Viewing Word Expert Parsing as Linguistic Theory

1. INTRODUCTION

The Word Expert Parser CWEP) is a computer program that analyzes fragments of natural language text in order to extract their meaning in context. The system has been developed with particular attention paid to the wide variety of different meaning roles of words when appearing in combination with other words. The theoretical position advanced by ~~P about the nature of individual words is that words have no meaning per se, but rather, that fragments of lexical items mean something through their interrelationships. Furthermore, the character of lexical relations runs the gamut from the simple direct kno~1ledge that some word sequence represents some remembered concept, to the more analytical knowledge that particular kinds of lexical sequences often represent certain classes of conceptual notions.

· The initial research motivating the ideas described in this report was supported by grant NSG-7253 from the National Aeronautics and Space Administration to the University of Maryland. During the writing of this paper, the author was supported by a Fulbright Lectureship in Artificial Intelligence at the Universit6 de Pari s, and computer faciliti es were provided by the Institut de Recherche et Coordination Acoustique/Musique. The s~pport of NASA, the Fulbright-Hays program, and IRCAM are gratefull y acknowledged.

The evolution of this perspective started with the observation that the understanding of a particular fragment of text depends fundamental ly on the disambiguation of the indi vidual word:> composing it. Knowing the contextual meanings of the words is tantamount to understanding the meaning of the overall fragment. Another war of saying the same thing [Rieger, 1977] is that language interpretation can be ultimately viewed as a process of word sense discriminati on. Unfortunately, this perspective does not eliminate the classic problems of deciding the nature of a distinct word sense, t·he difference bet1veen different usages, word senses, and idioms, and so forth. The solution to these problems comes in realizing that the process of understanding the meaning of words in context does not require reference to those notions at all. The design of a parsing procedure based on determining the meaning roles of individual words in context has led to the )rthogonal linguistic notions that are the subject of this paper.

The organization of WEP is founded on the belief that the grouping together of words to form meaningful sequences is an active process which succeeds only because of highly idiosyncrati c application of lexical knowledge. That is, we fragment text and understand the meaning of the pieces because we know how the particular words involved interact with each other.

Sometimes sequences of two or more words interact together to such an extent that they seem to behave as a single lexical i tern. Linguists have labelled such sequences idioms. The notion to which this definition gives rise, however, causes several problems for linguistic theory. First of all, rarely does such a sequence hold together so tightly that it can be truly treated theoretically as a single lexical item. Secondly, rarely does such a sequence have a unique meaning. More often than not, the meaning of an idiomatic expression must be determined by disambiguation. The sequence must be analyzed in context and be treated by comprehension processes as being either (a) a cohesive whole with idiosyncratic meaning, or (b) a sequence having meaning through less specific language knowledge. There is no a priori way of knowing the meaning of the sequence to be the one or the other.

The notion of idiom falls at one end of a spectrum, an idealized end that I claim does not exist. Lexical sequences can be more or less idiomatic, in the sense that the process interactions constituting the understanding of them includes greater or fewer idiosyncratic interactions. The WEP way of looking at the mos t idiomatic sequences is that the special interactions among the participating words take priority over any other potential interactions involving those words. The disambiguation of idiomatic expressions, i.e., the understanding of the sequences as either idioms or non-idioms (to use the popular distinction), generally requires other process interactions besides the strictly word-specific ones. The understanding of an idiom thus differs insignificantly, from the perspective of WEP theory, from comprehension of any other kind (according to whatever classification scheme) of lexical sequence.

The notion that all fragments of language are more or less idiomatic, while radical in some linguistic quarters, has been previous! y suggested. In his introductory textbook, Aspects of Language, Dwight Bolinger asks "whether everything we say may be in some degree idiomatic - - that is, whether there are affinities among words that continue to reflect the attachments the words had when we learned them, within larger groups" (Bolinger, 1975). After working wi thin what he calls "the prevailing reductionism", Bolinger began to suggest a positive answer to his pedagogical question, choosing to take "an idiomatic rather than an analytical view" (Bolinger, 1979) of langu!ge. The contribution of artificial intelligence in general, and of Word Expert Parsing in particular, is to develop theory from this informal view. The notion of process, and of process interaction, allows us to begin to do just that.

