Using Language and Context in the Analysis of Text

Using Language and Context in the Analysis of Text*

Yigal Arens

Computer Science Division Department of EECS

University of California, Berkeley Berkeley, California 94720

We will describe our model briefly her~, and will then go over the separate components In mor~ detail. A system called PHRAN (PHRasal ANalyzer) based on our view of the language centered comp?nent has been implemented, and will be ~xtensively described in most of the paper. An outline .of some of the ideas behind context centered processing wil l be given in section 5.

1.1 Brief Outline

According to our model, in understanding a text the language centered component reads .the text, recognizes the words and phrasal patterns In.It, and produces conceptual fragments representing the meaning of the patterns . Plus other contex tual information it may determ1ne from Its knowledge of the language . These fragments are accumulated In a buffer which is considered the context of the utterance. The contex t centered component looks at the conceptual fragments as they arc 1ntroduced to this buffer and uses some world knowledge to introduce to it yet other rel~ted conceptualizations. It mat ches the appropriate fragments to slots that we re left empty by the language centered process! and it clarifies the meaning of some fragments In the manner described 1n sec tion 5., thereby forming more complete conceptualizations in the buffer. These conceptual fragments are then used to repr~sent the compl ete meaning of the text the system 1s read1ng. Context is also used by the laneuage centered process wh en needed to resolve ambiguit1es.

In our model of language understanding, the language centered component ha~ knowledge about the meaning of words, but 1n add1t1on, much of the knowledge is about larger forms of utterances. This knowledge is stored in the form of pattern-concept pairs, where the pattern is a phrasal construct.ol vary1ng specificity, and the concept 1s a .notation that represents the meaning of the phrase . Together, this pair assoc1ates various forms .of utterances with their meaning, and poss1bly with other aspects of the context 1n whi ch they may appear.

The und erstanding process matches incoming utterances against known patterns and uses the concepts associated with the matched patte rns to represent the meaning of the text.

PHRAN (PHRasal ANalyzer) is a system we hove dPveloped that models the language .centered component of natural language understanding. It reads English text and produces structures that represent its meaning , so that a context centered component may act upon them. As 1t read s an utterance, PHRAN sea rches its knowledge base of pattern-concept pairs for patter~s that best Interpret the text. The concept port1on of these pairs is then used to produce the mean1ng representation for the utterance.

PHRAN is able to handle phrasal language units which are found with great frequency In ordinary speech and common natural language text -- in the same manner that it handles the r est of the text. Since control and representation are kept .separate, it is quite simple to add new 1nformat1on to the system. All one needs to do to extend PHRAN is add new pattern-concept pairs to the data-base. Mike Morgan, a graduate student at Berkeley, has

Abstract

We describe a theory of natural language understanding within which we identify two separate components, a language centered one and a context centered one. 1he former component uses a knowledge base consisting of pairings of phrases with the concepts associated with them to determine the meaning of utterances. The latter component clarifies the meaning found by the first one and makes it more specific by attempting to reconcil e it with the context of the utterance.

We have constructed a program called PHRAN (PHRasal ANalyzer) whi ch performs the task of thP language cente red component.

1.0 INTRODUCTION

A recent tendency among some researchers in natural language processing has been towards a uniform, single stage process of understanding text. In such an understanding process, all the knowledge the system has is available for use at any time, and any existing routine may be run at any point. One vers~on of this view has be~n ca~ledA by Schank et al \1~80) and Lebowitz \1980), the integrated parser .

At the root of this approach is the realization that previous attempts to process natural language in separate stages -- first finding a syntactic parse and then using semantic information to determine the meaning of an utterance -- were in error. Much research has been done in recent years demonstrating that semantic information is and should be used continually when processing text, and that it is wrong to assume that a separate syntactic parsing stage exists. But in arguing against the syntax-semantics distinction in processing, these natural language researchers have denied the existence of any discernible levels, or stages, in the process of understanding language and have argued that it is totally uniform.

