Lexicon, Ontology and Text Meaning

Boyan A. Onyshkevych
U.S. Department of Defense and
Computational Linguistics Program,
Carnegie Mellon University
bao@a, hi. es. cmu. edu
Sergei Nirenburg
Center for Machine Translation
School of Computer Science
Carnegie Mellon University
sergei @nl. cs. cmu. edu

Abstract
A computationally relevant theory of lexical semantics must take into consideration
both the form and the content of three different static knowledge sources -- the
lexicon, the ontological domain model and a text meaning representation language.
Meanings of lexical units axe interpreted in terms of their mappings into the ontology
and/or their contributions to the text meaning representation. We briefly describe
one such theory. Its precepts have been used in mONYSUS, a machine translation
system prototype.

1 Introduction
A comprehensive natural language processing system (NLP) must be able to analyze
natural language texts, and, having understood their content, generate other natural
language texts, depending on the nature of the underlying reasoning system. Thus, in a
machine translation system this "other" NL text will be a text in a different language but
with the same meaning; in a NL front end to a database, it will be the response to a query.
In order to produce an appropriate NL output, the NLP system must "understand" the
input text. In some systems this understanding is modeled as an acceptable match for a
pre-existing structural or literal pattern. In knowledge-based systems it is customary to
model understanding through representing the results of input text analysis as a set of
well-formed structures in a specially-designed formal artificial language. Such languages
come in a number of varieties and we are not arguing here about which ones are more
appropriate or promising. The crucial point is that in order to have any explanatory
power, the atoms of this meaning representation language must be interpreted in terms
of an independently motivated model of the world. Moreover, if any realistic experiments
have to be performed with such an NLP system, this world model (sometimes called
ontology) must be actually built, not only defined algebraically. The process of NL analysis
is then interpreted as putting lexical, syntactic, and prosodic units of the source text
in correspondence with elements of the text meaning representation language. During
generation, the representation language units serve as input and the target language units,
as output. (This view of language processing is, of course, highly schematic - - as will be
seen from the more detailed discussion which follows; but this is exactly the level of detail
needed at this point.)

238
 In our understanding, the purpose of the lexicon is to facilitate the above linking.
Of course, the lexicon contains not only mappings into the ontology. It also contains
information about morphological, syntactic, pragmatic, and collocational properties of
the lexical unit - - since all and any of these components serve as clues at various stages
of the process of mapping text into representation or vice versa. But the semantics of
most of the open-class lexical items is described in terms of their mappings into instances
of ontological concepts.
A generic view of the data flow in a system of the above class is given in Figure 1. The
experiences described in this paper relate to our work on DIONYSUS, a machine translation
system. The interaction between the ontology, the text meaning representation, and the
lexicon in this system was the central point of our study. This interaction takes place in
two spheres - - both during actual processing of text and during knowledge acquisition
phase. In what follows, we first briefly describe these three knowledge sources and their
interactions, as they are implemented in DIONYSUS. Next, we discuss some issues of
distinguishing lexical and non-lexical knowledge and their co-existence in a comprehensive
NLP application.
Justifying particular representation choices, details of coverage of specific phenom-
ena and comparisons with other semantic and pragmatic NLP processors are beyond the
scope of this paper. See Nirenburg and Defrise, forthcoming, a and b; Carlson and Niren-
burg, 1990; Meyer et al., 1990; Nirenburg, 1989; and Nirenburg and Goodman, 1990 for
additional details.
Each of the three knowledge sources have both content and form - - that is, a special
knowledge representation language has been developed for each of the knowledge sources
(TAMERLAN for text meaning representation, a constraint language for representing ontol-
ogy, and the structure of the lexicon entry for the lexicon). These languages are necessarily
connected in many ways (and in the implementation of DIONYSUS all are based on the
single underlying knowledge representation system FRAMEKIT (Nyberg, 1988). Our com-
ments will touch upon features from both the content and the form facets of tile knowledge
sources.

