Conventional Metaphor and the Lexicon 
James H. Martin 
Computer Science Department and 
Institute of Cognitive Science 
University of Colorado, 
Boulder, CO 
80309-0430 
martin @ cs. colorado, edu 
Abstract 
Metaphor and other forms of non-literal language are essential parts of language 
which have direct bearing on theories of lexical semantics. Neither narrow theories 
of lexical semantics, nor theories relying solely on world knowledge are sufficient to 
account for our ability to generate and interpret non-literal language. This paper 
presents an emerging approach that may provide such an account. This approach is 
based on systematic representations that capture non-literal language conventions, 
and mechanisms that can dynamically understand and learn new uses as they are 
encountered. 
1 Introduction 
In traditional approaches to Lexical Semantics a whole host of phenomena have been 
largely banished from consideration. Among the unwelcome phenomena have been irony, 
sarcasm, metaphor, metonymy, indirect requests, and productive idioms. These phenom- 
ena have been relegated to various areas of pragmatics, context, and world knowledge. 
Implicit in this banishment is the assertion that the mechanisms and knowledge necessary 
to account for these uses are fundamentally different from those used in ordinary semantic 
interpretation. Moreover, this additional knowledge is generally held to be that of a more 
general common-sense conceptual kind. 
This point of view has a number of useful properties. Firstly, it severely limits the range 
of uses that need to be accounted for by lexicM semantics. Therefore, the sheer amount 
of work faced by the lexical semanticist can be reduced. The job of lexical semantics is 
essentially finished once the semantics of a word's literal meaning have been adequately 
represented and tied to the observed syntactic behavior of the word. 
The second important property (really just a different view of the first) is the notion 
that a lexicon with a relatively small fixed number of senses per word can still hope to 
account for the much larger number of senses actually encountered in practice. This follows 
from the fact that the range of senses is tied to the amount of real world knowledge in the 
knowledge-base, and the ability of the extra-linguistic mechanisms to make use of this 
knowledge. It is this second property that is most relevant to approaches to computational 
lexical semantics. This state of affairs can be summarized as follows. 
• Lexical semantics is limited to individual lexical items paired with literal meanings. 
• Non-literal uses are dealt with by mechanisms that are not required for otherwise 
"normal" semantic processing. (Among the methods that have been used are ana- 
logical matchers, planners, and spreading activation.) 
56 
* Performance (the number of senses that can be accounted for) is dependent on the 
amount of general world knowledge. 
Unfortunately, as attractive as this formulation may be, a number of serious problems 
call it into question. These problems indicate that neither a narrowly construed version 
of lexical semantics, nor large amounts of world knowledge, nor a combination of the two, 
can account for the linguistic competence required to deal with non-literal language. The 
following two sections will outline two problems with the traditional formulation and its 
computational counterparts. 
2 Constraints from Psycholinguistic Research 
While it is difficult to apply results from psycholinguistics to semantic theories in a direct 
fashion, these results can nevertheless pose useful rough constraints. The results that are 
of interest here stem from research on the relative difficulty of understanding what has 
been called literal language versus various kinds of metaphorical, idiomatic, and indirect 
language \[Gerrig, 1989, Gibbs, 1984, Gibbs, 1989, Ortony et al., 1978\]. 
The basic result that will be used is that the time needed to process various kinds of 
non-literal language does not differ significantly from the time taken to interpret direct 
language in the appropriate context. Specifically, there is no experimental evidence to 
indicate that there is a radical difference between the time taken to interpret metaphorical 
language and that taken to interpret direct literal language. This constraint has been 
referred to as the total time constraint \[Gerrig, 1989\]. 
This rough equivalence of time to process must be taken into account by any realistic 
approach to semantic interpretation. While the empirical result of equivalent time to 
process does not necessarily imply that similar mechanisms are at work, in the absence 
of more fine-grained empirical results indicating that fundamentally different processes 
are at work, it seems reasonable to assume that the mechanisms will be similar. At 
the very least it must be shown that if different processes are at work then they should 
have roughly comparable computational complexity. The next section will show that the 
proposed methods for metaphor do not meet this criteria. 