We claim that in fact there are two distinct levels of processing in the language understanding process:

1. A langua~e centered component, involving the app Ication of knowledge about phrasal patterns and their meaning.
2. A context centered com~onent, involving the ~ppl1cat1on of knowle ge about the context of the utterance.

These two components together perform the task of understanding the language of the utterance . After these processes have done their job the full range of knowledge we have available to us may be applied to the result. This final application of knowledge will enable us to make inferences, etc., in order to comprehend the input.

* Research sponsored by the Office of Naval Research under contract N00014-80-C-0732 .

con~truct~d data-bases of pattern~concept pairs. that enable PHRAN to analyze Span1sh and Ch1nese sentences, with essentially no knowledge of PHRAN's control. The knowledge base used by PHRAN is declarative , and is shareable by a system for ~anguage production. Such a system, named PHRED ( PHRasal English Diction~, has been constructed at Berkeley by Steve Upstill.

2 . 0 PHRASAL LANGUAGE CONSTRUCTS

Many natural language processing systems assume that the meaning of utterances can be computed as a function of their constituents, w1th the cons tituents assumed to be words. Thus all the knowledge about the semantic~ of the language i s stor~d at the word level \Birnbaym ~nd Sel fridge) 1979! (Riesbec(< and Schank, 1975! \lrillksi 1973 (Woods, 1970). However, many natural anguage utterances have interpretations that cannot be found by examinin~ their components. Id1oms, canned phrases, lex1cal collocations, and structural formulas are instances of large classes of language utterances whose interpretation requires knowledge about the entl r e phrase lndependent of 1ts individual words \Becker, 1975) \Mitchell, 1971 ).

For this reason the language centered process needs to know more about the language than the meaning of words.

By the term "phrasal language constructs" we refer to those language units of whi ch the language user has specific knowledge. Our notion of a phrasal language construct is similar to a st ructural formula (Fillmore, 1979). Our constructs vary greatly tn tlexibilit¥,, from fix ed express1ons such as "a perfect stranger' to most general phrases expr essing the usage of a word sense. For example, to express one usage of the verb ki ck , the phrase "<person> <kick-form> <object>" --rs- used. (This denotes a perso9 followed by some verb form involvilg k1ck \e. g., kick, kicked, "would have kicked" foTIOWed by some utterance denot1ng an object .

3 .0 PHRAN**

PHRAN is a system which models the processing done by the language centered component while understanding ~glish text. PHRAN integrates both generat1ve and non-productive language abilities. It has knowledge about individual words combined with knowledge about longer utterances of the English language.

HerE' are some examples of sentences for which PI!RAN is able to produce conceptual fragmen t s r Ppresenting their mPaning. These fragments s hould t hen be passed on to the buffer upon whi ch a sys tem modeling the context centered process should act to reconci l e them with the context and produce more complete conceptualizations. Since no such system exists as yet, PHRAN has been extended to do some of i t s processing. The understanding of the following i s done , therefore, entirely by PHRAN.

• Oi lmen are encouraged by the amount of natural gas discovered in the Baltimore Canyon, an undersea trough about 100 miles off the New Jersey coast. (Newsweek , Feb 1980)
• Tenneco, one of 39 companies engaged in drilling in the area, thinks its leased tract contains a marke t able supply of gas.
• The young man has gotten into an another argument with his boss.
• Willa will drive Bill to The Big Appl e if she is given twenty five dollars .
• For more information on these projects s ee R. Wilensky's paper in this volume.
• ** For a more extens ive di scusnion of ~IR&N phrasal patterns see Wilensky und Arens ( 19i\O) . and

* Bill rewarded Mary for .marrying him by making a payment on her e~pens1ve cAr.

(These sentences are analyzed b~ an uncompiled version of PHRAN on the DEC KL10 syste~ at UC Berkeley in from 1/2 to 6 seconds of CPU time).