2. THE ~RO EXPERT PARSER

The WEP computer system maintains linguistic knowledge across a community of word-based structures called word experts, which represent the process of determining the contextual meaning and role of the individual words. A word expert must not Qe thought of as a representation for the various meanings, roles, and contributions of a word in context, but rather as a declarative representation (a network) of the process (which we shall call disambiguation) of determining these things. Certainly, it is the meaning contributions of individual lexical items that we wish to determine. Word experts are both data and process; they can be augmented, examined, and manipulated as data, yet parsing takes place through their interpretation as program by an expert evaluator, similar to the EVAL of Lisp.

The distributed parsing scheme of WEP works as follows. The IJEP reader examines a word of text and retrieves its word expert from memory. The word ~xpert starts executing, trying to determine the meaning role of its word in context, 1. e., interacting with other word experts and with higher-order system processes to acquire the appropriate contextual knowledge to make the correct inferences. Finally, all the word experts for a particular fragment of text come to mutual agreement on the meaning of the fragment, and the local distributed process terminates. Local, in the sense that as long as there remains input text, the overall parsing process continues, while the disambiguation of individual lexical sequences making up the larger text completes.

Interaction, between individuals in the world, or between distributed processes in a computer program, requires both (a) giving information and (b) receiv1ng information. In WEP, the experts exchange t1v0 kinds of information, called concept structures and control signals. Concept structures represent human concepts, such as "a book", "going fi shing", "the box of candy I gave Jeanie . for Val entine's daY in 1981", "some blue physical object", and the like. Control signals represent processing clues, such as "expect a ~lord that can begin a lexical sequence that can describe concept structure X", "send me the concept structure representing the agent of concept structure Y or a signal saying you cannot", "wait a second and yo;.~ will be sent a concept structure that will help you", and similar things. The representation and use of concepts and signals are described fully in [Small, 1980].

3. LEXICAL INTERACTIONS

I use the term 1ex"ica1 "interaction to denote the sending and receiving of control signals and concept structures by word experts in WEP. This includes interactions between individual experts, as well as those between a word expert and another kind of model process (e.g., a mechanism inferring the goals of a dialogue participant). This paper discusses lexical interactions by presenting four classes of required interaction, and then arguing for the necessity and giving examples of each. The categorization is by the kind of knowledge exchanged in the communication, and includes the following.

Idiosyncratic Interaction

Linguistic Interaction

Discourse Interaction

Logical Interaction

The least general class of lexical interactions are considered idiosyncratic since they are worct-specific and arise through simple recall memory. This type of interaction permits the understanding of idiomatic fragments. General knowledge about the syntax and semantics of some natural language gives rise to linguistic interactions, and are of course crucial to the understanding of lexical sequences not previously seen. Sometimes words interact with processes that monitor the development of an entire text (or parts thereof), or the goals of participants in discussion. These discourse interactions are often necessary for the meaningful cohesion of lexical fragments. Lastly, but certainly not least important are the logical interactions between words and the most general cognitive processes. Perceptions about the world, beliefs, inference-making skills, rote memory, and so forth, are basic to language understanding.

The classification of word fragments into categories such as "idiom", "collocation", "colligation", "noun phrase", "complement", and the like, does not make sense in WEP theory. Rather, individual words are viewed as having certain kinds and sequences of interactions with their neighbors to form meaningful pieces of text. Fragments often described as "idioms" are those that are understood principally through idiosyncratic lexical interactions. A non-idiomatic structure, diagnosed as a "noun phrase", is one that involves mostly linguistic interactions to understand. A so-called "noun-noun pair" can be thought of as a lexical sequence comprehend~d with the help of logical interactions, with recourse to common sense memory and skills.

3.1 Idiosyncratic Interaction

Since the emphasis of the WEP research effort is to construct a computer program to understand natural language, we are not qualified to make claims about just how much of human parsing involves idiosyncratic lexical interaction. Suffice it to say that no theory of language analysis can do without such a noti on, and that we have come acros s many examples, both in our own wo rk and in the linguistic literature, where it applies. Furthermore, our concepti on of idiosyncratic interaction (in conjunction with the other interaction types) , provides explanatory adequacy for many lingui stic phenomena, observed in di ver se camps.

