American Journal of Cornput ational Linguistics 
TEXT UiiDERSTANDING: 
A SURVEY 
ROBERT YOUNG 
Department of Computer Sciences 
University of Texas 
  us tin, Texas 78712 
Copyright 01978 
Microfiche 70 
Association for Computational Lingu~stics 
TABLE OF CONTENTS 
Introducti~n~..~~-~~~~--~~~~,.~.~...,..~- 
1, The Content of- Connected Discourse , . , , . , , . , , . , , . . 
1-1 The Nature of Discourse Content . , , . , , , . . , . . , , . 
1,1,1 Barclay, Bransford and FrAnks , , , . , . , , . , , , , . 
1-1-2 Frederikserl. . . a ., . , . , , . , , , , , . . . . . . 
1.1.3 Kintscll....., . ,...,,, ..,.,. .,,. 
1.1.4 Th~mdyke,,.-~*,,~,.-,,,~.,,....- 
1,2 The Structure of Discourse Content , . , . . , . , , . ., 
1-21 Bartlett -mm~,~~~..,,,,~.,,,~~,,~ 
1-2.2 Crothers . , , . .. . , , .. , . , . , , . , , . , . , , , 
1.2.3 Grimes and Meyer , . , , , , , , , , , . , . , , . , . , 
12.4 Kintsch andvanDijk , . , , , , . - a . . . . , . , 
1-3 Some Considerations Regarding Memory Experiments . , , . - . , 
1.4 Discussion and Conclusians - . , - . , . . . . , . . . . . . . 
2, The Form bf Connected Discourse , , , . , . , ' , . . , . , , , 
2.1 Structural Analysis of Text - , , . , , - . , , . . . , . , 
2.2 KintschandyanDijk - , - -. , , , . I - -. . . - - , a 
2.3 R~rnelhart~.~~,-,.~.,.~,,~-,--.~.,.,. 
2.4 Mandlerand Johnson - , , , , , , - , , . . . . , * . , . . 
2.5 Th~rndyke....,-~.~-,.,--,-,,,,~,..,- 
2.6 TextGramrnar .~,~.,-o,.,,,,.,..em,D,., 
2.7 Discussion and Conclusions . , . , . , , . . . , . , . , , , 
3, Computatiofial Models of Text Understanding . . , , . . , 
3.1 The Necessity of World Knowledge rn , , . , . , . . . * , , , 
3.11 Charniak ~a.,..,m...o., 
3.1.2 Rieger ,.,,,.....-.--e,-.-.--.,- 
3-1-3 Schank ,.,.m-....,-,...m,,,.m.,e 
3-2 The Organization of World Knowledge , . . . , a , , . , . , , a 
3.201 SchankandAbelson , , , ,, , , . . , . . . . . . , . 
3-202 Phillips 0 0-0 - .. - -0 -0 o om? 
3.3 Otherwork ..-.I.,,.,,--~-~,-~,.,, 
3-4 DiscussionandConclusions , . . , . , . , . -, a. . . . . . 
4, Final Observations , , , , - . , , . , . , , . , . , , . . , . , , 
 reference^^^...^^,^,,,,,,,,,^,.,^^,, 
Page 3 
INTRODUCTION 
The goal of this study is to examine work that has something to offer 
toward the construction of a computable model of text understandhg, and 
therefore toward cognitive models in genenil. The reason for the criterion of 
computability is that this rather strict requirernevt makes vague 
gcnernlizatiot~s i~nposisible (or at least more difficult), and forces exact 
specification of proccsscs being hypothesized- (However, this criterion is 
frequently difficult to apply, and in some instances the author s decision was 
undoubtedly a subject ivc one. ) 
It should be clearly poirltcd out that no attempt is made here to deal 
with the semantics of individual sentences. A familiarity with recent work in 
semantic representatian in cognitive psychology, linguistics and artificial 
intelligence is assumed (see Norman and Rumelhart [75], Schank and Colby [73] 
and Steinberg and Jakbbovifis [ 71 1 ) Although many problems in representing 
the meaning of individual sentences remain unsolved, this study focuses an 
those aspects of mcanlng that are conveyed only by groups of connected 
sentences - texts. Additionally, only work that attempts to deal with the 
semantics or understanding of texts, as opposed to statistical or syntactic 
analysis, is considered. This focus on text has also led to the omission of 
studies of non-textual mpmory and of computational systems that understand and 
solve problems stated in English or carry on dialogue. This omission is not 
meant. to imply that work in these areas has no relevance to text 
understanding, but the dif fcrent focus and additional constraints of these 
studies make such implications difficult to isolate. 