PHRAN is centered around a knowledge base of phrasal patterns. These ran~e from literal strin~s such as "by hook or by crook and "in a jiffy , through patterns such as "vice <title>" and "<parent> in law", to mos~ genera~ phrases such as "<number> <plural noun>' and '<person> <go> to <location>".

Associated with each phrasal pattern is a conceptual template' A conceptual template i s a p1ece of mean1n~ representation with poss ible references to p1eces of the associated phrasal pattern . For example, the conceptual templatf, associated with the pattern <parent> 1n law denotes the <parent> of the spouse of the person involved; and with the phrasal pattern "<person> <go> to <location>" is associated a template denoting the movement of <person> from the place s/he was to <location>.

4.0 HOW PHRAN WORKS

4.1 Overall Algorithm

PHRAN is made up of three parts - a database of patt e rn-concept pa1rs, a set of comprehension routines, and a routine which suggests appropriate pattern-concept pairs. PHRAN takes as input English text, and as it reads it from left to right, it compares the input against patterns from the database. Whenever a matching pattern is found, PHRAN interprets that part of the sentence that matched the pattern as describing the concept associated with the pattern in the pattern-concept pair, and passes the resulting conceptualizations on to the buffer where further processing will be performed by the context centered level.

4.1 .1 Overview Of Processing-

When PHRAN analyzes a sentence, it read s the words one at a t1me 1 from left to right. The pattern suggesting rout1ne determines if any new patterns should be tried, and PHRAN checks all the new patterns to see if they agree with that part of the sentence already analyzed, discardin~ those that don't. A word's meaning is determined s1mply by its matching a pattern consisting of that literal word. At this point a term is formed with the properties specified in the~ceptual template assoc1ated w1th the word, and this term is added to the buffer PHRAN uses. PHRAN checks if the term it just added to the list complete.s or extends patterns that had already been partially matched by the prev1ous terms. If a patt e rn is completely matched, the terms matching that pattern are removed from PHRAN's buffer and a new term, specified by the concept part of the patt ern-concept pair, is formed and replaces the terms the pattern matched in its buffer. In addition, an appropriate conceptual fragment is passed on to the buffer on which the context centered process operates, for further processing.

When PHRAN finishes processing one word it reads the next, iterat1ng the procedure just described, periodically passing conceptual fragments to the context centered level of processing.

4.1.2 Overview Of PHRAN Patterns-

A pattern-conce pt pair consists of a specification of the phrasal unit, an associated concept, and some additional information about how the two are related. When PHRAN instantiates a concept in its buffer, it creates an item called a term that includes the concept as well as some aaaTtional information.

A pattern is a sequence of conditions that must hold true for a sequence of terms. A pattern may speci fy optional terms too, the pla9e wher~ these may uppear, and wha t effect \if any their appea rance w1ll have on the properties of t e term formed if the pattern is matched. For example, consider the following informal description of one of the patterns suggested by the mention of the verb 'to take' in certain contexts.

pa~tern to recognize -

L< first term: represents a person>

<second term : i s an active form of TAKE>

<third term: represents a physical object>

<fourth term: the word FROM>

<OPTIONAL part:

<fifth term: represents a person>>]

\ PTRANS ACTOR <f1rst term>) OBJECT <third term>) FROM <fifth term 1f presenti oth~~w~se consu t context>) (TO <first term>)) 1

terlll to fori -

Notice that the fourth and fifth terms are marked as optional. If they are not present in the text, PHRAN will not fill the FROM slot, and the context centered process will use the context to find the right slot filler.

