G = Uζ .

32
Theorem 5. Let KG = (O, PG ) be a stratified ground para-MKNF knowl-
edge base, UK∞G be a set obtained from Definition 22 using any stratification,
and M= {I | I |=4 OBO,UK∞ }.
G

- If M ̸= ∅, then M is a paraconsistent MKNF model of KG .

- If KG has a paraconsistent MKNF model, then this model is equal to

M.

Proof. Let PG1 , . . ., PGζ be a stratification of PG used to compute UK∞G .

∪
Moreover, let PG≤i = j≤i PGj , let KG ≤i
= (O, PG≤i ), and Mi = {I | I |=4
OBO,Ui }, for 0 ≤ i ≤ ζ. Each Gi is a positive nondisjunctive program, so by
Proposition 2, it has at most a paraconsistent MKNF model corresponding
to Mi by Proposition 3.
Since each χi contains MKNF rules Kξ ← such that Kξ ∈ Ui−1 , then
Ui−1 ⊆ Ui . For each Kξ ∈ KA(KG ), σ(Kξ) = i, and each j > i, we have Mi
|=4 Kξ if and only if Mj |=4 Kξ. We denote this property by (†). In fact, if
OBO,Ui |=4 ξ, then OBO,Uj |=4 ξ, since Ui ⊆ Uj . If OBO,Uj |=4 ξ, but OBO,Ui
̸|=4 ξ. Since Uj \Ui ⊆ [Kξ]↑ , we have OBO,Ui ∪[Kξ]↑ |=4 ξ, which contradicts
Condition (2) of the definition of stratification. Therefore, OBO,Ui |=4 ξ if
and only if OBO,Uj |=4 ξ, which is equivalent to property (†).
We now prove by induction on 1 ≤ i ≤ ζ that both claims of this theorem
≤i
hold for KG and Mi . The base case i = 0 holds obviously, so we consider
the induction step.
(Claim 1) Assume that Mi ̸= ∅, then Mi−1 ̸= ∅ as well. By induction
≤i−1
assumption Mi−1 is a paraconsistent MKNF model of KG . Consider each
MKNF rule rG ∈ PG≤i−1 , for each rule head KH and each rule body KAt
(1 ≤ t ≤ m) and notBj (m + 1 ≤ j ≤ k), we have σ(KH) < i, σ(KAt ) < i
(1 ≤ t ≤ m) and σ(KBj ) < i (m + 1 ≤ j ≤ k). Then we have Mi |=4 Kξ if
and only if Mi−1 |=4 Kξ by property (†) , where ξ ∈ {H, At , Bj } (1 ≤ t ≤ m
and m + 1 ≤ j ≤ k). Since Mi−1 |=4 rG , we have Mi |=4 rG . Consider each
MKNF rule rG ∈ PGi , for each negative rule body notBj (m + 1 ≤ j ≤ k),
σ(KBj ) < i (m + 1 ≤ j ≤ k). By property (†), we have Mi |=4 rG if and
only if Mi |=4 rG [not, Pi , Ni ]. Thus we conclude Mi |=4 PGi , then Mi |=4
≤i
KG .
Assume that there exists a paraconsistent MKNF interpretation M′i ⊃
≤i
Mi such that (Ii′ , M′i , Mi ) |=4 KG for each Ii′ ∈ M′i . For each Kξ ∈
KA(KG ), such that σ(Kξ) < i, we have M′i |=4 Kξ if and only if Mi−1 |=4 Kξ.

33
 Rules DL= ϕ DL∈ PTime DL∈coNP
Non-disjunctive positive PTime PTime coNP
Stratified PTime PTime △p2
General coNP coNP Πp2
Disjunctive positive coNP/Πp2 coNP/Πp2 coNP/Πp2
Disjunctive general Πp2 Πp2 Πp2

Table 3: Data Complexity of Instance Checking in Para-MKNF KBs. DL indicates the

base DL.