Idioms

The comprehensi on of idioms requires both idiosyncratic lexical interactions and context-probing disambiguation (i.e., interactions between lexical processe s and higher-order memory processes) . As an example, examine the following sentence, analyzed easily by the prototype WEP system described in [Small, 1980].

(1) "The fellow throws in the towel."

Linguistic description of this sentence could take several different routes. The verb in the sentence could be seen to be either one of (a) throw, (b) throw in, or (c) throw in the towel, depending on the theoretical perspective of the lingui s ti c approach. 1-.'hile WEP does not itself (expl l e i tl yl make use of the notion of a verb, if forced to explain its behavior in these terms, I would say that WEP would consider the verb to be either (a) , (b), or (c), depending on the context.

The WEP computer system would behave as follows. The word expert for throw would send a message to the in expert, consisting of both a control signal and a concept structure. The signal would in effect tell in that throw would like to pair .up with it to form a cohesive fragment. The accompanying concept structure would tell in what kind of lexical sequence (or conceptual mapping of one) throw would expect t o find to the right of in for it to have confidence in such a pairing. Of course, the in expert has independent control of its own interactions, and could decide (based on the nature of the input) to reject any suggestions it receives. In the above example, in would l·lai t for a signal from the expert on its right (the the expert) indicating the formation of a cohesive fragment designating a towel, and then act on this information.

The concept structure sent by throw indicates that if the words to the right of in are the towel, then the words throw and in should pair up. However, the in expert has the prerogative to override this suggestion, and if query to the processes modelling the focus of di scour se attention (see the later section on discour se interactions) determined that some large towel wa s the location of some active event, would do so, transmitting a signal to throw rejecting its advice. Ordinarily, this would not happen, and the in expert would signal acceptance to throw. The throw expert would then query the discour se processes monitoring the activity context to determine if (a) someone is actually throwing a to~1el into some val ume, or (b) someone is conceding defeat to someone in a game. When this wa s determined, the understanding of the fragment (in this case a sentence) would be complete.

Col location and Col I igation

What is the nature of the advice given by one word expert to another, as from the throw expert to the in expert above, and wha t is its basis? These suggestions, transmitted as control signals and concept st ructures in WEP, are based on purely idiosyncratic criteria. While the idiomatic meaning of certain l exical sequences cannot be predicted from the parts, certain others have meanings that differ slightly from what might be inferred. These sequences are 5ometimes called collocations, and illustrate the idiosyncracy of lexical pairings and thus the basi s for certain interactions in WEP analysis.

In his recent paper, Bolinger [19791 quotes from T. F. Mitchell [1971] about the word work and its meaning in various fragments of lexical items: "Men specifical ly cement workers -- work in cement works; others of different occupation work on works of art; others again, or both, perform good works. No t only are good works performed but cement works are built and works of art produced" (italics in original l. Bolinger goes on to refl ect why "builders do not produce a building or author s invent a novel". From the vantage point of the distributed lexical actors of WEP, the se particularities form the basis of the inter-expert communicati ons we have been calling idiosyncratic interact ions.

Let us take Mitchell's example, the ambiguous word work, and see how the word experts of WEP can interact to determine its meaning in context. It should be easy to see from the discussion so far how WEP would handle the following fragment.

(2) "The Ensemble lntercontemporain performs a work."

Without knowing anything about the various entities mentioned in the sentence, it is clear that an ensemble of musicians played a musical piece. Lexical interactions between the word expert for perform and that for work make clear the meaning of the fragment. The perform expert would signal t o work its expectation of a lexical sequence denoting a conceptual entity that can be "performed", such as "a service", "a musical piece", or "a series of actions". It would then be up to work to use this (non-binding) advice to contribute to the meaning of the overall fragment in which it participates.

3.2 Linguistic Interaction

The syntax and semantics of natural language comprise the stuff of most lingui stic theory, and their use is the cornerstone of computer par s ing programs. However, WEP forces us to perceive the nature and use of this knowledge from an untraditional vantage pain~ that of the individual word and i ts active processing to form meaningful fragments of text with its neighbor s. From thi s perspective, the syntax and semantics of a language is that body of knowledge that helps us infer enough about the meaning contributions of new words to understand lexical sequences compl etel y different from ones that we have already seen. After process ing some lexical sequence, thi s linguistic knowledge provides us with certai n expectations ~bout upcoming lexical i tems and t he nature of their interactions, thus helping us to fragment those items into pieces and to infer the meaninR of the pieces from general informati on about their component parts.