4.1 . 3 Simple Example-

The following is a highly simphified example of how PHRAN processes the sentence John dropped out of high school":

First the word "John" is read. "Johg" m~tches the pattern consisting of the literal John', and the concept associated with this pattern causes a term to be formed that represents a noun phrase and a particular male person named John. No other patt erns are sug~ested. This term is added to *PHRAN-BUF*, the l1st of terms PHRAN keeps and in whi ch it does its processing. So at this point

"Dropped" is read next. It matches the literal "dropped", and an appropriate term is formed. The pattern suggesting rout1ne i nstruc ts PHRAN to consider the 'basic' pattern associated with the verb 't9 ~rop', whi ch is: I L <person> <DROP> <object>] [ ..· ] l

Its initial condition is found to be satisfied by the first term in *PHRAN-BUF* -- this fact is stored under that term so that succeeding ones will be checked to see if this partial match continues. The term that was formed after reading "dropped" is now adde~ to the list. *PH~AN-~UF* is now J < LJOHN1 - person, NPJ , LDROP - verb >

PHRAN now checks to see if the pattern stored under the first term matches the term just added to *PHRAN-BUF* too, and indeed it does. This new fact is now stored under the last term.

Next the word "out" is read. The pattern suggestion mechanism is alerted by th~ ocyurrence of the ve rb 'drop' followed by the word out , and it i9structs PHRAN to consider the pattern] 1 L <person> <DROP> "out" "of" <school> [ .·. ] l

The list in *PHRAN-BUF* is checked against this pat te rn to see if it matches its first two terms. It does , and this fact is stored under the second term. A term associated with 'out' is now added t o *PHRtN-BUF*: < JOHN1 - person, NP] , [DROP - verb] , [OUT] >

The two patterns that have matched up to DROP are checked to see if the new term extends them. Thi s is true only for the second pattern, and this [ act is stored under jhe next term. The pattern <person> <DROP> <object> is discarded.

Now the word "of" is read. A term is and added to *PHRAN-BUF*. The pattern that up to OUT is extended by OF so this fact is und er the fourth term. formed matched stored

The word "high" is read and a term is formed and added to *PHRAN-BUF*. Now the patt ern that matc h~d up t9 OF js compared ~g~inst HIGH. It doesn t Sat1~fy the next cond1t1on. PHRAN reads

"school", and the pattern suggestion routine present s PHRAN with two patterns:

1. I [ "h igh" "school" ] [ representation denoting a school (o~ 10th through 12th graders j I
2. I r <adj ec tive> <noun>] [ representation denotiqgl noun modified by adjectiveJ

Both patterns are satisfied by the previous term and this fact is stored under it. The new term is added to *PHRAN-BUF*, now:

< [JO~N1 -[ person, N~] '[[DROP - verb] , [OUT]], [OFJ , HIGH - adjJ , SCHOOL - school, noun >

The two patterns are compared against the last term, and both are matched. The last two terms are removed from *PHRAN-BUF*, and the ~atterns under OF are checked to determine which of the two possible meanings we have should be chosen. Patterns are suggested such that the more specific ones appear first, so that the more specific interpretation will be chosen if all patterns match equally well and if there is no possib1lity of further patterns being suggested that will clarify matters. That is the case here since we have reached the end of the sentence.

A term is formed and added to *PHRAN-BUF*, which now contains

The pattern under OF is checked against the last term in *PHRAN-BUF* . PHRAN finds a complet e mat ch, so all the ma tched terms are removed and replaced by the concept associated with thi s pattern.

*PHRAN-BUF* now contains this concept as the final result:

< [ ($SCHOOLING (STUDENT JOHN1) (SCHOOL HIGH-SCHOOL1)

(TERMINATION PREMATURE)) ] >

4.2 Pattern-concept Pairs In Some More Detail

4.2.1 The Pattern-

The pa ttern portion of a pattern-concept pair consists of a sequence of predicates. These may take one of several forms:

1. A word; which will match only a term representing this exact word.
2. 2 . A class name (in parentheses); will match any t~rm re~resenting a member of thi s class (e. g. (FOQD)" or "(PHYSICAL-OBJECT)").
3. A pair, the first element of which is a property name and the second is a value ; will match any term having the required v~lue of the property (e. g . "\Pa rt-Of-Speech VERB)").

In addition, we may negate a condition or specify that a conjunction or disjunction of several must hold.