2 Ontology
In formal semantics, one of the most widely accepted methodologies is that of model-
theoretic semantics in which syntactically correct utterances in a language are given se-
mantic interpretation in terms of truth values with respect to a certain model (in Montague
semantics, a "possible world") of reality. Such models are in practice never constructed
in detail but rather delineated through a typically underspecifying set of constraints. In
order to build large and useful natural language processing systems one has to go further
and actually commit oneself to a detailed version of a "constructed reality" (Jackendoff's
term). Interpreting the meanings of textual units is really feasible only in the presence of
a detailed world model; elements of this model may be linked to the various textual units
(either directly or indirectly, individually or in combinations) via is-a meaning-of links.
World model elements will be densely interconnected through a large set of well-defined
ontological links - properties- which will enable the world modeler to build descriptions
of complex objects and processes in a compositional fashion, using as few basic primitive
concepts as possible.
Knowledge bases in FRAMEKIT take the form of a collection of frames. A frame is

239
 NL Input

I Analyzer
Gmmmmrs

Meaning
repmsentmionI ~.~.~)J

Ontology +
Ressonlng System domain
models

Meaning
mpr~ent~ion

MTDs
Generator

Translations
Abstracts
Dialog turns
Updates of DB or KS

Figure 1: A Knowledge-Based NLP System.

240
a named set of slots. A slot is a named set of facets. A facet is a named set of views
and a view is a named set of fillers. A filler can be any symbol or a Lisp function
call. T h e above is the basic set of constraints and features of FRAMEKIT. T h o u g h
FRAMgKITactually specifies some e~tensions to this basic expressive power, it remains,
by design, quite general and semantically underspecified. The actual interpretation and
typing of its basic entities is supposed to take place in a particular application, in our case,
world modeling. The ONTOS constraint language is built on top of FRAMEKIT. This
strategy is deliberate and is different from many knowledge representation languages and
environments, in which the basic representation language is made much more expressive
at the expense of its relative unwieldiness, difficulty in learning and some format-related
constraints on application development.
In the ontology module of DIONYSUS F R A M E K I T frames are used to represent concepts.
A concept is the basic building block of our ontology. Examples of concepts are house,
four-wheebdrive-car, volunta~-olfacto~-event or specific.gravity. FRAMEKIT slots are
interpreted as a subset of concepts called properties. In practice some properties recorded
in concepts do not relate to the domain model but rather serve as administrative and
explanatory material for the human users (see examples below). The FRAMEKIT views
are used to mask parts of the information from a certain users. This mechanism is useful
when working with very large domain models serving more than one domain so that some
information will be different depending on whether a data element belongs to the domain
of, say, chemistry or the domain of software engineering. FRAMEKIT fillers are constrained
to be names of atomic elements of the ontology, expressions referring to modified elements
of the ontology, collections of the above, or special-purpose symbols and strings used by
the knowledge acquirers.
The example concepts from the ontology serve to illustrate some aspects of the con-
straint language used.
(MAKE-FLJ.ME
MACIKTOSH
(IS-A
(VALUE
(COMMOB PERSORAL-COMPUTER)))
(PRODUCED-BY
(SEN
(COMMOB APPLE) ))
(SUBCLASSES
(SEM
(COMMOK PLUS S E I I IIFX)))
(HAS-AS-PART
(SEH
(COMMOK DISK-DRIVE SYSTEM-U|IT MORITOR CPU MEMORY
EXTERKAL-D ISKETTE-DRI VE IRTERMAL-DI SKETTE-DRIVE) ) ) )

Other slot values inherit from P E R S O N A L - C O M P U T E R (e.g., MAXIMUM-NUMBER-

OF-USERS), or from C O M P U T E R , I N D E P E N D E N T - D E V I C E , DEVICE, ARTIFACT,
etc. In the above example, MACINTOSH is a frame name, IS-A is a slot name, VALUE
and SEM are examples of facet names, COMMON is the view name, and A P P L E and
MEMORY are fillers, for example.
(MAKE-FRAME
VOLUMTARY-OLFACTORY-EVEKT
(IS-A
(VALUE

241
 (COMN0| VOLUITARY-PERCEPTUAL-EVEIT)))
(AGE|T
(SEA
(COMM01 MAMXALBIRD REPTILE A~PHIBIA|)))
(I|STRUME|T
(SEN
(COMMO| OLFACTORY-0RGA|))))

Additional slots such as E X P E R I E N C E R and B E N E F I C I A R Y would be inherited

from among the ancestors of this concept, e.g., V O L U N T A R Y - P E R C E P T U A L - E V E N T
or E V E N T .

