Syntax-Driven and Ontology-Driven Lexical 
Semantics 
Sergei Nirenburg and Lori Levin 
Center for Machine Translation 
School of Computer Science 
Carnegie Mellon University 
{ sergei, lsl} @nl.cs.cmu.edu 
Abstract 
In this position paper we describe the scopes of two schools in lexicM semantics, 
which we call syntax-driven lexical semantics and ontology-driven lexical semantics, 
respectively. Both approaches arc used in various applications at The Center for 
Machine Translation. We believe that a comparative analysis of these positions and 
clarification of claims and coverage is essential for the field as a whole. 
There are different traditions in the study of lexical semantics. Two of them seem to 
be the most current in computational linguistics and its applications -- the one based 
on syntactic theory studies and the other, on the artificial intelligence approaches. The 
former seeks to discover semantic properties of lexical items from which syntactic behavior 
(such as subcategorization and participation in transitivity alternations) is predictable. 
(See Grimshaw, 1990; B. Levin & Rappaport Hovav, 1990.) The latter tries to establish 
the meaning of natural language texts with the help of an independently constructed 
"world model" (often called "ontology") which explicates relations among entities in the 
world rather than lexical units. 
It is customary to pitch the two approaches as competing. And in practice, researchers 
typically develop lexical-semantic theories and computer systems based on them using 
only one of the two. We would like to argue, however, that these approaches to semantics 
are much closer to one another in their aims and research methodologies than to any 
other schools of semantics ("logical," "philosophical," "formal," etc.) From a practical 
standpoint, we also have a different experience to report. We believe that neither on 
theoretical grounds nor with respect to computational applications is there a necessity 
to make one of the approaches prominent at the expense of the other. We have carried 
out both theoretical and practical work in which the approaches seem to coexist. The 
central role of the syntax-driven lexical semantics in the process of NLP is to decode the 
nature of the dependency between heads of phrases and their arguments in a particular 
language. This knowledge is then used in ontology-driven lexical semantics as a necessary 
set of heuristics which allow us to represent the meaning of a text in terms of a language- 
independent conceptual model. Thus, we believe that a comprehensive approach should 
combine the benefits of both approaches and that neither will on its own be sufficient for 
realistic NLP. 
1 Syntax-Driven Lexical Semantics 
In a knowledge-based system, the primary role of syntactic analysis is to decode knowl- 
edge that is syntactically encoded. An important component of such knowledge is the 
relationship between predicates and their arguments. For example, in order to know that 
the English sentence Max interviewed Hesler for a job means that Max was considering 
hiring Hester and not that Hester was in the position to hire Max, it is necessary to know 
that the interviewer role is expressed as the subject of the sentence, that the interviewee 
role is expressed as the object, that the subject precedes the verb and that the object 
follows it. Much of the information about the syntactic encoding of arguments can be 
placed in lexical entries of argument-taking predicates. Classes of predicates with simi- 
lar argument encodings usually have some semantic similarity, and thus define a kind of 
lexical semantics based on semantic features that are predictors of syntactic behavior. 
In light of these observations, our theory of lexical semantics includes a component 
of lexical knowledge that describes the syntactic encoding of arguments. This lexical 
knowledge consists of Lexical Conceptual Structures (LCS) and Linking Rules. We will 
assume that the LCSs of any given language are composed from a universal set of building 
blocks (including, but not limited to those described by Jackendoff, 1983, 1990). However, 
the presence of lexical gaps--meanings that are lexicalized in some languages but must be 
expressed phrasMly in another language--suggests that languages do not all share exactly 
the same inventory of LCSs. The most important property of the LCSs of any individual 
language is that they contain all of the information necessary for Linking Rules to calculate 
the syntactic encoding of predicates and arguments. Thus they contain syntactically 
relevant semantic information. 
Linking Rules describe how LCSs are syntactically encoded in a language and they 
partition LCSs into classes of predicates that have similar syntactic encodings. Linking 
Rules and the verb classes they define are language-specific (Mitamura, 1989), but gov- 
erned by language-universal principles. The particular version of linking rules and verb 
classes we have implemented is based on Lexicai Functional Grammar (L. Levin, 1987; 
Bresnan & Kanerva, 1989), and includes the treatment of the phenomena briefly discussed 
in the rest of this section. 
1.1 Which Grammatical Function Is Associated with Each Argu- 
ment? 