3 Problem-Solving Approaches 
The major approaches to dealing with the interpretation of non-literal language can all be 
characterized as "problem-solving" approaches. These include plan-based and analogy- 
based methods that attempt to find or infer the most likely meaning for a non-literal use 
when it is encountered \[Carbonell, 1981, DeJong and Waltz, 1983, Fass, 1988, Gentner 
et al., 1988, Indurkhya, 1987\]. In the case of metaphor, these approaches assert that 
metaphors arise from an underlying conceptual similarity or analogy between the concepts 
representing the literal meaning of the words and the concepts underlying the ultimate 
meaning of the utterance. They make no use of knowledge about the conventions of 
the language. The task of interpreting metaphoric language is seen as a special purpose 
problem-solving task requiring access to knowledge and inference techniques that are not 
otherwise a part of the normal language processing faculties. 
Note that the computational costs of these analogy mechanisms are radically higher 
than those posed for direct non-metaphorical language. While the details of each approach 
57 
differ, they are all fundamentally based on a two stage model where the literal meaning 
of the sentence is computed and judged to be ill-formed, and then an analogy system is 
employed to search for an appropriate target meaning. The major problem with these 
approaches is intractability. The proposed methods are for most practical purposes com- 
putationally intractable given any realistic representation of the domain. The problem is 
particularly acute when the system in intended to be used in some kind of user interface 
role where response time is critical. Moreover, such methods are directly contradicted by 
the empirical psycholinguistic evidence described above. 
4 An Emerging View 
These psycholinguistic and computational results have led to a surge of computational 
research in the area of semantic interpretation of non-literal language in recent years. This 
research has resulted in a number of systems that can handle one or more of the following 
phenomena: metaphor, metonymy, indirect requests, and somewhat productive phrasal 
idioms. \[Fass, 1988, Hinkelman, 1989, Martin, 1990, Pustejovsky, 1991, Zernik, 1987\] A 
careful analysis of these disparate efforts reveals that a common approach to the semantic 
interpretation of non-literal language may be emerging. At the heart of this approach 
is a concern with the tension between the conventionality of many non-literal uses, and 
the ability to understand and generate new uses that go beyond static conventions. This 
concern has generally led to a two-part approach involving systematic representations 
that capture conventions, and mechanisms that can dynamically understand and learn 
new uses as they are encountered. 
The work of Pustejovsky \[Pustejovsky, 1991\] is related in its attempt to reduce the 
size and complexity of individual lexical entries. The difference is that while the above 
efforts try to achieve a kind of lexical generativity through the use of extra-lexical and 
extra-grammatical language conventions, Pustejovsky's work can be seen as attempting 
to capture this generativity from within a lexicai knowledge base. 
Consider some of the following examples. 
(1) How can I kill a process? 
(2) My emacs just died. 
(3) Can you open the door? 
(4) The buses are going on strike. 
Each of these examples makes use of a completely conventional, yet non-literal, way of 
conveying the intended meaning. The use of the word kill in (1) is consistent with the 
conventional and productive metaphor that allows computer processes to viewed as living- 
things. This productive use also motivates the use of die in (2). Example (3) makes use of 
a conventional indirect way to make a request in English. In this case, asking about ones 
ability to perform some task by using Can you, indicates the speakers desire to have that 
task performed by the hearer. Finally, the last example demonstrates the widespread use 
of metonymy in English. In this example, the word bus is used to refer to tile distinct, 
yet obviously related, busdrivers. 
The unifying theme, in all of these examples, is the use of conventional yet productive 
mechanisms to produce these sentences. The remainder of this paper will introduce a 
unifying approach to these and similar phenomena that avoids some of the problems 
inherent in current approaches. 
58 
The following sections provide an analysis that sketches out the nature of this emerg- 
ing approach. In each of these sections I will first attempt to abstractly characterize the 
approach and then give a few details of how the approach has been instantiated in the MI- 
DAS \[Martin, 1990\] system. To be specific, Section 5 describes the nature of the knowledge 
that these systems use, and introduces a new knowledge construct called a Conventional 
Conceptual Mapping, that abstractly accounts for them all. Section 7 briefly describes 
how this approach has been applied to metaphor in MIDAS. 
5 Conventional Conceptual Mappings 
The fundamental construct that all these systems use is a knowledge structure that allows 
the system, under certain conditions, to map from one set of concepts to another set of 
concepts. In the case of MIDAS, these associations are used to represent the source and 
target concepts of conventional metaphors; in \[Fass, 1988\] they are used to capture the 
two component parts of metonymies; while in \[Hinkelman, 1989\] they are used to capture 
knowledge about indirect speech acts. Consider the following examples. 