Noun Phrases

How can the purely lexical WEP system require no notion of high order s tructural ph~nomena, ye t s t i ll be able t o account for them? The foll ow ing example (provided by Yorick W.ilks) illustrates the lexical interactions required to anal yze an interesting fragment of text.

(3) "Joanie washes the colorful di shes up."

The difficulty with this fragment is in determi ning that the word dishes contributes to the meaning of the fragment through interactions with the two words to its left, but that the word up contributes by association with the word washes, which precedes up by many intervening words. The reason that I am avoiding the use of traditional linguistic jargon for describing this phenomenon is the following belief: An understanding of WEP requires the viewing of language interpretation fr om t he vantage point of the individual word and its interactions. An important way to achieve this is to describe the analysis process with reference to the very noti ons (not the traditional ones) around which it i s organized.

In the analysis of the example fragment, WEP would find the referent of Joanie and then proceed as follows. The wash expert ~1ould begin executing, trying to determine its own meaning role in some lexical fragment, and at the same time, trying to provide information to other lexical agents to permit them to do the same. The meaning of wash in context depends on a number of factor s, including the natur·e of the words succeeding it, and their own actions in determining their meaning and role contributions in context. The wash expert mus t thus prepare for a number of contingencies, or different things that could happen in the text, and then. wait to see if any of them actuall y occur. If the word up appears to the right of wash, for example, the words could choose to pair up into a meaningful fragment (as in throw in above) . Under certain conditions, the word up could appear later on in the text, and still pair up with wash (as must occur for correct interpretati on of the example sentence).

What are the contextual conditions that would permit this? One of the contingencies that wash anticipates is the grouping of the words to its r ight into a meaningful fragment of their 01·m (1. e., a concept structure).. The wash expert knows that (a) the nature of this concept structure may be important for its own sense disambiguati on, and ~) that the word immediately following the meaningful fragment could pair up with it. In the jargon of WEP, one of the expert s in such a meaningful fragment reports a concept s tructure. Since the up expert does not reply to dishes with an acceptable message to continue the ongoing concept building activity, the dishes expert reports the structure. It is this report that triggers some new processing by the wash expert, namely the examination of the next word (i.e., up) .

The rest of the analysis take s place predictably~ The wash expert interact s with up as if up occurred to its immediate right in the text. The pairing up of the two words results fr om mutual accord, and the wash expert creates a concept structure to represent the meaning of the washing up of di shes. Next wash organizes the conceptual object, t.'"te co1orfu1 dishes, into the overall meaning of the sentence, and again waits for things t o happen. This time, the word expert for the peri od at the end of the sentence executes, ancl transmits an appropriate me ssage. The wash expert again executes, cleans up its business, and reports the concept structure representing the meaning, i n context, of the entire fragment (sentence) .

Passives and Relative Clauses

Sentences in the passi ve voice and those containing relative clauses are similar in being complex structural phenomena in natural language, and often suggestive of sentence-level rul es ·13 linguistic explanation. Furthermore, the understanding of such constructions by the distributed word-based approach of ~!E? may be far from evident, especially cons ider ing my claim t hat no explicit notions of structure are referenced by the computer system or used in the theory. Interpretation of textual fragment s contai ning complex syntactic structures takes place through complex patterns of lexical interacti ons emo~g the appropriate word expert s. The words t hat normal l y cue a reader about the presence of such structural relations in a fragment are the ones in WEP t hat coordinate the process of understanding them.

The analysis of a passi ve sentence i nvol ves linguistic interactions among the word expert s for the suffix en, the word by, and the other 1-:ords composing it. The following sentence has been parsed by the existing WEP system, ancl di scussed at length in [Small, 1980) .