The following is one of the patterns which may be suggested by the occurrence of the verb 'give' in an uttn·ance:

L\PERSON) (ROOT GIVE) (PERSON) (PHYSOB)]

4. 2.1 .1 Optional Parts-

To indicate the presence of optional terms , a li s t of pattern concept-pairs is inserted into the pattern at the appropriate place. These pairs have as their first element a sub-pattern that will mat ch the optional terms. The second part describes how the new term to be formed if the main pattern i~ found shou ld be modif1ed to rer ect the existence or the optional sub-pattern.

The concept corresponding to the optional part of a pattern is treated in a form slightly different from the way we treat regular concept parts of pattern-concept pairs. As usual, it consists of pairs of expressions. The first of each pair will be placed as is at the end of the properties of the t e rm to be formed, and the second Will be evaluated first and then placed on that list.

For example, another pattern suggested wh en give ' is seen is the following:

[(PERSON) (ROOT GtVE) ~PHYSOB) (tf6°(0~¥~~~t)~ CD-FORM))))]

The terms of this pattern describe a person, t he verb ~~~~~ and then some physical object. The las t term d Ibes the optional terms, consisting o f the word to followed b~ a person description. Associated with~is pattern IS a concept part that specifies what to do with the optional part if it is there. Here it specifies that the second term in the optional pattern should fill in the TO slot in the conceptualization associated with the whole pattern.

4. 2. 2 The Concept -

When a pattern is matched PHRAN removes the terms that match it from *PHRAN-BUF* and replaces them with a new term, as defined by the second part of the pattern-concept pair. For example, here IS a pattern-c9ncept pair that may be suggested when the verb 'eat is encountered:

([(PERSON) (ROOT EAT) U((FOOD)) ] ] (FOOD (OPT-VAL 1 CD-FORM)) )

[P-0-S 'SE~TENCE CD-FORr '(INGEST (ACTO~ ?ACTOR) (OBJECT ?FOOD)) ACTOR VALUE 1 CD-FQ~~ FOOD ' *FIND* (FOOD))J

jhe concept portion of this pair describes a term covering an entire sentence, ancj whose meanil_lg is the action of INGESTing some food (Schank, 1975). The next two descriptors specify how to fill in variable part~ of this action. The expression (VALUE n prop) specifies the 'prop' property of the g'th term 1n the matched sequel_lce of the pattern ~not counting optional terms). OPT-VAL does the s ame thing with regards to a matched optional sub-pattern. Thus the concept description above s pecifies that the actor of the action is to be the term matching the first condition. The object eaten will be the term corresponding to the optional pattern if it is found, or else it will be determined by the context centered process from context.

4. 3 Pattern l4anipulation In More Detail

4. 3.1 Reading A Word-

When a word is read PHRAN compares the patterns offered by the pattern suggesting routine with the list *PHRAN-BUF* In the manner described in the example in section 4.1.3. It discards those patterns that conflict with the information already In it. Then PHRAN tries to deterwine wh~ch meaning gf the word to choose, using the 'active patterns ~those ttat have matched up to the point where PHRAN has read). It checks if there IS a particular meaning that will match the next slot in some pattern or, . if no such definition exists, if there IS a meaning that might be the beginning of a sequence of terms whose meaning, as determined via . a pattern-concept pair 1 will satisf~ the next slot In one of the active patterns. If this is the case, that meanin~ of ~he word is chosen. If not enough information 1s available to facilitate a decision, all meanings are recorded with the hope that a pattern suggested at a later point will make a choi ce possible then.

In principle, of course~ the context centered level should be able to nelp in such situations, using the context.

A new term is formed and is both passed the buffer of the context centered level and recorded in *PHRAN-BUF*. If it satisfies the next condition in one of the active patterns, the appropriate pattern i s moved to the pattern-list of the new term. If the next condition in the pattern indicates that the term specified is optional, then PHRAN checks for these optional terms, and if it is convinced that they are not present, it checks to see if the new term satisfies the condition following the optional ones in the pattern.