3 Representation of Text Meaning

Results of source text analysis in DIONYSUS are represented in a formal, frame-based
language, T A M E R L A N (Nirenburg and Defrise, forthcoming). The types of knowledge
which are represented in T A M E R L A N include the meanings of natural language units and
knowledge about the speech situation, including speaker attitudes. Meaning proper is
represented through instantiating, combining, and constraining concepts available in the
ontology. These processes crucially rely on lexical-semantic information in the appropriate
lexicon entries. This information relates not only to ambiguity resolution in "regular"
compositional semantics (i.e.,straightforward monotonic dependency structure building),
but also the identificationof idioms, treatment of metonymy and metaphor, and resolution
of reference ambiguities.
It is well known, however, that the intent of a text is typically not capturahle by
just instantiating ontological concepts; what is additionally needed is the representation
of pragmatic and discourse-related aspects of language -- speech acts, deictic references,
speaker attitudes and intentions, relations among text units, the prior context, the physi-
cal context, etc. As most of the knowledge underlying realizations of these phenomena is
not society-general but is rather dependent on a particular cognitive agent (a particular
speaker/hearer in a particular speech situation), the pragmatic and discourse knowledge
units were not included in the ontology (which is supposed to reflect, with some vari-
ability, a societal view of the world - - indeed, if every speaker/hearer had a totally id-
iosyncratic world view, communication would not be possible). The representation of this
"meta-ontological" information is thus added to the representation of meaning proper
to yield a representation of text meaning; a resulting representation thus reflects both
lexically-triggered ontologically-based semantic information and non-ontological informa-
tion reflecting pragmatic and discourse factors.
The propositional part of text meaning is obtained by instantiating relevant ontological
concepts, accessed through corresponding lexicon entries, which list mappings between
word senses and ontological concepts. In order to carry out the mappings of complex
structures with dependencies, information about argument structure of lexical units (also
stored in the lexicon) is used.
However, many lexical units in the source text contribute to the nonpropositional
meaning of the text. The structures for the nonpropositional components of text meaning
are also derived from the lexicon, where appropriate patterns are stored (see example
lexicon entries below). The nonpropositional information is not stored in the ontology for
reasons explained above. In what follows we motivate the most important non-ontological
components of our text meaning representation formalism (for more detailed discussion

242
see Nirenburg and Defrise, forthcoming, a) - - speaker attitudes, stylistic features and
rhetorical relations.
A critical facet of capturing the intent of a speaker in a meaning representation is
rendering of the attitudes that the speaker holds toward objects or situations represented
in the propositional (compositional-semantic) component of text meaning representation.
Indeed, the speaker may convey attitudes about producing the text in the first place
(the speech act), elements of the speech context, or even other attitudes. An attitude is
represented by a structure which includes the type of attitude (from the list below), a
value (a decimal value taken from the interval [0, 1] indicating the polarity or strength
of the attitude), an indication of to whom the attitude can be attributed (typically the
speaker), a scope (identifying the entity towards which the attitude is expressed), and a
time (representing the absolute time at which the attitude was held). Types of attitudes
recognized in T A M E R L A N are: epistemic, evaluative, deontic, expectation, volition and
saliency.
The stylistic overtones of a lexical entry, even if not contributing directly to the compo-
sitional semantics of a text, serve a crucial role in conveying the speaker attitudes and in
achieving his intent in producing the text. Thus we identify that the stylistics of a lexeme
needs to be encoded in the lexicon entry in addition to the lexical semantic information.
In encoding lexicons for languages with rich social deictics, such as Japanese, the issue of
stylistics becomes even more acute.
We are utilizing a set of style indicators which is a modification of the set of pragmatic
goals from Hovy 1988. This set consists of six stylistic indicators: formality, simplicity,
color, force, directness, and respect. In our implementation, we need to label lexical entries
(including idioms, collocations, conventional utterances, etc.) with appropriate represen-
tations of values for these stylistic indicators; values for these factors are represented as
points on the interval [0,1], where 0.O is low, 1.0 is high, and 0.5 represents a default,
neutral value. In NL generation, these factors are used in lexical selection; in NL analysis,
the values are available for assisting in disambiguation (relying on expected values for
the factors and utilizing the heuristic that typically the stylistics will be consistent across
words in an utterance).
Some examples of English lexical entries that would need to include style features are:

upside: formality - low

color - high
delicious: formality - somewhat high
color - high
drop by: formality - low
great: formality - low
one: personal-reference - low (pronominal sense)
formality - high
force - low

Relations in TAMERLANare a mixed bag - some of them refer to real-world connections

and are, therefore, defined in the ontology while some others (such as rhetorical relations,
e.g., conjunction and contrast) refer to properties of text itself and are, therefore, not a
part of the world model. At present, the major classes of relations in TAMERLAN include
causal, conjunction, alternation, coreference, temporal, spatial, textual, reformulation and
domain-text relations.