≤i−1
Otherwise, Mi−1 is not the paraconsistent MKNF model of KG . But then
(Ii′ , M′i , Mi ) |=4 χi , which contradicts the fact that Mi is the paraconsistent
≤i
MKNF model of Gi . Therefore, Mi is a paraconsistent MKNF model of KG .
≤i
(Claim 2) Assume that KG has a paraconsistent MKNF model M′i . For
each Kξ ∈ KA(KG ), such that σ(Kξ) < i, we have M′i |=4 Kξ if and only if
Mi−1 |=4 Kξ. Otherwise, Mi−1 is not the paraconsistent MKNF model of
≤i−1
KG . But then M′i is an MKNF model of Gi . Therefore, Mi = M′i .4

As discussed above, an algorithm for computing paraconsistent MKNF

models follows from the fixpoint operators. From Theorem 2 in Section 4, we
obtain that the data complexity in our paradigm is not higher than that for
the classical knowledge base. Therefore, we obtain Table 3 from [33], which
presents the data complexity of instance checking in different para-MKNF
knowledge bases.

6. Related Work
A well-founded semantics [21] for hybrid MKNF has been proposed for
better efficiency of reasoning. It extends two-valued semantics to three valued
semantics, and is able to detect inconsistencies occurring in the knowledge
base, however, cannot handle inconsistencies as such. Since [21] is restricted
to nondisjunctive rules, it is less expressive than the four-valued semantics.
The advantage of [21] is that when the considered DL is of polynomial data
complexity, the combined approach remains polynomial.

4
This proof is only a slight variation of the proof of Theorem 4.19 in [33], which we
include for completeness of our treatment.

34
 There are alternatives to four-valued paraconsistent semantics for DLs,
due to the inherent limitation of four-valued logic. In fact, quasi-classical se-
mantics [41], based on quasi-classical logic [18], is a paraconsistent semantics
for ALC to handle inconsistencies. A weak semantics was proposed, which
was actually identical to ALC4. The problem that Modus Ponens, Modus
Tollens, and Disjunctive Syllogisms fail, which is inherent in ALC4, also oc-
curs in the weak semantics. A strong semantics which was built upon the
weak semantics does not suffer from these problems.
Paradoxical description logic ALC LP [42] is an extension of ALC with
semantics of logics of paradox [36]. Paradoxical entailment satisfies the ex-
cluded middle rule and intuitive equivalence that are not valid in four-valued
logics. Moreover, ALC LP has strong inferential power than ALC4 on asser-
tions and material inclusions.
Nevertheless, the distinct advantage of four-valued paraconsistent seman-
tics is allowing classical reasoners to derive sound but non-trivial conclusions
from even inconsistent knowledge bases by embedding them into the classical
framework.
Chiaki Sakama and Katsumi Inoue [38] proposed a paraconsistent stable
semantics for extended disjunctive programs. They introduced a fixpoint
operator for disjunctive programs, which is the inspiration of our work on
the fixpoint operator. Moreover, by substituting a nine-valued lattice for
FOUR as truth value set, they proposed semi-stable models, which is also
used in [9] to cope with instability. The program P = {a ← nota} has
no paraconsistent stable model, but has a semi-stable model. This method
deserves further discussion in our work, in which there is no paraconsistent
MKNF model for the MKNF rule Ka ← nota.

7. Conclusion
In this paper we have presented a paraconsistent semantics of hybrid
MKNF knowledge bases that is sound w.r.t. the classical two-valued se-
mantics defined in [33], which restricts to the paraconsistent semantics of
extended disjunctive program [38] and to the paraconsistent semantics of
OWL [28], when the DL-part and LP-part is empty, respectively. Further-
more, we characterized paraconsistent MKNF models via fixpoint operators,
and showed that the complexity of our paradigm is not higher than that in
[33].

35
 There are a number of paths to further develop this work. First of all, in
[21, 20], a well-founded semantics was introduced for hybrid MKNF knowl-
edge bases which has better complexity properties, and our paraconsistent
approach could be carried over to this paradigm. Moreover, inconsistency is
not the only problem that occurs in the real world, some other problems such
as vagueness and probabilistic uncertainty, which cannot be coped with by
classical reasoners either, may deserve some discussions and research. Then
it is necessary to extend fuzzy semantics and probabilistic semantics to hy-
brid MKNF knowledge bases. Finally, an even tighter paraconsistent and
non-monotonic integration of OWL and rules could furthermore be investi-
gated. In [4, 22, 23, 24], nominal schemas are introduced as an extension to
DL-based ontology languages, which provide sufficient expressivity to incor-
porate rule-based modeling into ontologies. Inconsistency handling is also an
interesting problem which deserves to be discussed in the context of nominal
schemas.