(5) How can I get out of Emacs? 
(6) Can you speak Spanish? 
(7) Plato is on the top shelf. 
In Example (5), MIDAS uses a set of mappings from the domain of enclosures to the 
domain of computer processes to process this question. In (6), Hinkelman's system may 
be viewed as making use of structured associations that link features and concepts asso- 
ciated with this surface form, to a concept representing a request to perform the intended 
meaning. In this example, this association links questions pertaining to the ability to 
perform an action to requests for that action to be performed. Finally, in (7), Meta5 
makes use of mappings that represent conventional metonymies. In this case, the system 
makes use of a mapping that allows certain kinds of products to be referred to through 
their producer. In this case, the works of an author can be referred to by referring to the 
author. 
There are a number of common characteristics that are shared among these efforts. 
The most important of which is the idea that these associations may be represented 
at a variety of levels ranging from linguistic features to concepts. This has a number 
of important ramifications. The first is that it allows these mappings to capture the 
generalizations at the right level of conceptualization. For example, it is not necessary in 
(7) to have a new word sense for the word Plato. Rather, the system merely needs to know 
the fact that Plato was an author to use the proposed metonymic map linking producers to 
products. It is also important to note that while these conventional associations interact 
with world knowledge, they can not be conflated with it. 
The second important characteristic shared among these efforts is the idea that these 
structured associations have explicitly represented constraints on their use. When the 
constraints on the use of the association call be satisfied in context, then the association 
can be used to map from one set of concepts to another. 
I call a mapping that has these characteristics a Conventional Conceptual Mapping 
(CCM). More specifically, a CCM is a structured mapping that has the following charac- 
teristics. 
• It allows disparate sets of concepts to be associated in a structured manner. 
59 
• It can be represented at varying degrees of abstraction. 
• It can have constraints on its use explicitly represented. 
Not surprisingly, none of the systems mentioned above fits this description precisely. 
Most deal with only one of the phenomena, while ignoring the others. For example, 
Hinkelman's system deals only with speech act level conventions, while Martin's MIDAS 
system deals only with conventional metaphors. While Fass's META5 system does handle 
metaphor and metonymy, only metonymy is handled in a way that uses ccM's. 
The following section will give a high level discussion of how the this approach can 
be applied to the problem of metaphor. Section 77 will then detail the current state of an 
implementation of this approach. 
6 A Sample Metaphor Analysis 
To make this discussion more concrete, consider some of these issues in terms of the 
following examples. 
(s) \] 
(9) I 
(10) 
(11) (12) 
(13) 
(14) 
(15) 
(16) 
(17) 
(18) 
came to the conclusion that Mike was right. 
reached the conclusion that Mike was right. 
I arrived at the conclusion that Mike was right. 
I was led to the conclusion that Mike was right. 
I was dragged kicking and screaming to the conclusion that Mike was right. 
That Mike was right came to me. 
It came to me that Mike was right. 
It hit me that Mike was right. 
It struck me that Mike was right. 
Mike struck me as being right. 
Mike came across as being right. 
Each of these examples expresses the idea that someone is changing state with respect 
to some belief. More specifically, they express the idea that the believer now believes 
some proposition that had not been previously held. They all express this notion through 
the use of a spatial metaphor that has the following entailments: beliefs are external 
objects with locations, believers have locations, shared location between a believer and a 
proposition indicates belief, and finally movement to a shared location indicates a state- 
change resulting in belief. \[Barnden, 1989\] 
Given this general metaphor, we can narrow the discussion further by isolating on the 
behavior of a single lexical item. 
(19) The idea that Mike was smart came to me. 
(20) It came to me that Mike was smart. 
(21) I came to the conclusion that Mike was smart. 
(22) ?I came to the idea that Mike was smart. 
(23) I came upon the idea that Mike was smart. 
These examples illustrate at least two broad semantic senses and four different syn- 
tactic environments for the word come. This leads to the big question: 
60 
Do these meanings, and these syntactic environments, have to be listed in the 
lexical entry for come, or can these facts be motivated by other independent 
language conventions?~ 
The following discussion will raise a series of more specific issues surrounding this 
question and provide some flavor of the research program we are pursuing. 
