Interpretation without Semantics 
Stephen Helmreich 
Computing Research Laboratory 
Box 30001/3CRL 
New Mexico State University 
Las Cruces, NM 88003-0001 
shelmrei@nmsu, edu 
Abstract 
I suggest that Geoffrey Nunberg's work can make two contributions to the issue 
of the relationship of lexical semantics to knowledge representation. The first is 
the claim that Lexical Semantics and Knowledge Representation are the same. The 
second (which is both more controversial and more difficult to implement) is that the 
connection between this world knowledge and lexical items must be quite flexible, if 
not, in fact, non-existent and constructed anew for each interpretive act. I outline 
several arguments for this position and present a formal method of incorporating 
this within a standard Montague-grammar framework. 
1 Introduction 
In this paper, I attempt to relate the NLP issue of the relationship between lexical se- 
mantics and knowledge representation to the claims made by Geoffrey Nunberg in The 
Pragmatics of Reference \[1\]. His basic claim is that interpretation of lexical items in an 
utterance is almost entirely a pragmatic matter and not one related to language-specific 
semantics. He uses a theory of referring functions to account for the variety of uses to 
which any one lexical item may be put. 
Though phrased in linguistic terminology, the corollary of this thesis for NLP is that 
purely linguistic lexical information is quite limited (perhaps only syntactic part of speech 
and phonological information), with encyclopedic and world-knowledge information con- 
stituting most of the knowledge base of the system. In addition, Nunberg claims that the 
connection between these linguistic symbols and the information in the knowledge base 
must be quite flexible (and, perhaps unspecified). 
The first claim echoes those viewpoints which see no justification for differentiating 
lexical semantics and world knowledge. It is the second claim that differentiates what 
might be called a "radical semantics" position from a "radical pragmatics" approach. In 
both cases, lexical knowledge is treated identically with world knowledge. In the first 
case, however, this information is regarded as "semantics", knowledge about the meaning 
of the lexical items. (Langacker \[2\], Lakoff \[3\], Martin \[4\] and Pustejovsky \[5\] can be seen 
as typical of this approach.) In the second, it is regarded purely as information about 
the world and the connection between the use of lexical items in language and this world 
knowledge is the result of pragmatic interpretation. 
Since it is this second claim that would require the most radical change in NLP systems, 
I shall concentrate on this aspect of Nunberg's approach. In the remainder of this position 
paper I will first restate some of the arguments of Nunberg for such a pragmatic approach 
to lexical interpretation. Then I shall set out in brief the approach I take in formalizing 
this approach within a typical Montague Grammar. 
34 
2 Argumentation 
The first argument points to the many uses of lexical items that cannot be "stored" as 
part of the semantics of the item, but rather must be inferentially derived as part of the 
interpretation of the utterance. These include self-referential usages, ad hoc usages, and 
metonymical and metaphorical uses. 
Self-referential uses. Example: Rice starts with an "r'. Since these types of usages 
(referring to the phonological, orthographical, phonetic or other characteristics of the 
lexical item itself) are universal for every lexical item in every language, it is clearly 
wasteful to store this idiosyncratically as part of the "meaning" of the lexical item. 
Ad hoc uses. Example: Your coat is on the refrigerator door. In these cases, the 
interpretation depends on highly context-specific information. In the example sentence, 
the intent of the speaker was to refer to the dry-cleaner's claim ticket which was attached 
to the refrigerator door by a magnet. 
Metonymy and metaphor. More complex issues arise with various common "literary" 
usages such as metaphor or metonymy. There are claims on both sides for idiosyncratic 
lexicalization and also for free generation (Martin \[6\], Fass \[7\], and Pustejovsky \[5\]). 
Taken together these types of usage are only suggestive. They require an extension, 
an analytic leap: if these types of usage are be handled pragmatically, perhaps all types 
of usage should be so handled. 
At this point, it should be pointed out that simply because particular information is (or 
is most conveniently) stored rather than derived is not an argument that this information 
is semantic and not pragmatic. As long as it can be so derived, it can be regarded as 
stored pragmatic information. 
A second argument in favor of treating the relationship between lexical items and 
world knowledge as pragmatic rather than semantic is an Ockham's razor argument: 
If one can exhibit a pragmatic system that performs lexical interpretation as well as 
an idiosyncratic language-specific semantic system can, then it would be best to opt 
for the pragmatic system. This is because the sub-systems needed by the pragmatic 
system (world-knowledge, context knowledge, and inferencing, at a minimum) are all types 
of information and processing that are needed for other independent and non-linguistic 
activities. Why duplicate them with a specific linguistic-semantic system? Note that this 
argument in particular requires a demonstration of the existence and workability of such 
a system. 
Yet a third argument of Nunberg that I feel has not been sufficiently understood and 
internalized is that of the idealized speaker/hearer of the Chomsky model. Nunberg claims 
that this model is sufficient for phonology and syntax, but fails for semantics. Clearly at 
an individual level, the experiences on which world-knowledge is based varies widely from 
one to another, and in addition, (even given a relative commonality of experience) the 
way in which this knowledge is organized can be expected to vary widely from one person 
to another. Therefore, there seems to be no reason for lexical knowledge to be in any way 
identical or identically organized from person to person. 
On yet another level, Nunberg claims that the way in which this lexical knowledge is 
utilized requires varying ways of looking at this knowledge. Thus the meaning that one 
attributes to a word in a particular context varies with the type of interaction, the social 
grouping in which one finds oneself, and the particular assumptions one makes about the 
interlocutors. Examples of this kind of varying usage are such terms as "classical music" 
or "jazz". In one context, the term "classical music" might include Broadway musicals, 
35 
while in another it may refer only to the music of Mozart and early Beethoven. 