Page 4 
The study is divided into three parts, which occasionally overlap, The 
first section deals with the conterlt of coni~ccted text - that is, what exactly 
does text commtlnicate. The work in this section is primarily that of 
experimental psychologi~ts interested in memory and recall- The second 
section deals with the _strucrurq of text, apart from its specific content, 
This work has been done mairlly by cultural antl~ropologists and those 
interested in the theory of literaturc, but is also being investigated by 
experimental psychologists. The third section discusses computational models 
prop-osed by computational linguists, which include attempts to implement in 
computer programs some of the processes discussed in the other sections- Each 
section is concluded with a dlsc~~ssion of the emerging model of text 
understanding. Finally, some directions for needed research are suggested. 
It is hoped that this kind of interdisciplinary survey will call attention of 
workers 112 one area to work in other areas addressing the same problems, This 
type of communication is valuable in two distinct ways. When similar 
conclusions are reached in different disciplines frequently using different 
methods and with somewhat different goals, these conclusians must be given 
special credence, On the other hand, it is sometimks the case that one 
discipline completely overlooks a problem or some aspect of it due to their 
own interests and biases. Such omissioi~s need to be brought into sharp focus. 
Page S 
1.0 THE CONTENT OF CONNECTED DISCOURSE 
This scctfon will present a brief survey of the recent work of some 
experimental  psychologist,^ concerned with thc process of understanding 
corlncctcd discourse. 
Thc details of their cxperi~n~nts will not be presented, 
but the types of cxperirneals and general conclusions will be discussed, as 
well as any madcls proposed by the investigators. Th~~se results seem to 
cotlve rge toward a single model, in spite of apparent cont radictions, 
Psychological work on text understanding is relevant to computational models 
because humans are still the only good tcxt undcrstnndcrs. Also, although 
some aspects of psychological research, such as forgetting, are not currently 
included in any serious way in computational models, it would appear that such 
models have a very useful role to play in pointing out where significant 
details of the process of text understanding are being handled intuitively by 
the theorist, and have not really been defined in the tlleory, 
1.1 The Nature Of Discourse Content 
The nature (general characteristics and features) of the meaning of 
connected text will be examined first. The principle focus of these studies 
is to compare the understanding of a text to the actual text and attempt to 
charnctefize the types of differences that are found. 
1.1.1 Harrl~ly, Brainsford And Franks - 
I\,~rc*l,ty, Bra~~sford and Franks in n large body of work (Bransford, Rarclay 
irnd l'rrwks [72], l3nrcl~y [73], Bransford and Franks (731, Branrrforrl atld 
Ptt*('*n 1,-11 1741 nnd Frn~~ke and Rraneford [74]) argue for the cxletence of 
II t t c rcy~rcscnt nt ions indepcndcnt of the surface input, and propose active 
11 1 t'haiCS 011(~11t it18 on thc input. They describe eontencea as Information 
whl(\r is ttscld by tlrc undcrstnnder to construct a descript-ion of a situation. 
F n t11oy clnim tl\r~t the itrformnt ion used to construct the selnantic 
dt~*~-lpt [ot~ is not wholly contained in the sentences, but that an underatander 
lr is ~rrtsvious knowlcdgc. to a great extent. They do not make any definite 
01 s r r 11 tlrc Form of the setnant ic representat ions. Typical 
cxl~l1~tt~~wt:: pcrformtbd to test thcsc hypotheses consist of recognition or 
r I 1 QP I-ompoi~nci rtrntt\ncea 1 ikc: 
I. Tllr~c turt lcs rastcd hoaidc u floating log, and a fish swam beneath 
t 111-in. 
3. 'l'1rrt.c turtlm r~8t~d on a floating log, and a fish swam beneath ttiem. 
Tht- plryhi c;\ t sittint ion dc~cribod by the accond sentctlcc is essentially the 
t 1 
I f L rot thrm" is rcplaccd uieh "Jt". This situation differs from 
t I),\ t (it-st.- rl \red i n t lrc f i rs t sent crlct', howcve r. Whcn sub jc ct s we re p resented 
wt r 11 .I hcnt c~~ci\ in whicll tllc p ronor11\ substi rut ion had bcrn made, bhosc who had 
II(*,II tl tlrc* f l rrt scbtrt cl~cc- were able to make thc distinct ion, whf Ic those who 
I \\vnrtf I r;ccot\d mnctc no di~tinction. The explanation offered is that 
r;\~ll jr'ct rz i~r;r\cl thc2 t spntinl kt.rowlcdgc to crclate situatiot~ dcscriprions, and 
when 
two different sentences produce the same description, it is difficult to 
distinguish them. 