(4) "The case was thrown out by federal court . "

The en expert begins executing before throw, and the normal attempts by the throw expert t o coordinate the analysis of the fragment in which it participates are intercepted by en. The actions of enallowthrowtopairup with out, as outlined above for throw in and wash up, but its lexical interactions to determine the nature of the object being "thrown out", and the agent doing t he "throwing" are all intercepted by the en expert, which provides throw with the correct replies t o its queries. Please refer to [Small, 1980) and [Small, 1981) for a fuller discus s i on.

Relative clauses beginning with the word who are analyzed by WEP through the interactions among the who word expert and the experts for the other words in the clause and the larger fragment containing it. The following sentence is an example of such a fragment.

(5) "The man who throws the game likes t o lose."

The who expert in this sentence has t he responsibility for interacting wi th the word expert for 1ikes to inform 1ikes about the man doing the "1 ilcing". Ordinarily, this expert would ·expect to find a meaningful lexical sequence to its lef t representing the needed concept. Ho~:eve r, the particular structure of the fragment means that who must be at the other end of the relevant linguisti c interactions of 1ikes, rather than the expert fo r the word to its immediate left, whi ch h'Oul d normally perform the needed service.

The WEP interpretati on of the exampl e fragment proceeds as follows. The word experts for the and man agree to form a meaningful sequence and construct a concept structure to represent it s meaning. The who expert begins executing, get s hold of this concept, and waits for the throw expert to start expl oring the nature of the lexical sequence on its 1 eft. In additi on, the who expert anticipates that anot he r wo rd expe r t fur ther down the 1 ine (in the example, the expert for 1il<es) will al so seek out informati on about t he sequence to its left, in exactl y the way t hat throw cloes . The who expei"t, 1 ike every word expert in WEP, plans a strategy to interact with involved in both its prior context subsequent context, cooperatively to fragments of text. the expert s and its interpret The throw expert begins executing and investigates the nature of the lexical sequence to its left. The who expert provides the appropriate information, i.e., the concept structure representing the man, and throw begins to disambiguate its meaning in context. The experts for a and game mutually agree on their local meaning, and through 1 inguistic a:1d idiosyncrat i c interactions with throw help it determine its meaning as the "throwing of a contest". The likes expert starts executing, and its me ssages in search of the person doing the "liking" are intercepted by the who expert, which has been on the loe>kout for such interactions since the beginning. Since the who expert knows the unique name of the concept structure representing the man, it sends this concept to likes, which proceeds normally, knowing nothing of the s tructural complexities preceeding it.

The word experts for both throw and for likes can be expected to explore the underlying meaning of the lexical sequences preceding them. Note the way that WEP applies this linguistic knowledge to the interpretation of fragments of natural language text containing these words. Rather than saying that throw and likes act as finite verbs in certain context s (which are described in some relational representational scheme, such as grammar rules or logic), we say instead that these words carry on linguistic interactions with the active processes modelling the other words making up the (local linguistic) context to arrive at a mutually acceptable characterization of their individual contributions to textual meaning. The advantage of this perspective comes from the fact that linguistic interactions constitute but a portion of all possible lexical interactions that represent in WEP the process of understanding.

3.3 Discourse Interactions

While it is clear that certain lexical sequences cannot be understood solely through recourse to syntax and semantics, namely those fragments for which idiosyncratic interactions are required (i.e., specific remembered contexts), why do we need other kinds of general knowledge? We have already seen examples suggesting the answer to this question. In trying to understand the meaning of throw in the towel, the relevant word experts must find out some things about the person perjorming the described action, before knowing what action he is in effect performing.

If the discourse describes some sort of competition between two people (or teams), for example, throw in the towe1 could indicate a concession of defeat by one of them. The following fragment illustrates such a contextual situation.

(6) "Rick and Joanie play chess. Rick throws in the towel."

On the other hand, if the discourse has recently made reference to a place where one might dispose of a towel, throw in the towe1 might be signifying the putting of some towel in that place. The following example illustrates this case.

(7) "Joanie drops a penny in the pit. Rick throws in the towel."

I am not claiming that knowledge of the discourse context i s sufficient to disambiguate the meani 11g s of the exampl e sentence, but rather, that such knowledge i s required to understand it.