Linking Rules associate grammatical functions with arguments in a LCS. For example, 
the English verb interview associates its interviewer role with the grammatical function 
SUBJECT and its interviewee role with the grammatical function OBJECT. Argument- 
taking predicates themselves fall into subcategoriies (or subcategorization classes) based 
on the grammatical functions that they assign to their arguments. For example, those 
that assign SUBJECT and OBJECT, those that assign SUBJECT and OBLIQUE, etc. 
1.2 Which Syntactic Position, Case Marker, or Agreement Mor- 
pheme Encodes Each Argument? 
After linking LCS arguments to grammatical functions, the next step in the syntactic 
encoding of LCS arguments is to assign a syntactic position or morphological marker to 
each argument. For example, in English, the interviewer role is syntactically encoded 
to the left of the verb. In another language, the verb might specify a case marking for 
the interviewer role. Default information about syntactic encoding of arguments (e.g., 
SUBJECTs have nominative case) can be specified in the syntactic rules of the language, 
instead of in the lexicon. However, after these defaults are specified, there remains a 
10 
considerable amount of lexically specific information--which prepositions mark oblique 
arguments, which verbs take dative objects, and so on. 
1.3 Which Predicates Participate in Alternate Assignments of 
Grammatical Functions? 
Many predicates allow multiple assignments of grammatical functions or case markings to 
their arguments. For example, the well-known spray/load alternation involves the assign- 
ment of the grammatical function OBJECT or OBLIQUE to an argument that is filled or 
covered, as in spray the wall with paint or spray paint on the wall. The passive construc- 
tion also results from an alternation in grammatical function assignment to arguments. 
An argument that has the OBJECT function in an active sentence has the SUBJECT 
function in a passive sentence. 
Argument-taking predicates fall into classes according to which alternations they allow. 
When a set of LCSs is identified that all allow the same alternations, it is usually possible to 
identify a semantic feature that they have in common that correlates with the transitivity 
alternation. Hence LCSs that undergo the same transitivity alternations appear to form 
semantic classes -- for example, the change of state verbs, the spray-load verbs, verbs of 
removing, verbs of creation, etc. (See B. Levin (1989) for an extensive list of such classes.) 
1.4 In Which Argument-Structure-Building Operations Does the 
Predicate Participate? 
Talmy (1975, 1985), Jackendoff (1990), and B. Levin & Rapoport (1988) among others 
have proposed rules that conflate two lexical meanings to create a new lexical meaning. 
Such argument-structure-building rules in English include incorporation of manner of mo- 
tion and directed motion (the argument structure for jump and the argument structure for 
go combine to form the argument structure for jump into the room); resultative secondary 
predication (the argument structure for become and the argument structure for hammer 
combine to form the argument structure for hammer the nail fiat); and many others. For 
these operations, we are assuming that two LCSs are combined to form a new LCS, and 
this new argument structure undergoes assignment of grammatical functions and syntac- 
tic encoding of arguments. The allowable types of LCS conflation and the classes of verbs 
that undergo them vary from language to language. 
All of the operations mentioned in this section identify syntactic classes of verbs--the 
verbs that undergo the same transitivity alternations, the verbs that map their arguments 
onto the same grammatical functions, the verbs that undergo a certain argument structure 
building operation, the verbs that encode their arguments syntactically in a certian way. 
It is usually possible to find semantic features of LCSs that correlate with these classes-- 
e.g., a change of state, an effect on a patient, etc. In this sense syntactic patterns of verb 
classes define a lexical semantics. This lexical semantics is language-specific in that the 
membership of verb classes and the semantic features that they are based on are different 
in different languages. 
11 
2 Ontology-Driven Lexical Semantics 
While the focus of attention in syntax-driven lexical semantics is on components of lexical 
meaning that determine syntactic behavior, the task of ontology-driven lexical semantics is 
to determine and specify meanings of lexieal units, suggest how they must be represented 
(in the lexicon) and how they contribute to the overall representation of text meaning. We 
view this task as a component of the more general task of representing the meaning of a 
text in a computational application. This view has been developed mostly in the tradition 
of natural language processing within artificial intelligence. (It is impossible to reference 
all of the contributors to this, by no means monolithic, paradigm; a very incomplete list of 
references might include Wilks (1975), Schank (1973), Sowa (1984), Hirst (1987), articles 
in Hobbs and Moore (1985) or Lehnert and Ringle (1982)). 