While speakers of the same language use the same lexical items and the same syntactic 
structures, their internal representation of world knowledge and the way they connect 
these representational structures with linguistic items may vary greatly from person to 
person and from social group to social group. Speakers of a language must take this into 
account in trying to communicate. They cannot assume that all other speakers have the 
same shared knowledge, organized in the same way, and connected identically to the same 
lexical items. 
The implications of this for NLP is that any language processing system should be able 
to work equally well with world-knowledge representations that may vary significantly in 
format, structure, and contents. If lexical items are strictly tied to elements in a particular 
knowledge base, this is not possible. 
3 Implementation 
It remains, however, to demonstrate that such a system is feasible. In the rest of this 
position paper, I shall outline the manner in which I have attempted this. Formalizations 
of Nunberg's referring functions have been made by Sag \[8\] (on a small scale) and by 
Fauconnier \[9\] (in a more cognitive context). 
I have taken a typical Montague Grammar (truth-conditional/model theoretic) ap- 
proach (Dowty, Wall, and Peters \[10\]) and eliminated the function which provides inter- 
pretations for each lexical item in the language. Instead, this is replaced by an Interpretive 
Component (which can be viewed ill tile large as related to Kamp's Discourse Represen- 
tation Theory \[11\]). 
The only requirement on the model is that it contain elements that map onto the lexical 
items of the language (thus allowing for self-reference). In fact, the only fixed "semantics" 
of the system associates lexical items in the language with these representative elements 
in the model. 
In the Interpretive Component are Montague Grammar operators which function like 
Nunberg's referring functions. That is, they operate on items (and sets of items, and 
intensions of sets of items) and return other items, sets of items, and intensions of sets of 
items. These referring functions (MG operators) also contain usage/context and speaker 
indicators. 
Certain referring functions operate on the lexical items in the model and return sets 
of things that particular speakers (in particular contexts) often refer to by using that 
particular lexical item. This is the closest to a "meaning" that the system contains. 
However, it is not part of the language, may vary greatly from speaker to speaker of the 
language, and is subject to certain rules of application before an interpretation can be 
generated. 
These particular referring functions are not directly inserted into the sentence to pro- 
vide its interpretation. Rather, it is a matter of co-determining a context and an inter- 
pretation. Part of the context is already given, while part is abductively added by the 
referring functions themselves. 
I can show that this type of approach is useful for several purposes. First of all, it 
provides an existence proof for a semantic system that is both compositional and has no 
lexical semantics, only a lexical pragmatics-a system in which lexical semantics is highly 
constrained and the connection between the lexical items and the encyclopedic knowledge 
36 
is flexible. Second the concepts involved can be used to usefully distinguish between 
metonymy and metaphor (by determining whether or not the referring functions used 
have 1-1 or 1-many extensional equivalents). It also allows a definition of "same lexi~al 
item," that is, two senses are the same lexical item if they~are reachable through the set 
of given referring functions. 
Although not an implementation of this approach, I would like to conclude by making 
a brief reference to the work that Eric Iverson and I have been doing on meta5/metallel- 
a program of word sense disambiguation that also identifies instances of metaphor and 
metonymy (Helmreich et al \[12\]). Although our research has not been directly reflective of 
the concerns of lexical knowledge versus world knowledge, we are working with a program 
which treats both types of knowledge as equivalent. That is, the structure of the lexicon 
contains both encyclopedic and what would be called strictly lexical information. In 
addition, metaphorical and metonymical usages are not directly stored in the lexicon, but 
are derived inferentially on the basis of the linguistic and world-knowledge context. 

References 
\[1\] G. Nunberg, The Pragmatics of Reference, Indiana University Linguistics Club, Bloom- 
ington IN, 1978. 
\[2\] R. Langacker, Foundations of Cognitive Grammar, Stanford University Press, Stanford 
CA, 1987. 
\[3\] G. Lakoff, Women, fire, and dangerous things, University of Chicago Press, Chicago IL, 
1987, University of Chicago Press. 
\[4\] J.H. Martin, "A unified approach to conventional non-literal language", Proceedings 
of the Fifth Rocky Mountain Conference on Artificial Intelligence (RMCAI-90), Las 
Cruces, NM, 1990, pp. 5-40. 
\[5\] J. Pustejovsky, "The semantic representation of lexical knowledge" Proceedings of the 
First International Lexicai Acquisition Workshop, 1989. 
\[6\] J.H. Martin, A Computational Model of Metaphor Interpretation. Academic Press, Cam- 
bridge MA, 1990. 
\[7\] D. Fass, Met*: A Method \]for Discriminating Metonymy and Metaphor by Computer. 
CSS/LCCR TR 89-15, Centre for Systems Science: Burnaby, B.C., 1989. 
\[8\] I. Sag, "Formal Semantics and Extralinguistic Context", Radical Pragmatics, P. Cole, 
ed., Academic Press, New York NY, 1981. 
\[9\] G. Fauconnier, Mental Spaces: Aspects of Meaning Construction in Natural Language. 
MIT Press, Cambridge MA, 1985. 
\[10\] D. Dowty, R. Wall, S. Peters, Introduction to Montague Semantics, Synthese Language 
Library 11, D. Reidel, Dordrecht Holland, 1981. 
\[11\] H. Kamp, "A Theory of Truth and Semantic Representation", J.A.G. Groenendijk, 
T.M.V. Jansen, and M.B.J. Stokhof, eds., Formal Methods in the Study of Language, 
Mathematisch Centrum, Amsterdam, 1981. 
\[12\] S. Helmreich, E. Iverson, and F. Laroche, "Modular Meta5: Further Research in Colla- 
tive Semantics", Memoranda in Computer and Cognitive Science, MCCS-90-192, Com- 
puting Research Laboratory, New Mexico State University, Las Cruces NM, 1990. 