4.3. 2 A Pattern Is Matched

When a pattern has been matched completely, PIIRAN continues checking all the other patterns on the pattern-list. When it has finished, PHRAN will take the longest pattern that was matched and will consider the concept of its pattern-concept pair to be the meaning of the sequence. If th.ere are several patterns of the same length that we re matched PHRAN will group all their meani ngs together.

New patterns are suggested and a disambi!uation process follows, exactly as in the case o a new word being read, with the resulting meaning(s being passed to the context centered process.

For example the words "the big app~e", when recognized, wili have two possible mean1ngs; one being a large fruit, the other being New York City. PHRAN will check the patterns active at that time to determine if one of these two meanings satisfies the next condition in one of the patterns,and if so, that meaning will be chosen. Otherwise, both meanings will be entertained until future information confirms one of them.

4,3.2.1 Possible Extension Of A Pattern-

It sometimes happens that we have two patterns where the smaller one ends before the longer one. For examp~e considef ------ L {PERSON) ROOT SIT) (EVENT) OUT ]

and

when PHRAN rea~s the sentence "Jane sat the gawe out'. In this case upon analyzing the word "sit PHRAN realizes that the second pattern has matched completely, but it is also aware of the fact that there exists a pattern which has matched up to the same point and which also may be present in the text. Since PHRAN is intended to match the longest pattern possible, it will not treat the shorter pattern as matched yet. PHRAN will continue reading and analyzing the text until either the longer pattern is found to be present, in which case it will be used to find the meaning of the sentence, or until the longer pattern fails and PHRAN concludes that it is not present. In the latter case PHRAN will back up--ana consider the shorter pattern matched.

r (PERSON) (ROOT SIT) ],

4.4 Indexing And Pattern Suggestion

Retrieving the phrasal pattern matching a particular utterance from PHRAN's knowledge base is an important problem that we have not yet solved to our complete satisfaction. We find some consolation in the fact that the problem of indexing a large data base is a necessary and familiar problem for all knowledge based systems.

In the current version we index the pattern-concept pairs of the database in a tree. As words are read the pattern suggesting mechanism travels down this tree, choosing oranches according to the meanings of the words. It suggests to PHRAN the patterns found at the nodes at which it has arrived. This list of nodes is remembered, and when the next word is read the routine continues to branch from them, in addition to starting from the root. In practice, the number of nodes in the list is small.

For example ~ whenever a noun-phrase is foll owed by an active rorm of some verb, the .node the suggesting routine arrives at in the tree Instructs PHRAN to consider the simple declnrAtive f o rms of the verb. The phrasal pattern that will recognize t ~e expression "by and lArge" is found at the node reached only after seeing those three word s consecut ively. In this manner this pnttern will be sugges ted only when necessary.

The main problem with this scheme is that it does not lend itself we ll to allowing contextual cues to influence the cho1ce of patterns PHRAN should try.

5.0 CONTEXT CENTERED PROCESSI NG AND RECONCILIATION

1. 1 The Context, And Reconciling Conce ptuAl Fragments With It

Consider the disambiguation: well known example of

1. The old man's glasses were filled with sherry.

When mo~t people read this senten9.e, the wor,d glasses is .Interpreted as ml;aning . eye-~lasses., as opposed to drinking glasses · This.decision I~ based on the fact that the mention of the old man activates some knowledge concerning older people and their normal physical appearance.

The difficulty people have with this sentence supports the claim that in understanding text people form a context, which includes the concepts appearing up to the point whe re the~'ve read and related knowledge. Upon processing future utte rances and producing additional conce ptual fragments, the understander attempts to reconcil e these fragments -- whethe~ they describe objec ts or higher level concepts -- With the context. We call this process reconciliation.

The existence of a context will als9 help und erstanding system in the processing of following sentence: the the

(2) Mary cut the salami.