Some of the desiderata for an ontology-driven lexical semantics are discussed in the 
remainder of this section. 
2.1 It Must Be Constructive 
Among our differences from a typical practice in model-theoretical semantics is our "con- 
structive" attitude to the models. We believe that in order to be able to discuss model- 
theoretic semantics, the models (that is, the ontologies) must first be constructed and put 
in correspondence with the lexicon. This premise also sets us apart from some work in 
linguistic semantics, for instance, Jackendoff's. At the same time, we do not strive to use 
a very limited set of primitive concepts. 
One way of constructing a model is to come up with a set of properties describing 
things in the world, define value sets for these properties and then describe each concept 
in the world as a set of particular property-value pairs. For reference purposes such sets 
can be named. The result is a world model, or computational ontology. A number of 
ontologies has been recently suggested and built, though not al of them with language 
processing in mind (cf. Lenat and Guha, 1990, Dahlgren and McDoweil, 1989, etc.). 
If some of the properties reflect links among ontological units, then the ontology can 
be viewed as a multiply interconnected network. Such a network can support "reasoning" 
about the world encoded in it, if special heuristic rules for network traversal (including 
inheritance) are defined (see, e.g., Woods (1975) Brachman (1983), Horty et hi. (1990)). 
2.2 It Must Support Disambiguation 
In a computational environment, syntax-driven lexical semantics helps distinguish mean- 
ings but not represent them. Indeed, in some cases knowledge of predicate-argument 
structure can disambiguate a lexical unit entirely. Thus, the three senses of spray in 
LDOCE (illustrated by spray paint on a wall; the water sprayed out over the garden and 
spray a wall with paint) can be completely disambiguated on the basis of differences in 
their predicate-argument structure. However, after the choice has been made, the correct 
meaning needs to actually be represented. That is, we need a representation of mean- 
ing that shows how the different senses of spray are different, not simply that they are 
different. 
However, in some cases a lexical unit cannot be completely disambiguated after the 
predicate-argument analysis is performed. Consider the English verb hack. We call dis- 
tinguish two senses of the word -- the cutting and tile programming one. The difference 
12 
between these senses cannot be expressed in terms of the predicate argument structure. 
We will need to specify, for instance, that in one of the senses the instrument of the event 
is a knife or an axe and in the other, a computer. This type of information is, in fact, a 
familiar selectional restriction. Specifying selectional restrictions is different from specify- 
ing predicate-argument relationships. The specification of selectional restriction falls into 
the purview of ontology-driven lexical semantics. 
In the lexis realm, ontology-driven lexical semantics provides the residue of mean- 
ing beyond the predicate-argument structure. Additional disambiguating power is to be 
provided on top of the verb class distinctions of the syntax-driven lexical semantics-- for 
instance, selectional restrictions have to be formulated. Therefore, the lexicon should con- 
tain knowledge that is not required for decoding the syntactic realization of arguments. 
Naturally, this calls for additional expressive power in the formalism and, crucially, for 
reliance on an independently developed model of the world as the source of meaning 
interpretation for lexical units (see next item). 
2.3 Metalanguage Must Be Different from Object Language 
No matter how extensive a computational dictionary is prepared for an application, there 
will always be cases when a given set of selectional restrictions will fail to help disambiguate 
a text. This means that even more disambiguating power should be added to the arsenal of 
a lexical-semantic theory. The cases where selectional restrictions are not powerful enough 
are typically cases of metonymy or metaphor. Note that classification of phenomena as 
treatable by selectional restrictions or only by metonymy/metaphor processors crucially 
depends on the content of the dictionary -- the less selectional restriction information in 
the dictionary and the fewer word senses distinguished, the more work remains for the 
metonymy/metaphor processor. 
Since processing metaphor and metonymy is more effort-consuming than using selec- 
tional restrictions, it is natural that lexical semanticists have been searching for ways of 
reducing the need for the former. One way is to require very large and detailed lexicons. 
In part, this is the reason why some researchers suggested that word senses from a natural 
language could be used as elements of a metalanguage. 
We believe in the distinction of metalanguage and object language and therefore claim 
that the meanings of natural language units should not be represented using unadorned 
lexical units of the same language. Ill the most general case, the meaning-encoding met- 
alanguage should be a full-blown language-independent conceptual world model. 