Acknowledgments
The authors thank Matthias Knorr and the anonymous referees for their
valuable comments, which improved the quality of this paper. The authors
acknowledge the support provided by the National Basic Research Program
of China [2010CB334706, 2011CB311808], and the National Natural Science
Foundation of China [11071061]. The last named author acknowledges sup-
port by the National Science Foundation under award 1017225 “III: Small:
TROn – Tractable Reasoning with Ontologies.”

References
[1] Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Patel-Schneider,
P. F. (Eds.): The Description Logic Handbook: Theory, Implementation
and Applications. Cambridge University Press, Second Edition (2007)

[2] Belnap, N. D.: A Useful Four-valued Logic. Modern uses of multiple-

valued logics. 7–73 (1977)

[3] Berners-Lee, T., Hendler, J., Lassila, O.: The Semantic Web. Scientific
American. 284 (5), 35–43 (2001)

36
[4] Carral Martinez, D., Hitzler, P.: Extending Description Logic Rules. In:
Simperl, E. et al. (eds.) Proceedings of the 9th Extended Semantic Web
Conference (ESWC2012). LNCS, vol. 7295, pp. 345–359. Springer (2012)
[5] Carral Martinez, D., Krisnadhi, A. A., Hitzler, P.: Integrating OWL
and rules: A syntax proposal for nominal schemas. In Proceedings of
Proceedings of OWL: Experiences and Directions Workshop, Heraklion,
Crete, Greece, May 27-28 (2012)
[6] Donini, F. M., Lenzerini, M., Nardi, D., Schaerf, A.: AL-log: Integrat-
ing Datalog and Description Logics. Journal of Intelligent Information
Systems (JIIS). 10(3), 227–252 (1998)
[7] Eiter, T., Lukasiewicz, T., Schindlauer, R., Tompits, H.: Combining An-
swer Set Programming with Description Logics for the Semantic Web. In
Proceedings of the 9th International Conference on Principles of Knowl-
edge Representation and Reasoning (KR’04), pp. 141–151. AAAI Press
(2004)
[8] Eiter, T., Ianni, G., Schindlauer, R., Tompits, H.: Effective Integration
of Declarative Rules with External Evaluations for Semantic Web Rea-
soning. In Proceedings of the 3rd European Semantic Web Conference
(ESWC2006). LNCS, vol. 4011, pp. 273–287. Springer (2006)
[9] Eiter, T., Fink, M., Moura, J.: Paracoherent Answer Set Programming.
In Proceedings of the 12th International Conference on Principles of
Knowledge Representation and Reasoning (KR’10), pp. 486–496. AAAI
Press (2010)
[10] Fitting, M.: First-Order Logic and Automated Theorem Proving, 2nd
Edition. Texts in Computer Science. Springer (1996)
[11] Gelfond, M., Lifschitz, V.: The Stable Model Semantics for Logic Pro-
gramming. In Proceedings of the 5th International Conference and Sym-
posium on Logic Programming (ICLP/SLP’88), pp. 1070–1080. MIT
Press (1988)
[12] Grosof, B. N., Horrocks, I., Volz, R., Decker, S.: Description Logic Pro-
grams: Combining Logic Programs with Description Logic. In Proceed-
ings of the 12th International World Wide Web Conference (WWW2003),
pp. 48–57. ACM (2003)