Here, reconciling the general notion of cutting with this particular context w1ll enable the s~stem to realize that, among other th1ngs 1 the. cutt~ng was done with a knife, it was aone 1n a d1rect1on perpendicular to . the length of the sal~mi, or at a slight angle to 1t, and that the salam1 was probably peeled too.

On the other hand, when we hear mention of cutting in the context of a meal, the reconciliation process enables the system to realize that the 1nstrument used was a knife, and that with the help of a fork the person doing the cutting separated bite sized p1eces from the food and proceeded to ingest them.

In comparing the limited number of inferences that can be made from a generalized concept of cutting to what can be understood from the fully specifled one the system arr1ves at after reconciling it with the context , we see the utility of this process to the natural language understander.

5 . 2 The Context Centered Process

In view of the previous d1scussion we have decided to include in our model of natural language understanding another component, the context centered component. This component -aTtempts to reconc1Ie conceptual fragments which are introduced to the buffer with what is already known about the context of the utterance.

No program has this component of this point state expected to do: yet been written implementing our model. However, we can at in general terms what it is

1. Add recognized concepts to the context.
2. Recall r elAted con~e pts from memory and ndd them. tQ context \making use of other routlnesJ .
3. ) . Wheneve r add i ng n concept to context, check if context requir~s a pnrtl cular interpretntion of it. (Thus resolving some ambiguities that the language ce~tercd level is unable to resolve by itself.)
4. Determine references (based on the known context).

In performing its task the context centered process makes use.of the context of the utt e rance as previously r ecognized. In determinlng references , for instance, it will not be able to resolve a reference if the thing refered to has not been previously int roduced to the context.

Unlike the knowledge base used by the language centered process, much of the knowledge used at this stage is not relevant only to understanding ideas communicated to us. We apply here, among other things, the same knowledge we might use in order to understand someone's obse rved actions, or to plan our own. For exampl e , the knowledge about cutting sn]Ami used to understand the sentence dealt wi t h earlier may be helpful when we decide to serve some to friends, or when we are trying to figure out what someone is doing when we see them bending over a salami with a sharp knife in their hand. Thus, the context centered component may call on other routines to supply it with further r elated conceptunlizntions --which will contribute to the context of the utternnces.

One should keep in mind that the two processes described in this paper are constantly influencing each othe r. Obviously, the language centered component supplies the context cent~red compon~nt with fragments to work on; but there IS Interaction in the other direction too. As we have seen fragments recognized by the language centered process may be used as constituents in further processing by it. The context centered level, however, may add information to the fragment found by the previous level, and in this ~anner . modify . the meaning of an~ fragments of wh19h this one 1s a part. In addition, the information concerning the context, whi ch is collected during context centered processing, is used, when necessary, to determ1ne the appropriate meaning of a word or phrase the language centered component has found.

5.3 Example

In order clarify some of the ideas described. in the previous sec tion, let us simulate the processing of the following story:

It was duck hunting season. Mary went hunting. She aimed at a bird and pulled the trigge r. She got the bird.

When the first sentence is processed by the language centered level, an appropriate conceptua l fragment is produced, namely one representing the fact that the time of year described was that during whi ch one could hunt ducks. When the context centered process sees that duck hunting is mentioned, 1t consults memory about ducks . and about hunting them. It adds to the context, 1n addit1on to the fragment representing the first statement, conceptualizations representing:

Prototypical knowledge about ducks, knowledge about hunt1ng, an~ in particular,

guns and how they are used e. g. representations for 'aiming ' and 'shooting' ,

and perhaps others (the p~ecise conceptu~lization s introduced to context w1ll be a funct1on of hQw memory is organized; cf. MOPS in Schank, 1980) . Among the co9ceptualizations ~dqed there will be one of the form ($HUNT (HUNTER ?YJ (OBJECT DUCK)), with some indication that it has not yet been seen in the t ext. This fragment will be of use later.