Constraints like selectional restrictions and distinctions among word senses have to be 
expressed using symbols -- either words of a natural language or elements of an artificial 
"language of thought" (in the sense of Fodor, 1975). One problem with using words of a 
natural language is that they are typically themselves ambiguous and therefore a computer 
program which has to use them as disambiguation heuristics will run into additional 
disambiguation requirements. 
Another reason for not basing an ontology on word senses of a natural language is that 
there are cases where different word senses, as specified in a human-oriented dictionary, 
seem to be divided quite artificially. Consider, for instance, the entry for preempt in 
LDOCE. It has four senses all of which conform to a single predicate-argument structure. 
All of them have something to do with having authority for preferential treatment with 
respect to goods or services. Ill fact, if one is so bent, one can think about distinguishing 
more than four such senses, at about the same level of generality. We believe that such 
13 
cases are prime candidates for merging the senses into one and treating the semantic 
analysis process for them in the metonymy/metaphor mode. 
The above seems to argue that if automatic seeding of a computational lexicon is 
undertaken from an MRD, a nontrivial manual pass over the resulting lexicon should be 
carried out to make sure that the grain size of sense distinction is commensurate with the 
lexical and conceptual detail in the corresponding domain model. 
2.4 Procedures for Assignment of Meaning Are an Integral Part 
of the Theory 
Ontology-driven lexical semantics defines the meaning of a lexical unit through a mapping 
between a word sense and an ontological concept (or a part thereof). The process of text 
analysis, then, consists essentially of instantiating the appropriate ontological concepts or 
modifying property values in concepts that have already been instantiated. 1 
2.5 It Must Offer a Treatment of the Phenomena of Synonymy, 
Hyponymy and Antonymy 
In a theory where lexical meaning is expressed in terms of mappings into instances of 
ontological concepts, these standard lexical-semantic relations are explained in terms of 
relations among elements of a world model. Thus, synonymy will be defined as a relation 
among the lexical units at least some of whose senses map into the same ontological 
concept. 2 Lexical unit A is defined to be a hyponym of lexical unit B if the concept in 
which at least some of the senses of A map stands in the taxonomical is-a relation to the 
concept into which lexical unit B is mapped. Antonyms are defined on lexical units which 
map into regions on an attribute scale. Specifically, antonyms map into regions that are 
symmetrically positioned around the center point of the scale. 3 
2.6 Text Meaning Representation Language Should Not Be Too 
Close to the Syntax and Lexis of Any Natural Language 
Since the purpose of an LCS in syntax-driven lexical semantics is to reflect the syntacti- 
cally relevant semantic properties, that is those properties that determine how an LCS is 
mapped into a syntactic dependency representation, the LCSs are usually closely tied to 
the syntax and lexis of a specific language. They are therefore not as useful in multilin- 
gual computational environments as in monolingual ones. This is because languages differ 
in their lexicalization patterns and in patterns of syntactic dependency. Following are 
some examples that call for departure from the lexical semantics of individual languages 
in muir-lingual processing enviroments. 
1 Lexical-semantic information is, of course, only one of several components of a text meaning represen- 
tation. In practice, additional information (e.g., pragmatic and discourse-related) is encoded in lexicon 
entries and is represented as part of text meaning. 
2It should be noted that in our application of lexlcal semantics, the I~ltmvsus project, such mappings 
are allowed to be constrained, not just univocal (for instance, the English verb taxi maps, in one of its 
senses, to the concept move-on-surface but only if the theme of this instance is constrained to the concept 
of airplane or one of its ontological descendents) -- see Onyshkevich and Nirenburg (1991) for additional 
detail. 
aNote that synonymy on scalars is defined as mapping onto the same region of a scale. 
14 
A head-modifier relationship can be realized in opposing ways in different languages; 
thus, a morpheme carrying the aspectual meaning can be realized as the head of 
a phrase (like the verb continue or finish), with a word denoting an event as its 
complement; alternatively, the word denoting the event can be realized as the main 
verb while the aspectual meaning can be realized through an adverb or even a affix. 
Lexical Gaps: Russian has no word that corresponds exactly to the English word 
afford (as in I can't afford X or I can't afford to Y). In a multilingual process- 
ing environment, there might be a concept corresponding to a sense of the English 
word afford. A Russian sentence Jane mogu sebe etogo pozvolit' (I can't allow my- 
self this), uttered in a context of acquisition -- which could have been assigned a 
straightforward lexical-semantic representation if we were building a lexical semanr 
tics for Russian alone -- should involve the concept that represents afford. This 
means that if the units of the representation language are chosen so that they are 
based on Russian lexis, the meaning of afford will be missing. But this meaning 
seems sufficiently basic to be included in an ontology. As a result, if lexical patterns 
of many languages are used as the clues for building ontology, the quality of the 
latter should increase. 