Consider first the two broad senses that are evident here. In the first sense, the 
new proposition is the mover, changing its location to the location of the believer. One 
implication of this sense is that the believer did not consciously expend any effort to 
produce this belief. This sense is illustrated in (19) and (20). Examples (21) through (23) 
illustrate the second sense that has the believer actively moving to the new proposition. 
This sense has more of an implication that the believer is taking an active part in the 
process. 
We are currently investigating the proposition that these two senses result from the 
interaction of a single broad spatial metaphor, the core spatial semantics of the word 
come, and facts about the situation. It is, therefore, not necessary to have either two 
lexical senses or two conventional metaphors. 
Now consider the syntactic variation exhibited in examples (19) and (20). In both of 
these examples the proposition is in motion. In (19) it appears in the subject position, 
while in (20) it is extraposed. Our current hypothesis is that it is not necessary to list in 
the lexical entry for come the fact that it permits extraposition. Rather this flows from 
the general requirements of the extraposition construction combined with the semantics 
of target concept of this metaphor. 
Next consider the apparent well-formedness variation between (21) and (22). The 
use of come to with conclusion is clearly preferred. Again the metaphorically focussed 
approach advocated here would say that the metaphorical semantics of the word conclusion 
lead it to naturally fit into the spatial metaphor governing the use of come. 
Finally, consider the difference between example (23) and (21). This example seems 
to imply that, while the believer was in motion towards the proposition, the discovery 
was in fact something of a chance occurrence. But of course, this would also be implied 
by a literal use of come upon. Therefore, this difference is attributable to a difference in 
the spatial semantics of come. The question remains as to whether it is necessary to list 
this fact in the lexical entry for come. 
7 Overview of MIDAS 
This section provides a brief overview of our current computational realization of this 
approach to metaphor. In particular, it introduces the following issues. 
Representation: The explicit representation of the conventional metaphors in a 
language in the form of explicit associations between concepts. 
Interpretation: The correct and efficient application of this metaphoric knowledge 
to the interpretation of metaphoric language. 
Under this view, the proper way to approach the study of metaphor is to study the 
underlying details of individual metaphors and systems of metaphors in the language. 
This approach follows on the metaphor work of Lakoff and Johnson \[Lakoff and Johnson, 
1980\] and the computational approaches to metaphor described in \[Jacobs, 1985, Norvig, 
1987\]. 
61 
This approach has been embodied in MIDAS (Metaphor Interpretation, Denotation, 
and Acquisition System) \[Martin, 1990\] MIDAS is a set of computer programs that can 
be used to perform the following tasks: explicitly represent knowledge about conven- 
tional metaphors, apply this knowledge to interpret metaphoric language, and learn new 
metaphors as they are encountered. 
7.1 Knowledge Representation 
Consider the following simple example of a conventional UNIX metaphor. The metaphori- 
cal use of the word in reflects a systematic metaphorical structuring of computer processes 
as enclosures. 
(24) I am in Emacs. 
Metaphors like this may be said to consist of the following component concepts: a 
source component, a target component, and a set of conventional associations from the 
source to target. The target consists of tile concepts to which the words are actually 
referring. The source refers to the concepts in terms of which the intended target concepts 
are being viewed. In this example, the target concepts are those representing the state of 
currently using a computer process. The source concepts are those that involve the state 
of being contained within some enclosure. 
The approach taken here is to explicitly represent conventional metaphors as sets of 
associations between source and target concepts. The metaphor specifies how the source 
concepts reflected in the surface language correspond to various target concepts. In this 
case, the metaphor consists of component associations that specify that the state of being 
enclosed represents the idea of currently using the editor, where the user plays the role 
of the enclosed thing, and the Emacs process plays the role of the enclosure. These 
associations also serve to delimit the particular parts of the various source and target 
domains that are relevant to particular conventional metaphors. 
Note also that these source-target associations are represented at the conceptual and 
not the lexical level. Any single lexical item or expression that can be construed as 
referring to the source concept of a known metaphor, may invoke that metaphor. In 
this example, the source component of the metaphor is attached to the concept of being 
enclosed, not to the lexical item in. 
These sets of metaphoric associations, along with the concepts that comprise the source 
and target domains, are represented using the KODIAK \[Wilensky, 1986\] representation 
language. KODIAK is an extended semantic network language in the tradition of KL-ONE 
\[Brachman and Schmolze, 1985\] and its variants. 