243
4 The Lexicon
An entry in the DIONYSUS lexicon is comprised of a number of zones (each possibly hav-
ing multiple fields), integrating various levels of lexical information. The zones are CAT
(syntactic category), ORTH (orthography - abbreviations and variants), PHON (phonol-
ogy), MORPH (morphological irregular forms or class information), SYN (syntactic fea-
tures such as attributive), SYN-STRUC (indication of sentence- or phrase-level syntactic
inter-dependencies such as subcategorization), SEM (lexical semantics / meaning repre-
sentation), LEXICAL-RELATIONS (collocations, etc.), PRAGM (pragmatics hooks for
deictics, for example, and stylistic factors), and STUFF (user information such as modi-
fication audit trail, example sentences, etc.)
In this paper we concentrate on the SEM zone of the lexicon, since it is the locus of
lexical-semantic analysis. This zone contains two fields - - MEAN-PROC and LEX-MAP.
The MEAN-PROC field invokes procedural semantic functions. One such function is
"intensify-range" used to realize the meaning of intensifiers such as very. This function
operates on scalar attribute ranges as input, returning compressions of the input ranges.
For example, the English word old may involve a range (> 0.8) of the AGE attribute. The
intensify-range function would be invoked in the phrase very old, for example, returning
(> 0.9). Similarly, for very young this function would return (< 0.1), when given (< 0.2).
For very average (perhaps (0.44< x < 0.6)), this function would return (0.45 < x< 0.55).
The LEX-MAP field contains mappings indicating the links between lexical units and
ontological concepts. The simplest version of such a mapping is the straightforward link
between a concept in the ontology and the meaning of the lexical unit in question. In
analysis the link identifies which ontological concept to instantiate as the representation of
a lexical unit. So, for example, the lexical semantics for the primary sense of the English
word clog would be a pointer or link to the concept of "dog" in the ontology: (marked,
by convention, *dog). In DIONYSUS, the semantic analyzer instantiates an instance of the
dog concept (e.g., ~dog3~) when processing a sentence containing the word "dog" or a
reference to it. This mechanism works only in the case of one-to-one mapping between
words and ontological concepts, so that the ontology becomes essentially a collection of
word senses.
There is no consensus in the field with respect to size and composition of the onto-
logical world model. The "word-sense" view of the ontology, in addition to the obvious
disadvantage of ontology size, leads to a problems in multilingual applications - - often
roughly comparable words in different languages do not "line up" the same way, which
leads to proliferation of new concepts with every new language. This makes the trans-
lation process difficult. Another well-known approach prefers a small restricted set of
primitives which are combined or constrained in an attempt to render the meaning of any
lexical unit. This decision has been amply criticized as it leads to difficulty in building
large-scale world models and capturing shades of meaning; additionally, this approach can
yield enormous, unmaintainable lexicon entries for complex concepts. Further discussion
of this issue (and pointers to the literature on this subject) can be found in Meyer et al.,
1990.
The approach taken in DIONYSUS lies somewhere in between the two extremist positions
outlined above. Therefore, in some cases, it is possible to find one-to-one correspondences
between lexeme and concept. Unfortunately, in the majority of cases, this is not possible
without excessive proliferation of concepts. Therefore, mappings are allowed to concepts
with "similar" but not identical meanings - to the most specific concept that is still more

244
general than (i.e., that subsumes) the meaning of the lexeme in question. Once the most
directly corresponding concept is determined, additional information is specified (thus
constraining the general concept) in the LEX-MAP field in the lexicon. This modified
mapping may either add information to the concept as it is specified in the ontology,
override certain constraints (e.g., a selectional restriction), or indicate relationships with
other concepts expected in the sentence.
Before proceeding to illustrate the constrained mappings, we first introduce the knowl-
edge representation tools used to express them. In the DIONYSUS formalism, ontological
concepts are interpreted as frames, and their properties, as slots (see the discussion of
FRAMEKITabove). It is natural, then, to use the facet mechanism to encode constraints
introduced in the LEX-MAP mappings. The constraining or "specialization" facets used
in the DIONYSUS lexicon representations are as follows:

. VALUE - a specific value (e.g., number of sides for a triangle = 3, sex of a man =
male). This is the facet where actual information is represented. When this meaning
pattern is instantiated by use in a sentence, semantic dependency structure links will
be indicated typically in a VALUE facet.