Different languages describe the same event using different lexical semantics. For 
example, the ritual act of washing one's body in Islam is expressed with a change of 
state verb in Arabic (to get washed), as a verb of receiving in Indonesian (to receive 
ablutions), and with an agentive verb in French (to make ablutions) (Mahmoud, 
1989). 
For reasons such as the above, we argue that, at least in some applications, such as 
multilingual MT, a more language-independent representation scheme is preferable. 
3 Our Position on Some of the Questions Posed in 
the Call for Papers 
On the basis of the above discussion, our opinions on some of the questions posed in the 
call for papers are as follows. 
3.1 What is World Knowledge and What Is Knowledge of Lan- 
guage 
Lexical knowledge of a language is relevant for mapping phrases in syntactic structures 
onto argument positions of a predicate. World knowledge is relevant for representing the 
distinctions among the senses of particular predicates and arguments which cannot be 
distinguished by distributional characteristics. 
3.2 Cross-Linguistic Evidence for the Specificity of Lexical Se- 
mantic Representation 
Our experience in medium-scale KBMT projects shows that verb classes defined by linking 
rules in different languages are not identical. This means that the syntax-driven lexical 
semantics component of work on a computational application will be different for every 
15 
language involved. For example, in Arabic there is a transitivity alternation that is similar 
to the causative/inchoative alternation (Someone broke the glass / The glass broke) in En- 
glish. However, The English and Arabic transitivity alternations apply to different classes 
of verbs. The Arabic rule, for instance, applies systematically to verbs of psychological 
change of state (excite, please, stun, entertain, confuse, etc.), but the English rule does 
not apply to these verbs. Since our goal in developing KBMT systems is to support mul- 
tilinguality, we preferred to create a single, deeper, interlingual meaning representation 
which will not be subject to the divergences between verb classes in individual languages 
(Mitamura, 1989). 
4 An Example: Two Semantic Structures 
In this section we will show the differences in analyzing a single sentence of English ac- 
cording to each of the two methods of semantic analysis. This sentence has been extracted 
from a journalistic text, in which it was actually a clause in a compound sentence. 
Revenues and earnings will begin to post double digit growth in 1992. 
/,From the standpoint of syntax-driven lexical semantics, the desiderata of analysis 
include preserving argument structure and producing a semantic analysis in such a manner 
that syntactic realization of predicates and their arguments will be attainable through 
linking rules. We will concentrate solely on verbs. 
Following Jackendoff (1976, 1983, 1990), we will analyze begin as a change of circum- 
stance. In other words, revenues and earnings enter tile circumstance of posting double 
digit growth. For verbs in the change-of-circumstance class, the argument that changes 
circumstance links with the grammatical function SUBJECT. The circumstance can be 
realized as a nonfinite complement. We will treat post as a conflation whose meaning is 
similar to "show by posting" and which is a member of the same verb class as show or 
display in the sense of to have an observable property. 4 The linking rules for this class 
link the property with the OBJECT grammatical function and link the object that has 
the property with the SUBJECT grammatical function. 
The LCS for the above example shows two predicates -- BEGIN and SHOW-BY- 
POSTING. Notice that we have avoided the habitual thematic role names for their argu- 
ments and adjuncts. Linking to grammatical functions is shown by markers in parentheses. 
BEGI| 
thing-that-changes-circumstance: revenues and earnings (SUB J) 
nee-circumstance: SHOW-B¥-POSTIIG (KCONP) 
thing-that-has-a-property: revenues and earnings (SUB J) 
property: double digit gro.th (0BJ) 
time: 1992 (Adjunct) 
/,From the standpoint of ontology-driven lexical semantics, the desiderata of analysis 
include a) independence of the syntax of representation from the syntax of the source text, 
b) explicitly representing all implied meanings which are referred to in the source text - 
either lexically or through deixis or ellipsis or other means, c) keeping the set of atomic 
4 A conflatlon is a combination of two LCSs, A and B, into a new LCS, C. The type of conflation in this 
exmnple is that C has the meaning of "B by A-lug" and follows the linking rules and subcategorization 
patterns of B. More on the phenomenon of conflatlon see in Jackendoff (1990), B. Levln and Rapoport 
(1988), Talmy (1975, 1985). 