These sets of metaphoric associations representing conventional metaphors are full- 
fledged KODIAK concepts. As such, they can be related to other concepts and arranged in 
abstraction hierarchies using the inheritance mechanisms provided by KODIAK. The hier- 
archical organization of conventional metaphoric knowledge is the primary means used to 
capture the regularities exhibited by the system of metaphors in the language. Specifically, 
KODIAK is used to represent specialized domain specific metaphors, pervasive high-level 
metaphors, and the systems of relations among related metaphors. 
62 
7.2 Interpretation 
The interpretation process in MIDAS is basically one that views a given input sentence 
as providing a set of constraints on possible interpretations. MIDAS checks the input 
constraints against all the possible interpretations that can be conventionally associated 
with the input. Interpretations that are coherent with the constraints are returned. The 
possible conventional interpretations may include direct non-metaphoric interpretations, 
as well as all the conventional metaphors that are invoked by the input. 
Consider the details of the following shortened trace. In this example, MIDAS must 
find a coherent interpretation for this use of enter. MIDAS finds, and attempts to apply, all 
the conventional metaphorical and non-metaphorical concepts associated directly with, or 
inherited by, this concept. In this case, it finds that the only conventional interpretation 
that is consistent with the input is the one that results from the application of the known 
Enter-Lisp metaphor. 
| 
> (do-sentence) 
Interpreting sentence : 
How can I enter lisp? 
Interpreting concreted input. 
(A EnteringSO (T Entering) 
(enterer50 (T enterer) (A I203 (~ I))) 
(entered50 (~ entered) (A Lisp58 (~ Lisp)))) 
A parser first produces a syntactic analysis and a preliminary semantic representation 
of the input. At this point in the analysis, uc calls upon MIDAS to begin a deeper analysis 
of this initial representation. 
Failed interpretation: Entering50 as Entering. 
Failed interpretation: EnteringSO as Enter-Association. 
Valid known metaphorical interpretation: EnteringSO as Enter-Lisp. 
The case structure of this preliminary representation is checked against the semantic 
constraints of all the interpretations conventionally associated with the Entering concept. 
In this case, MIDAS finds that the direct interpretation and one of the other possible 
entering metaphors can be rejected before the appropriate Enter-Lisp metaphor is found. 
It is important to realize that the order of tile search performed here is arbitrary. 
MIDAS is exhaustively finding all conventional interpretations that are consistent with the 
input. The determination of consistency for any given interpretation is independent of the 
consistency of any of the other possible interpretations. In particular, the well-formedness 
of a direct, or literal, interpretation has no effect on whether or not a metaphorical in- 
terpretation will be found. It follows from this that the order of the search through the 
possible interpretations has no effect on which interpretations will ultimately be produced. 
53 
Applying conventional metaphor Enter-Lisp. 
(A Enter-Lisp (T Container-Metaphor Metaphor-Schema) 
(enter-lisp-res enter-res -~ lisp-invoke-result) 
(lisp-enterer enterer -~ lisp-invoker) 
(entered-lisp entered--+ lisp-invoked) 
(enter-lisp-map Entering -~ Invoke-Lisp)) 
Mapping input concept EnteringSO 
to concept Invoke-Lisp30 
Mapping input role entererSO with filler I203 
to target role lisp-invoker30 
Mapping input role enteredSO with filler LispS8 
to target role lisp-invoked30 
Yielding interpretation: 
(A Invoke-Lisp30 (~ Invoke-Lisp) 
(lisp-invoked30 (~ lisp-invoked) (A Lisp58 (~ Lisp))) 
(lisp-invoker30 (T lisp-invoker) (A I203 (T I)))) 
MIDAS then begins the process of mapping from the given source concepts to the 
appropriate target concepts based on the constraints imposed by the metaphor. The 
mapping process, called metaphoric unviewing, creates a new instance of the metaphor 
itself along with the attendant source and target concepts. In this example, the source 
concept of Entering is mapped to the target concept Invoke-Lisp as specified by the 
metaphor. 
Final Representation: 
(A How-q207 (T How-q) 
(topic206 (T topic) 
(A Invoke-Lisp30 (T Invoke-Lisp) 
(lisp-invoked30 (T lisp-invoked) (A Lisp58 (T Lisp))) 
(lisp-invoker30 (~ lisp-invoker) (A I203 (T I)))))) 
8 Conclusions 
The competence required for the frequent and conventional use of idiom, metaphor, 
metonymy, irony, and sarcasm can not be accounted for by either a narrow theory of 
lexicM semantics or any theory relying on world knowledge alone. Moreover, empiricM 
evidence suggests that these phenomena constitute an integral part of our linguistic com- 
petence. It is therefore clear that our semantic theories must be broadened beyond their 
current scope to adequately deal with the challenges posed by these phenomena. The 
work outlined in this paper is a start in this direction. 