The discourse interactions required t u interpret the above example take place (a) betl-leen the throw expert and a higher order process modelling the activity context, and (b) between the in expert an a process modelling the di scour se focus of attention.* There are two aspect s to the processing of the activity mechani sm, the unsolicited sending of control signals to indi cate the anticipation of certain actions in the text and concept structures to represent them, and the more data-directed interactions with word expert s (and other understanding processes) to determine t he nature of the actions that actually do occur. The throw expert must carry on activity context interactions to determine if the discourse could be seen as discussing some competitive activity. If so, the "concession of defeat" interpretation of the example sentence is plausible. The in expert carries on focus of attention interactions to find out if some location has recently been described i n the text in which something might be thrown.

While the WEP system has t·een directed t OIV<Jrcl the understanding of fragments of text occurring i n textual discourse, the issues arising in the interpretation of dialogue are very similar. The difference between the · two tasks involves the nature of discourse interactions. In interpreting fragments of dialogue from the vantage point of one of the participants, word experts must interact with model processes monitoring the goals of the other participant. The following example (provided by James Allen) illustrates the question.

(8) "When is the Windsor train?"

In trying to understand this question from the perspective of the person at the information desk of a train station, the question could be directed at eliciting either of two pieces of informati on [Allen, 1978], i.e., the time of the next arri va l from Windsor, or the time of the next departure t o Windsor.

By saying that the Windsor train is a "noun- noun pair", we get nowhere in trying t o understand it. In WEP, the word experts for Windsor and train would interact locally and determine the range of possible interpretations for the fragment. In the case of textual discour se, the train expert would carry on discour se interactions with the activity process to find out if discussion of some particular train were anticipated in the text. In the case of dialogue, these interactions would occur between train and an intention mechanism, which might determine that t he speaker in the dialogue is concerned with the trains coming from Windsor, and not with the tra ins leaving for Windsor. If the processes modelling the activity context or the speaker intenti ons cannot provide help to the train expert, the word experts for the sequence would construct a concept structure to represent the disjunct of the t wo possibilities, but continue to await the information that would decide between them.

* The term activity context describes a noti on similar to the scripts of Schank and Abel son [1 975] and to the frames of Charniak [1977]. The noti on of focus of attention has been taken directl y fr om the work of Grosz [1977].

3.4 Logical Interactions

The understanding of fragments of natural language text by a particular individual (or computer program) often requires knowledge of the beliefs of that individual. How can this be true given the fact that, even when the dis~ourse context has not been made to render a particular fragment unambiguous, the majority of reader~ still interpret it the same way? The answer lies ~n the common experience brought to ~nderstand1ng . bY readers from the same culture. Th1s not1on carries over to the common experiences of people in .all sorts of sub-cultures as well, such as sc1ent1f1c communities religious groups, age and cla~ s groups, and' so forth. There are pieces of text that would be understood in common by members uf these groups and not by people outside these groups, and other fragments that would be understood in common by almost everyone.

A psychological experiment at Stanford University [Smith, Lance, and Shoben, 1974], described in the paper by Bolinger [1979] that I have made such extensive use of here, tested the meaning of various words in terms of class membership. Subjects were asked to judge the truth or falsity of certain sentences, such as A robin is a bird and A c/1icl<en is a bird. The fact that it generally takes longer to judge the latter true than the former led the researchers to conclude that there are degrees of birdiness. Clearly, different people with different conceptions of notions like birdiness would understand the same sentence in different ~lays. Furthermore, the data of these psychologists, when interpreted in terms of language understanding, suggests that a lexical sequence could be viewed as representing, to a greater or Jesser degree, some particular conceptual notion.

It is my belief that the understanding of many sequences of words takes place through a process of comparing new sequences to already interpreted ones. Since the linguistic experience of each individual is different, language comprehension must necessarily take place in a person-specific manner. Examine the following sentence for an example of what I mean.

(9) "Joanie Caucus throws a seminar."

Most people reading this sentence will understand the notion of throwing a seminar by comparison with the other kinds of · things they have (linguistically) seen thrown. When someone throws a party, for example, he is organizing a social event. When someone throws a chess match, he is losing the match on purpose. My claim is that by comparing throwing a seminar with other fragments involving throwing, people come up with the intended interpretation of the fragment.