16 
meaning representation elements as low as possible (this latter goal does not imply a desir¢ 
to postulate a small fixed set of meaning primitives). An ontology-driven lexical-semantic 
representation for the above example can look as follows: ~ 
clause 
head: increase-1 
aspect: 
phase: begin 
duration: prolonged 
iteration: single 
time relative: ''after time of speech'' 
absolute: ''during 1992'' 
increase-1 
theme: ''revenues and earnings'' 
rate: ''greater than 10~ and less than 100~'' 
relation-1 
type: temporal-before 
. ?: . ,. 
from: ''time of speech'' .-:¢" , ,,, ." '..:!~'- 
to 1992 . -. ~ ........ ,,:.....~ ~.- -. 
Note that the verbal meanings are factored out rote propos!t!oa_~l ~$~gai~!~/~i;; ( ' 
clause head and the aspectual and temporal meanings. Note also that the ClaUse head" is- 
increase which means that the meanings of begin and post got i~a~rpo%ated i~to othe¢ _ 
representations -- the former gave rise to the value beg:i.a~f thei:'~pectga~"ir~a~gi~i'Of ;i "~ . 
the clause. The latter was understood as a functional verb whose only perpoae:w~::t0~}: . 
allow the English realization of the concept of increasing in a nominalized f~hig~:~~!~':!~ i:~{ 
post is treated as a collocant of growth (see e.g., Mel'~uk, 1981, "Nirenbu~g et~ g~!'i~\[~~!~i~:; :;:: 
Smadja and McKeown, 1989 for various treatments of colloe0.ti0ns): Proper~i~e~ ¢O~d to ' : 
describe events and objects (such as :i.ncrease) are predefined in an on\[ologi~al g!0m,a~ia ..... 
model. Phrases in double quotes are inserted as placeholders, ~ince for the ~!~rpos,~s 0fief::-}.: 
this comparison their further analysis is immaterial. • :.?::::!o.~ .. 
5 Conclusion 
Our approach to lexical semantics combines the benefits of both syntax-driven and ontology- 
driven semantics. We have been able to use both types of knowledge in several integrated 
applications. We plan to use this experience to formulate a distinct lexical-semantic the- 
ory based on this work. As a preparatory step toward formulating such a theory we have 
suggested the structure of ontological domain models, lexicons for particular languages 
and a text meaning representation language (see Carlson and Nirenburg, 1990; Meyer et 
hi., 1990, Nirenburg and Defrise, 1991a, b). 

References 
\[1\] Brachman, R. 1983. What IS-A and Isn't: An Analysis of Taxonomic Links in 
Semantic Networks. Computer, 16(10), October, 30-36. 
Sin this representation we didn't try to follow any particular knowledge representation formalism. 
J 
\[2\] Bresnan, J. and J. Kanerva. 1989. Locative Inversion in Chichewa: A Case Study 
of Factorization in Grammar. Linguistic Inquiry, 1-50. 
\[3\] Carlson, L. and S. Nirenburg. 1990. World Modeling for NLP. CMU CMT Technical 
Report 90-121. 
\[4\] Dahlgren, K. and J. McDowell. 1989. Knowledge Representation for Commonsense 
Reasoning with Text. Computational Linguistics, 15, 149-170. 
\[5\] Fodor, J. A. 1975. The Language of Thought. New York: Crowell. 
\[6\] Grimshaw, J. 1990. Argument Structure. Cambridge, MA: MIT Press. 
\[7\] Hirst, G. 1987. Semantic Interpretation and Resolution of Ambiguity. Cam- 
bridge University Press. 
\[8\] Hobbs, J. and R. Moore (eds.) 1985. Formal Theories of tile Commonsense 
World. Norwood, N J: Ablex. 
\[9\] Horty, J. F., Thomason, R. H., and Touretzky, D. S. (1990) A skeptical theory of 
inheritance in nonmonotonic semantic nets. Artificial Intelligence 42(2-3), 311-348. 
\[10\] Jackendoff, R. 1976. Toward an Explanatory Semantic Representation. Linguistic 
Inquiry, 89-150. 
\[11\] Jackendoff, R. 1983. Semantics and Cognition. Cambridge, MA: MIT Press. 
\[12\] Jackendoff, R. 1990. Semantic Structures. Cambridge, MA: MIT Press. 
\[13\] Lenat, D. and R. Guha. Building Large Knowledge-Based Systems. Reading, 
MA: Addison-Wesley. 
\[14\] Lehnert, W. and M. Ringle (eds.) 1982. Strategies for Natural Language Pro- 
cessing. Hillsdale, N J: Lawrence Erlbaum. 
\[15\] Levin, B. 1989. English Verbal Diathesis. Lexicon Project Working Papers 32. 