37
[13] Haase, P., van Harmelen, F., Huang, Z., Stuckenschmidt, H., Sure, Y.:
A Framework for Handling Inconsistency in Changing Ontologies. In:
Gil, Y., Motta, E., Benjamins, V.R., Musen, M.A.(Eds.) Proceedings of
the 4th International Semantic Web Conference (ISWC2005). LNCS, vol.
3729, pp. 353–367. Springer (2005)
[14] Hitzler, P., Krötzsch, M., Parsia, B., Patel-Schneider, P.F., Rudolph, S.
(Eds.): OWL 2 Web Ontology Language: Primer. W3C Recommendation
27 October 2009 (2009), available from http://www.w3.org/TR/owl2-
primer/
[15] Hitzler, P., Krötzsch, M., Rudolph, S.: Foundations of Semantic Web
Technologies. Chapman & Hall/CRC (2009)
[16] Horrocks, I., Patel-Schneider, P. F., Boley, H., Tabet, S., Grosof, B.,
Dean, M.: SWRL: A Semantic Web Rule Language Combining OWL
and RuleML. W3C Member Submission 21 May 2004, available from
http://www.w3.org/Submission/SWRL/
[17] Huang, S., Li, Q., Hitzler, P.: Paraconsistent Semantics for Hybrid
MKNF Knowledge Bases. In: Gutierrez, C., Rudolph, S. (Eds.) Pro-
ceedings of the 5th International Conference on Web Reasoning and Rule
Systems (RR’11). LNCS, vol. 6902, pp. 93–107. Springer (2011).
[18] Hunter, A.: Reasoning with Contradictory Information Using Quasi-
classical Logic. Journal of Logic Computation. 10 (5), 677–703 (2000)
[19] Inoue, K., Koshimura, M., Hasegawa, R.: Embedding Negation as Fail-
ure into a Model Generation Theorem Prover. In Proceedings of the 11th
International Conference on Automated Deduction. LNCS, vol. 607, pp.
400–415. Springer (1992)
[20] Knorr, M., Alferes, J. J., Hitzler, P.: A Coherent Well-founded Model
for Hybrid MKNF Knowledge Bases. In: Ghallab, M., Spyropoulos, C.
D., Fakotakis, N., and Avouris, N. M.(Eds.) Proceedings of the 18th
European Conference on Artificial Intelligence (ECAI2008). Vol. 178, pp.
99–103. IOS Press, Patras, Greece (2008)
[21] Knorr, M., Alferes, J., Hitzler, P.: Local Closed-world Reasoning with
Description Logics under the Well-founded Semantics. Artificial Intelli-
gence 175(9–10), 1528–1554 (2011)

38
[22] Knorr, M., Hitzler, P., Maier, F.: Reconciling OWL and Non-monotonic
Rules for the Semantic Web. In Proceedings of the 20th European Confer-
ence on Artificial Intelligence (ECAI2012), pp. 474–479. IOS Press (2012)

[23] Krisnadhi, A., Maier, F., Hitzler, P.: OWL and Rules. In: Polleres,
Al, d’Amato, C., Arenas, M., Handschuh, S., Kroner, P., Ossowski, S.,
Patel-Schneider, P.F. (Eds.) Reasoning Web. Semantic Technologies for
the Web of Data. 7th International Summer School 2011, Galway, Ireland,
August 23-27, 2011, Tutorial Lectures. LNCS, vol. 6848, pp. 382-415.
Springer (2011)

[24] Krötzsch, M., Maier, F., Krisnadhi, A. A., Hitzler, P.: A better un-
cle for OWL: Nominal schemas for integrating rules and ontologies. In:
Sadagopan, S., Ramamritham, K., Kumar, A., Ravindra, M., Bertino,
E., Kumar, R. (eds.) Proceedings of the 20th International World Wide
Web Conference (WWW2011), Hyderabad, India, March/April 2011, pp.
645–654. ACM, New York (2011)

[25] Krötzsch, M., Rudolph, S., Hitzler, P.: ELP: Tractable Rules for OWL 2.
In: Sheth, A., Staab, S., Dean, M., Paolucci, M., Maynard, D., Finin, T.,
Thirunarayan, K. (Eds.) Proceedings of the 7th International Semantic
Web Conference (ISWC2008). LNCS, vol. 5318, pp. 649–664. Springer
(2008)

[26] Levy, A. Y., Rousset, M.-C.: Combining Horn Rules and Description
Logics in CARIN. Artifician Intelligence. 104(1-2), 165–209 (1998)