2. DEFAULT - usual (i.e., typical, expected) value (e.g., color of diapers = white).
3. SEM - akin to a traditional selectional restriction (e.g., color of diapers has to be a
COLOR), this is essentially a constraint on what VALUE may be.

4. RELAXABLE-TO - maximum relaxability, if any, of SEM restrictions.

. SALIENCE - a scalar value in the range [0.0, 1.0] designating the significance of a
specific attribute slot or role (partly reflecting the notion of "defining properties"
vs. "incidental properties").

The structure below, for the lexeme "eat-v1" (the primary transitive verb sense) illus-
trates a simple case of lexical semantic mapping. The SYN-STRUC zone of the lexical
entry establishes the subcategorization of the verb, and the structure is used in parsing
or generation. In the course of parsing, the variables S v a r l and Svar2 are bound to
a "placeholder" for the lexical semantics of the subject and object of the verb, respec-
tively. Once the lexical semantics of the syntactic constituents in those syntactic roles is
determined, it becomes available for the compositional purposes - - to build a semantic
representation for a higher-level text component (e.g., a sentence).
(~ingeet
(AGE|T (value "Snarl) ;s u b j e c t o f l e x e m e v i i i g e t bound t o
(eem * a n i m a l ) ) ;Snarl, uhich ¢orresponde to
;AGE|T s l o t o f t h e o n t o l o g i c a l c o n c e p t
;INGEST
(THEME ( v a l u e " $ v a r 2 ) ;o b j e c t o f l e x e m e u i l l g e t bound t o $ v a r 2 ,
(sem * i n g e s t i b l e ) ; u h i c h must c o r r e s p o n d t o t h e THEME s l o t
; of I|6EST
(relaxable-to *physical-object)))) ; indication of
;extent of allouable semantic constraint
; relaxation

As an illustration of the process of semantic dependency structure building, consider

the analysis of the sentence The man visited the store. This analysis will involve the estab-
lishing (among others) of TAMEFtLANentities ~visit132 and ~human135 (the numbers are

245
added for unique reference, the percent sign marks an instance of an ontological concept),
which are instances of the ontological concepts *visit and *haman. This is made possi.
ble through the content of the LEX-MAP fields of lexicon entries "for (the corresponding
senses of) the English words "visit" and "man." The semantics of the entire sentence will
include the instantiated concepts, connected by using Shaman135 to fill the AGENT slot
in ~visit18~. The knowledge necessary for establishing this dependency is found in the
syntactic component of the lexical entry for "visit" (where the t v a r variables would be
shown as corresponding to subject and object positions); this information is then linked to
the semantic sphere using the following notation in the LEX-MAP field of the appropriate
sense of "visit:"
( ~ v i s i t (AOEI! (VALUE " $ v a r l ) ) . . . )

The caret is an operator (akin to an intension operator) which retrieves the meaning
of the syntactic constituent bound to the variable. The resulting TAMEItLAN structure in
will, thus, include the following:
(~visit132 (AQE|T (VALUE~human135)) . . . )

At this point we are ready to illustrate constrained mapping between word senses and
ontological concepts. An example of constraining the sense of an ontological concept is
the lexeme taxi-d, identifying the sense of taxi as an instance of the ontological concept
*move-on-surface, but with important further constraints - - the theme can be only an
aircraft, (e.g., The plane taxied to the terminal). The LEX-MAP field for taxi-v1 is, then,
as follows:
(~move-on-surfac*
(SURFACE (SEN *uater *ground))
(THEKE (SEN * a i r c r a f t )
(RELAXABLE-TO* v e h i c l e ) ) ) ) )

In this example the thing that is moving-on-the-surface is being constrained to be an

aircraft, possibly relaxable to include any vehicle (perhaps in more florid speech). Note
that there may be second-order constraints (i.e., constraints on constraints) in addition
to the above: jet-vl (the literal sense of 'to travel by jet', e.g., The presidential candidate
spent most of the year jetting across the country from one campaign rally to another.)
(~mov@
(IISTRUME|T (SEM * a i r c r a f t
(PROPELLED-BY
(value ~ j e t - e n s i n e ) ) ) ) ) ) )

This constrains the vehicle used in jetting to be a jet-propelled vehicle.