Multiple Choice Perspectives

This notion of comparison with already kno11·n concepts is a fundamental method of 1nterpretation in the Word Expert Parser. Individual word experts interact with a memory of real-world knowledge to determine whether certain conceptual notions can be perceived as other ones. The paradigm for these interactions is based on multiple choice - - of all the fragments of text that have already been understood (the finite choice), which most closely resembles the one now being examined? For the example sentence above, the question in WEP would be put forth by the throw expert and would be some variation on the following (depending on the knowledge o.f throw stored in the throw word expert): "Is a seminar better viewed as a party, a tantrum, a chess game, a legal case, or a

baseball?".*

This type of logical interaction (interaction between a word expert and a process modelling beliefs about the world) has a fundamental role in UEP language understanding. The queries like the example are called (multiple choice) view 1nteractionL It is important to note that these multiple choice queries often take place without including a "none of the above" opt10n. The understanding process must be directed toward the goal of providing some interpretation on each portion of input text. This means that a reader must do the best that he can to understand it, whether or not he has sufficient linguistic and cultural experience to come to the correct (intended) interpretation. Sometimes we make mistakes.

Plausibility of Propositions

The multiple choice view interactions allow word experts to use general knowledge about conceptual items in the world to understand fragments of natural language text. Another kind of general knowledge concerns structures more complex than single items, namely relat1ons among several such conceptual items. When we percei ve such a relation as being convertible into some sort of truth value, we call it a proposition. In WEP, propositions may have truth values along a w1de range, from something we might call completely disbelieved to the opposite extreme we could call completely believed. I say that propositions in UEP work like this not because of the existence of a computer program to operate on them, but rather, because such a program must exist for language understanding within the WEP framework.

The reason involves one class of logical interactions, those between individual word experts and a process maintaining beliefs about the world. A word expert may interact with this belief modelling process to determine the relative plausibility of two propositions. Consider the following sentences, both of which have been successfully interpreted by the existing WEP system (with the user acting as the belief modelling process).

(10) "The man eating tiger growls."

(11) "The man eating spaghetti growls."

The difference between these two fragment s from the perspective of WEP involves the relative plausibility of tigers that eat men and men who eat tigers in the first case, and of spaghetti thdt eat men and men who eat spaghetti in the second. Of course, in certain contexts, the problem i s resolved through discourse interactions; the activity context or focus of attention could make clear the appropriate meaning of the sentences without any need at all for more general knowledge. Clearly, however, we can understand t~ese fragments perfectly well without any guid1ng discourse context.

* Note the important effect such logical interactions have on motivating general inference processes in understanding. The augmentation of a real-world knowledge base depends on making appropriate inferences, and ~EP motivates these by forcing the system to find relationships between previously unrelated notlons. For example, by asking to relate seminar and part~ the parser instigates the construction .of a new co?,cept t o represent "an organized act1v1ty of people ·

The understanding of these fragments i s coordinated in WEP by the word expert for the affix in~ The ing expert interacts linguistically ~ith the experts for the words around it, helping them form meaningful sequences, and carries on logical interactions with the belief modelling proces s to determine the relative plausibility of the t~o propositions possibly signified by the larger sequence. In the first case above, the ing expert begins executing after the and man ha·:e already started constructing a concept structure to represent the meaning of the man. It awaits the report of this concept structure, as well as the one to be reported by the tiger word expert. Furthermore, ing carries on linguistic interactions with eat to arrive cooperatively at a concept structure representing its meaning. The ing expert then has a plausibility interaction with the belief modeller, and coordinates the remainder of the understanding process based on this important len owl edge.

4. SUMMARY

Word Expert Parsing is a linguistic theory based on a lexical organization of linguistic knowledge represented procedurally in word experts. The comprehension of fragments of natural language text is viewed as a process of word interactions, where active lexical agents cooperate to form meaningful sequences of interrelated lexiral items. Lexical interactions are of four types, idiosyncratic, 7 inguist ic, discourse, and logical. Idiosyncratic interactions allow WEP to explain the understanding of idiomatic (more or less idiomatic) lexical sequences, by comparing new sequences with explicitly remembered ones (called prefabs by Bolinger [1979]). Linguistic interactions tnable the use of syntactic and semantic generalizations to interpret fragments, and discourse interactions provide word experts with knowledge of discourse activities and foci of attention. Logical interactions allow word experts to use knowledge about the real-world, especially about the multiple perspectives of individual conceptual objects within it and the relative plausibility of propositions about it.