[27] Lifschitz, V.: Nonmonotonic Databases and Epistemic Queries. In: My-
lopoulos, J. and Reiter, R. (Eds.) Proceedings of the 12th International
Joint Conference on Artificial Intelligence (IJCAI-91), pp. 381–386. Mor-
gan Kaufmann Publishers (1991)

[28] Ma, Y., Hitzler, P., Lin, Z.: Algorithms for Paraconsistent Reasoning
with OWL. In: Proceedings of the 4th European Semantic Web Confer-
ence (ESWC2007). LNCS, vol. 4519, pp. 399–413. Springer (2007)

[29] Maier, F.: Extending Paraconsistent SHOIQ. In: Hitzler, P., and
Lukasiewicz, T. (Eds.) Proceedings of the 4th International Conference
on Web Reasoning and Rule Systems (RR’09). LNCS, vol. 6333, pp. 118–
132. Springer (2010)

39
[30] Maier, F., Ma, Y., Hitzler, P.: Paraconsistent OWL and Related Logics.
Semantic Web journal. DOI : 10.3233/SW-2012-0066, 2012.

[31] Motik, B., Cuenca Grau, B., Horrocks I., Wu, Z., Fokoue, A., Lutz,
C., Hitzler, P., Krötzsch, M., Parsia, B., Patel-Schneider, P.F., Rudolph,
S. (Eds.): OWL 2 Web Ontology Language: Profiles. W3C Recommen-
dation 27 October 2009 , available from http://www.w3.org/TR/owl2-
profiles/

[32] Motik, B., Rosati, R.: A Faithful Integration of Description Logic with
Logic Programming. In Proceedings of the 20th International Joint Con-
ference on Artificial Intellegence (IJCAI-07), Hyderabad, India, Janaray
6-12, 2007, pp. 477–482. AAAI Press (2007)

[33] Motik, B., Rosati, R.: Reconciling Description Logics and Rules. Journal
of the ACM, 57(5), 1–61 (2010)

[34] Motik, B., Sattler, U., Studer, R.: Query Answering for OWL-DL with
rules. Journal of Web Semantics. 3(1), 41–60 (2005)

[35] Patel-Schneider, P. F.: A Four-valued Semantics for Terminological Log-

ics. Artificial Intelligence. 38, 319–351 (1989)

[36] Priest, G.: Logic of Paradox. Journal of Philosophical Logic. 8, 219–241

(1979)

[37] Rosati, R.: DL+log: Tight Integration of Description Logics and Dis-
junctive Datalog. In Proceedings of the 10th International Conference
on Principles of Knowledge Representation and Reasoning (KR’06), pp.
68–78. AAAI Press (2006)

[38] Sakama, C., Inoue, K.: Paraconsistent Stable Semantics for extended
disjunctive programs. Journal of Logic and Computation. 5, 265–285.
Oxford University Press (1995)

[39] Schlobach, S., Cornet, R.: Non-standard Reasoning Services for the De-
bugging of Description Logic Terminologies. In: Gottlob, G., Walsh, T.
(Eds.) Proceedings of the 18th International Conference on Artificial In-
telligence (IJCAI-03), pp. 355–362. Morgan Kaufmann Publishers (2003)

40
[40] Van Gelder, A., Ross, K., Schlipf, J. S.: The Well-Founded Semantics
for General Logic Programs. Journal of the ACM. 38 (3), 620–650 (1991)

[41] Zhang, X., Lin, Z.: Paraconsistent Reasoning with Quasi-classical Se-
mantic in ALC. In: Calvanese, D., Lausen, G.(eds.) Proceedings of 2th
International Conference on Web Reasoning and Rule Systems (RR 2008).
LNCS, vol. 5341, pp. 222–229. Springer (2008)

[42] Zhang, X., Lin, Z., Wang, K.: Towards a Paradoxical Description Logic
for the Semantic Web. In: Link, S., Prade, H. (eds.): Proceedings of the
6th International Symposium on Foundations of Information and Knowl-
edge Systems (FoIKS2010). LNCS, vol. 5956, pp. 306–326 (2010)

41