When reading the second sentence, the language centered l~vel p:odu~es a fragment representing tne f act discr1bed 1n ~t· 1. ~· something like $HUNT HUNTER MARY) \OBJECT ?XJ). Then the context centere p:ocess, in attempting to reconcile the fragment w1th the context, 1dentifies the hunting in ~he cont~xt (i. e. hu~ting duck, plus the additional 1nformat1on there) w1th the hunt1ng Mary was doing. This will happen be9ause in tbe context we b~ve as we recall, \$HUNT \HUNTER ?Y) (OBJECT DUCK)), ~hich the previous fragment will match.

When the word 'she' is read the person is identified as Mary, since the context contains ~now~edge.of only.one perso~ .. The concept of aiming 1s 1dent1f1ed w1th the a1m1ng of a gun already in the context, in a fashion similar to that descr1bed f~r the prev~ous sentence. The bird is identified w~th a duck, s1~ce the system is told that Mary a~m~d at a b1rd, conte~t ~ncludes the concept of a1m1ng at a 9uck, and spec1fy1ng the bird as a duck 1s both poss1ble -- 1n th1s case because a duck is a bird -- and enhances the match between the text and the context. Since the aiming in the context is mark~d as part ~f the shooting w~i~h is part of hunt1ng 1 we w1ll get that the a1m1ng ment1oned in the text.is also understood as part of the process of hunt1ng duck. The pulling of the trigger is, again, understood in terms of using the gun and hunting.

When reading the last sentence, the language centered component comes up with several possible meanings, among them that 'Mary received the bird from an unknown party', and that 'Mary succeeded in doing what she wanted to do to the bird'. The information about hunting already in the context enables the context centered component to decide in favor of the second meaning, because we know nothing about anyone giving Mary anything, but we have represented the fact that Mary want~d to do something to the bird (i. e. k1ll it). This conceptualization will match part of the second p~ssible meaning of th~ phrase, so the last sentence w1ll be understood as Mary succeeded in killing the bird'.

Becker 1 Joseph D. (1975). The phrasal lexicon. In Theoretical Issues in Natural L?n~ua'e Processing : H. Schank an~ ~an-webber e s .. Cambr1dge, Mass.

Birnbaum, L., and Selfridge, M. (1979). Problems in concept~al analysis of natural language. Yale Un1vers1ty Department of Computer Science Research Report #168.

Fillmore, C. J. ( 1979). Innocence: A Second Idealization for Linguistics. In Proceedings of the F1fth Berkeley Language Symposium, Berkeley, 'CliT!To rfii a ·

Lebowitz, Michael (1980). Generalization and Memory 1n.an Integrated Understanding System. Yale V.n1vers1ty Dept. of Computer Science Research Report #186.

Mitchell! T. F. (1971). Linguistic "Goings On"; Collocat1ons and Other Matters Arising on the Syntact1c Record. Archivum Linguisticum 2 (new series) 35-69.

Riesbeck, C. K., and Schank, R. C. (1975). Comprehension by compu~er: expectation-based analys1s of sentences 1n context. Yale University Research Report #78.

Schank, R. C. (1975). Conceptual Information Processing. American Elsev1er PUbl1sh1ng Company Inc., New fork. '

Schank, R. C. (1980). Language and Memory. Cognitive Science, vol. 4, no. 3.

$chan~, R. C., Lebowitz, M., and Birnbaum, L. (1980). An Inte~rated Understander. Am. J. of Computational L1nguistics vol.6, no.~----

Wilensky, R., and Arens, Y. (1980). PHRAN- a Know1e9ge Ba~ed Approach to Natural Language Analys1s. Un1vers1ty of California at Berkeley Electronic Research Laboratory Memorandum No ' UCB/ERL MB0/34. .

Wilksi Yorick (1973). An AI Approach to Machine Trans ation. In ComEuter Models of Th?u~ht and Lan~ua~e, R.C. Scna k and~COTby e s.)~.H. Pre rna and Co., San Francisco, 1973.

Woods, W. A. (1970). Transition Network Grammars fo~ Natural Language Analysis. CACM 13, 591-606.