5. ACKNOWLEDGEMENTS

The theory presented in this paper was originally conceived in wonderful cooperation with Chuck Rieger over the past several years. Some of the sensible perspectives in the paper have benefitted from much appreciated written and spoken suggestions and criticisms of Yorick Willes, Dick Hudson, Dwight Bolinger, Pat Hayes, and James Allen. The Groupe d'Intelligence Artificielle of the Universit6 Paris VIII -Vincennes has provided an excellent environment for research. Thank s t o Patrick Greussay, Harald Wertz, Daniel Goosens, Annette Cattenat, and Gerard Paul. Many extra thanks to Patrick for all the personal and professional help he has given in making my year at Vincennes both valuable and enjoyable.

6. REFERENCES

Allen, James (1978), Recognizing Intention in Dialogue, Technical Report, Department of Computer Science, University of Toronto.

Bolinger, Dwight (1975), Aspects of Harcourt Brace Jovanovich. Language,

BolInger, Dwight (1979), Meaning and Memory, in Experience Forms, Haydn (ed.), Mouton.

Charnia~, Eugene (1977) , Ms. Malaprop, .ll Language Comprehension Program, Fifth International Joint Conference on Art if i cia I I nte I I i gence.

Chomsky, \oam (1<165), Aspects of the ·Theory of Syntax, MIT Press.

Dik, Simon C. (1978), Functional Grammar, 1\'orth Holland.

Fillmore, Charles J. (1968) , The Case for Case, in Universals in Linguistic Theory, Bach and Harms (eds.), Holt.

Grosz, Barbara (1977), The Representation and Use of Focus in Dialogue Understanding, Techni cal \o te #151, Stanford Research Institute.

Hudson, Richard A. (1979), Pan-Lexica/ism, h·orki ng Paper, Department of Phonetics and Lingustics, University College London.

Jackendoff, Ray (19 72), Semantic Interpretation in Generative Grammar, MIT Press.

Jackendoff, Ray (1976), Toward an Explanatory Semantic Representation, Linguistic Inquiry, Y. 7 , n. 1.

Kaplan, Ronald M., and Joan lv. Bresnan (1980) , Lexical-Functional Grammar: A Formal System for Grammatical Representation, Occasional Faper #1 3, MIT Center for Cognitive Science.

Marcus, Mitchell P. (1979), An Overview of a Theory of Syntactic Recognition for Natura/ Language, AI Memo #531, MIT Artificial Intel igence Laboratory.

Miller, George A. (1978), Semantic Relations among Words, in Linguistic Theory and Psychological Reality, Halle, Bresnan, and Miller (eds.), MI T Press.

Mitchell, T. F. (1971), Linguistic "goings on"; co11ocations and other lexical matters arising on the syntactic record, Archivum Linguisticum, v. 2.

Rieger, Chuck (1977), Viewing Parsing as Word Sense Discriminatio~ in A Survey of Linguistic Science, Dingwall (ed.), Greylock Publishers.

Riesbeck, Christopher K., and Roger C. Schank (1976), Comprehension by Computer : Expectation-Based Analysis of Sentences in Context, Resear ch Report #78, Department of Comput er Science, Yale University.

Schank, Roger C., and Robert P. Abel son (19 75l , Scripts, Plans, and Knowledge, Fourth Internationa l Joint Conference on Artificial Intelligence.

Small, Steven (1980), Word Expert Parsing : II Theory of Distributed Word-Based Natural Language Understandin~ Technical Report #954, Department of Computer Sc i ence, Univer s ity of Maryland.

Small, Steven (1981), Toward a Cognitive Mechanics for Distributed Mode71ing, Technical Repor t , Department of Computer Science, Uni \'ers i ty of Rochester (to appear).

Smith, Ed~ard E., Lance J. Rips, and Ed~ard J . Shoben (19 74) , Semantic Memory and Psycho1ogica1 Semantics, 1n The Psychology of Learning and Motivation, v. 8, Bower (ed.), Academic Press .

Wilk s, Yori ck (1980) , Some Thoughts on Procedural Semantics, Cog[1itive Studies Centr e Report #1, Univer sity of Es sex.