64 

References 
\[Barnden, 1989\] John Barnden. Belief, metaphorically speaking. I In Proceedings of the 
First International Conference on Principles of Knowledge Representation, San Mateo, 
CA, 1989. Morgan Kaufmann. 
\[Brachman and Schmolze, 1985\] Ronald J. Brachman and James Schmolze. An overview 
of the kl-one knowledge representation system. Cognitive Science, 9:346-370, 1985. 
\[Carbonell, 1981\] Jaime Carbonell. Invariance hierarchies in metaphor interpretation. 
In Proceedings of the Third Meeting of the Cognitive Science Society., pages 292-295. 
Cognitive Science Society, August 1981. 
\[DeJong and Waltz, 1983\] Gerald F. DeJong and David L. Waltz. Understanding novel 
language. Computers and Mathematics with Applications, 9, 1983. 
\[Fass, 1988\] Dan Fass. Collative Semantics: A Semantics for Natural Language. PhD 
thesis, New Mexico State University, Las Cruces, New Mexico, 1988. CRL Report No. 
MCCS-88-118. 
\[Gentner et al., 1988\] D. Gentner, B. Falkenhainer, and J. Skorstad. Viewing metaphor as 
analogy. In D.H. Helman, editor, Analogical Reasoning. Kluwer Academic Publishers, 
1988. 
\[Gerrig, 1989\] Richard J. Gerrig. Empirical constraints on computational theories of 
metaphor: Comments on indurkhya. Cognitive Science, 13(2):235-241, 1989. 
\[Gibbs, 1984\] Raymond W. Gibbs. Literal meaning and psychological theory. Cognitive 
Science, 8:275-304, 1984. 
\[Gibbs, 1989\] Raymond W. Gibbs. Understanding and literal meaning. Cognitive Science, 
13(2):243-251, 1989. 
\[Hinkelman, 1989\] Elizabeth Hinkelman. Linguistic and Pragmatic Constraints on Utter- 
ance Interpretation. PhD thesis, University of Rochester, Rochester, NY, 1989. Tech- 
nical Report No. UR-CS-288. 
\[Indurkhya, 1987\] Bipin Indurkhya. Approximate semantic transference: A computa- 
tional theory of metaphors and analogy. Cognitive Science, 11:445-480, 1987. 
\[Jacobs, 1985\] Paul S. Jacobs. A Knowledge-Based Approach to Language Production. 
PhD thesis, University of California, Berkeley, Computer Science Department, Berkeley, 
CA, 1985. Report No. UCB/CSD 86/254. 
\[Lakoff and Johnson, 1980\] George Lakoff and Mark Johnson. Metaphors We Live By. 
University of Chicago Press, Chicago, Illinois, 1980. 
\[Martin, 1990\] James II. Martin. A Computational Model of Metaphor Interpretation. 
Academic Press, Cambridge, MA, 1990. 
\[Norvig, 1987\] Peter Norvig. A Unified Theory of Inference for Text Understanding. PhD 
thesis, University of California, Berkeley, Computer Science Department, Berkeley, CA, 
1987. Report No. UCB/CSD 87-339. 
\[Ortony et ai., 1978\] A. Ortony, D. Schallert, R. Reynolds, and S. Antos. Interpreting 
metaphors and idioms: Some effects of context on comprehension. Jounai of Verbal 
Learning and Verbal Behavior, 17:465-477, 1978. 
\[Pustejovsky, 1991\] James Pustejovsky. Towards a generative lexicon. Computational 
Linguistics, 17(1), 1991. 
\[Wilensky, 1986\] Robert Wilensky. Some problems and proposals for knowledge represen- 
tation. Technical Report UCB/CSD 86/294, University of California, Berkeley, Com- 
puter Science Division, May 1986. 
\[Zernik, 1987\] Uri Zernik. Strategies in Language Acquisition: Learning Phrases from 
Examples in Context. PhD thesis, University of California, Los Angeles, Computer 
Science Department, Los Angeles, CA, 1987. 