Center for Cognitive Science, MIT. 
\[16\] Levin, B. and M. Rappaport Hovav. 1990. Wiping the Slate Clean: A Lexical Se- 
mantic Exploration. ms. Northwestern University and Bar Ilan University. 
\[17\] Levin, B. and T. Rapoport. 1988. Lexical Subordination. Proceedings of the 24th 
Annual Meeting of tile Chicago Linguistic Society. 
\[18\] Levin, L. 1986. Operations on Lexical Forms: Unaccusative Rules in Ger- 
manic Languages. Ph.D. Dissertation. MIT. 
\[19\] Levin, L. 1987. Toward a Linking Theory of Relation Changing Rules in LFG. CSLI 
Report No. CSLI-87-115, November 1987, Center for the Study of Language 
and Information, Stanford, California. 
\[20\] Mahmoud, A.T. 1989. A Comparative Study of Middle and Inchoative Alternations 
in Arabic and English. Ph.D. Dissertation. University of Pittsburgh. 
Mel'6.uk, I.A. 1981. Meaning-Text Models: A Recent Trend in SOviet Linguistics! 'i 
The Annual Review of Anthropology. 
Meyer, I., B. Onyshkevych and L. Carlson. 1990. ~Lexicographic Principles amd 
Design for Knowledge-Based Machine Translation. CMU CMT Technical Report 
90-118. 
Mitamura, T. 1989. The Hierarchical Organization of Predicate Framesf0r 
Interpretive Mapping in Natural Language Processing. Ph.D. Dissertation. 
University of Pittsburgh. 
Nirenburg, S., E. Nyberg, R. McCardell, S. Huffmann, E. Kenschaft, and I. Niren, 
burg. 1988. DIOGENES-88. Technical Report CMU-CMT-88-107. Center for Ma- 
chine Translation, Carnegie Mellon University, Pittsburgh, PA. . ., %iii:~i~. ~i~ !~ ~i 
Nirenburg, S. and C. Defrise. 1991a. Application-Oriented Computational Sem~n~::~i~'i ~!'~! ' 
tics. In: R. Johnson and M. Rosner (eds.) Formal Semantics and Computa- ~ 
tional Linguistics. Cambridge University Press (in press). 
Nirenburg, S. and C. Defrise. 1991b. Aspects of Text Meaning. In: J, PusteJ0ysky : 
(ed.). Semantics and the Lexicon. Dordrecht, Holland: Kluwer. 
Onyshkevych, B. and S. Nirenburg. 1991. Lexicon, Ontology and Text Meaning\]"~; : i, 
This volume. 
Schank, R. 1973. Identification of Conceptualizations Underlying Natural Language. 
In: R. Schank and K. Colby (eds.), Computer Models of Language and 
Thought. San Francisco: Freeman. ~: ..... ~ : " ~'i ~ 
Smadja, F. and K. McKeown. 1990. Automatically Extracting and Representing 
Collocations for Language Generation. Proceedings of the 28th Annual Meeting of 
the ACL. 
Sown, J. 1984.Conceptual Structures: Information PRocessing in Mind and 
Machine. Reading, MA: Addison-Wesley. 
Talmy, L. 1975. Semantics and Syntax of Motion. In: J.P.Kimball, ed. Syntax and 
Semantics, 4. New York: Academic Press. 181-238. 
Talmy, L. 1985. Lexicalization Patterns: Semantic Structure in Lexical Forms. In: 
T. Shopen (ed.) Language Typology and Syntactic Description, vol. 3. Cam- 
bridge University Press. 57-149. 
Wilks, Y. 1975. A Preferential, Pattern-Seeking Semantics for Natural Language 
Inference. Artificial Intelligence, 6, 144-147. 
Woods, W. 1975. What's in a Link: Foundations of Semantic Networks. In: 
D.Bobrow and A. Collins (eds.) Representation and Understanding: Stud- 
ies in Cognitive Science. New York: Academic Press. 