There does not need to be a regular correspondence between the syntactic category
(part of speech) of the lexeme and the general high-level partition into which the corre-
sponding ontological concept falls. The highest node in the ontology, ALL, is partitioned
into the subclasses EVENT, OBJECT and PROPERTY (the latter being partitioned into
ATTRIBUTE and RELATION), which one might expect to correspond to verbs, nouns,
and adjectives or prepositions, respectively. However, there is a great deal of variability in
the correspondence between ontological partition and linguistic part of speech. See Meyer
et al., 1990 for further discussion of this issue.

246
 In some cases, the constraint of an ontological concept may be achieved through the
use of scalar attributes: SHOP = = > STORE of a certain size; CHILD = = > HUMAN
of a certain age; FRESH-BREWED = = > modifies COFFEE or TEA of a certain age.
The RANGE facet is used to access particular information about scalar attributes in
the ontology, such as AGE, TEMPERATURE, LUMINANCE, and SIZE. Any attribute
which can be measured on a scale can also be described in relative terms. For example,
we can say That house is 4000 square feet or That house is very large; That water is 200
degrees (F) or That water is very hot. By means of the RANGE facet, we can establish a
correspondence between measured and relative values for various ontological concepts.

5 Integration of Multiple Knowledge Sources in Lex-

icon Entries
As is evident from the brief discussion above, the lexicon can be seen as the nexus through
which a number of knowledge sources are integrated in order to produce an appropriate
representation of the "meaning" of the input. The clearest case of this is the triggering
of instantiation of ontological concepts by a "meaning template" in the LEX-MAP field
of the lexicon entry. Furthermore, the additional information (augmenting and constrain-
ing information) presented in the LEX-MAP field provides a mechanism of tailoring the
cross-linguistic ontological (world model) information to language-specific, lexeme-specific,
representation of a conceptual notion.
In some lexicon entries LEX-MAP contains a mapping not into an ontological concept
but rather into the representation of a pragmatic or discourse entity, such as a relation
or an attitude. In other cases, lexical units have mixed ontological and non-ontological
meaning components. In both cases, this is a mechanism for providing for the integration
of non-ontologicM (hence "non-semantic") information with ontological, as triggered by
lexical entries. The following examples will illustrate such phenomena.
The following is an example of how an evaluative attitude can be used for describing
the meaning of a lexeme:

excsllent-adjl
(ATTITUDE
(type ( v a l u e e v a l u a t i v e ) )
(attitude-value (value 1))
(scope " $ v a r l ) ) ; $ v a r l corresponds to the o b j e c t f o r which the
; a t t i t u d e i s h e l d , as i d e n t i f i e d s y n t a c t i c a l l y
(attributed-to (value [$SS].deictic-indices.speaker))
; the holder of the a t t i t u d e i s the speaker ( i . e . , the
; p r o d u c e r of the t e x t )

The entry for smell-v5 (defined as perceive something negative intuitively, as in I could
smell trouble brewing or Everyone could smell the hostility in the air) demonstrates the
integration of ontological and non-ontological structures:

(~perceptual-event
(EXPERIE|CER ( v a l u e " $ v a r l ) ; l i n k t o t h e s y n t a c t i c s u b j e c t
(sem *human))
(THEME ( v a l u e " $ v a r 2 ) ; l i n k to the s y n t a c t i c o b j e c t
(sem * m e n t a l - o b j e c t ) ) )
(ATTITUDE
(type ( v a l u e e v a l u a t i v e ) )

247
 (value ( v a l u e (< 0 . 2 ) ) ) ; l e s s than . 5 , s o n e g a t i v e f e e l i n g s
(scope (Zporceptual-event.theme)) ;link to the contents of
; the indicated slot in the ontological
; concept ~perceptual-event referenced above.
(attributed-to
(value [$SS].deictic-indices.speaker))))) ; u pointer
; t o t h e d e i c t i c indQx i n t h e c o n t e x t frluae

Here there is a restriction on the EXPERIENCER, who must be a human, and on the
THEME, which must be a MENTAL-OBJECT. A further item of information, however,
is that there is a negative attitude towards the theme (e.g., one does not normally say
things like I could smell happiness in the air, Everyone could smell the positive attitude she
had, etc.) The attribution of the attitude is a hook into the deictic indices of the speech
situation of the utterance - the formula in the attribution identifies that the speaker of
the current speech situation be identified, and that the speaker be linked in as the holder
of attitude.
The semantic mapping mechanism as a representation language in the LEX-MAP field
of the lexicon (outlined above) allows for a fluent integration of encyclopedic knowledge
into the framework without "corrupting" the linguistic lexicon with this type of infor-
mation. The approach in question is the formation of an "episodic memory" knowledge
source, consisting of a set of data structures identical to lexicon entries. However, the
person Ronald Reagan would be represented in the episodic memory as a named instance
of the concept *human, or some more specific concept, along with other information
known about him rendered in various slots of the concept instantiation, not by a trigger
to produce a new instantion of the concept as would be the case by a linguistic lexicon
entry.

6 Status and Future Work

All three static knowledge sources described briefly above are still under development.
In fact, only when the ontology and the lexicon reach nontrivial size, can one put a
computationally relevant lexical-semantic theory to a real test in a large-scale application.
A complete theory of lexical semantics must, of course, include the description of the
actual processes which lead from a natural language text to its meaning representation
(NL analysis) and in the opposite direction (NL generation). We continue to develop these
processes, in the tradition of knowledge-based NLP.

Acknowledgements
We gratefully acknowledge the contributions of Christine Defrise, Ingrid Meyer, and Lynn
Carlson.

References
Carlson, Lynn and Sergei Nirenburg. 1990. World Modeling for NLP. Technical Report
CMU-CMT-90-121. Center for Machine Translation. Carnegie Mellon University.
Goodman, Kenneth and Sergei Nirenburg (eds.) 1989. KBMT-89. CMU-CMT Project
Report.

248
Hirst, Graeme. 1988. Semantic Interpretation and the Resolution of Ambiguity. Cam-
bridge: Cambridge University Press.
Hirst, Graeme. 1989. Ontological Assumptions in Knowledge Representation. In Pro-
ceedings of the First International Conference on Principles of Knowledge Representation
and Reasoning (Toronto, May 1989).
Hovy, Edward. 1988. Generating Natural Language Under Pragmatic Constraints. Ph.D.
Yale University.
Meyer, Ingrid and James Steele. 1990. The Presentation of an Entry and of a Super-entry
in an Explanatory Combinatorial Dictionary. In James Steele (ed.), The Meaning-Text
Theory of Language: Linguistics, Lexicography, and Practical Implications. Ottawa,
Canada: University of Ottawa Press.
Meyer, Ingrid, Boyan Onyshkevych, and Lynn Carlson. 1990. Lexieographic Principles
and Design for Knowledge-Based Machine Translation. Technical Report CMU-CMT-90-
118. Center for Machine Translation. Carnegie Mellon University.
Monarch, I. and S. Nirenburg. 1988. ONTOS: An Ontology-Based Knowledge Acquisition
and Maintenance System. Proceedings of the Second Workshop on Knowledge Acquisi-
tion. Banff, Canada.
Nirenburg, Sergei, and Christine Defrise (forthcoming, a). Practical Computational Lin-
guistics. In R. Johnson and M. Rosner (eds.), Computational Linguistics and Formal
Semantics. Cambridge: Cambridge University Press.
Nirenburg, Sergei, and Christine Defrise (forthcoming, b). Lexical and Conceptual Struc-
ture for Knowledge-Based Machine Translation. In: J. Pustejovsky (ed.), Semantics and
the Lexicon. Dordrecht, Holland: Kluwer.
Nirenburg, Sergei and Kenneth Goodman. 1990. Treatment of Meaning in MT Systems.
Proceedings of the Third International Conference on Theoretical and Methodological Is-
sues in Machine Translation of Natural Language. Linguistic Research Center, University
of Texas at Austin. June 1990.
Nyberg, Eric. 1988. The FrameKit User's Guide, Version 2.0. CMU-CMT-MEMO. Center
for Machine Translation. Carnegie Mellon University.

249