Similar investigations used a sequence of sentences, each comparing two 
objects from a set of five according to some dimension (e.g. positjon, 
height, speed, weight). Subjects who knew that a single situation was being 
described had a better memory for the situation, but were less able to 
distinguish sentences they had heard from those implied by the described 
situation. Once again, the conclusion is that a non-linguistic representation 
is created when a meaningful text is understood. 
A final group of studies involved descriptive texts that were ambiguous 
or difficult to comprehend without an appropriate context. The context might 
be a picture of the described situation or a meaningful title for the text. 
In every case the results indicated the understander attempted to build a 
description of the whole situation, and when this was very difficult or 
impossible, comp rehension was low. (Related studies by Anderson, Reynolds, 
Schallert and Coetz [76] and Scallert [76] utilize texts, each with two 
totally different meanings. Evidence is obtained that one or the other 
alternative is almost always chosen for the entire text. The effect of 
subjects backgrounds on interpretations of these ambiguous texts Is also 
Gtudied, and a correlation between background and chosen interpretation is 
claimed. ) 
The principal conclusion of these studies is that the process of 
understanding text involves constructing a consistent unified meaning, wlli ch 
is not simply the set of individual sentenee meanings. It diffcrs in thht 
Page 8 
additional knowledge and orga11ir.n t i 011 i .; Int rod~iccd by tire uilderstander. 
1 Frcderikscn - 
Frederikscn [75p, 7Sb1 contin~~cs this line of thought, treating the 
undcrqtanding of n disccl\irsc as the. scmarlt i c knowlcdgs that is acquired in 
listening to the disco\rrsc. Thi r; knowlc.~l~~,~~ f o r~cqutrcd by processes that 
utilize prior knowlcdgc, cont cxt , c t c. nnd includcs inf armat ion which is 
inferred as well AS that explicitly prcs~b~~tctl. Fredcuiksen also argues thaf: 
the processes involved in undcrstnndinp, a discourse are direqtly related to 
the process limitntior~s of thc PL'OCCSSO~. The two processes that he discusses 
are overgcnera1i;cat ion and infert>~~cr. Ovc.rgcncirnlizetion is the discarding of 
detail informat ion resulting in n more p,th1'icr~1 concept. 'tie suggests that 
overgenera1,ization reduccs the amcwnt of infornlotion to be understood, thus 
reducing the processing Zond on the UII~C ~-stni\de Inferred information is 
infannation which is assunlcd to hc. t even if it- is not explicitly present. 
This process is clhimcd to reduce the pruccssing load by eliminating the 
necessity of completely andc rst lrndiny, t\vtary sctlteilce. FredarAksen performed 
neutral recall cxpcrimcnts nnd roct~ll cxl~rrin~~nts in wh.lch the context - the 
task assigwcd to the subjc.ct prlor to pli-:;c~itntion of the story - was changed,. 
His conclusfons arc that. ovc rgcncr;tliz,~r 1011 rind inferences are incorporated 
into the semantic reprcsi-ntnticrl~ of tl~cb story as it is ut~derstood, and that 
the context can inf lu~nccl tl~c nii)ou~~t of infc rcnciny, that is done during 
Page 9 
Frederiksen uses a simple, in£ ormally def incd semantic rep resentation 
in 
the above studies. It consists of set Inclusion, identity, and logical 
implication relationships that hold among the concepts of the text,. He 
indicates the need for& detailed model of the sciwntic representation, and in 
other work, Frederiksen [75cJ begin's to clef ine such a rep resentational scheme. 
He propcses the use of two networks: a semantic network and a logical 
network. The semantfc network comains the rep resent at ion of individual 
propositions while the logical network is composed of relations that hold 
between propositions that exist in the semantic network. ~rederiksen's system 
is probably the most elaborate yet proposed using the basic ideas of case 
grammar represented by semantic networks. He begins with the fundamental 
ideas of object and action hierarchies, although his are quite large end 
contain numerous distinctions. He proposes a system that distinguishes 
sixteen verb, cases, and a large numhc r of relations specifying states, 
quantification locati~n, manner, time, order, proximity, tense and aspect. 
His logical network consists of relations drawn from propositional logic plus 
causality, and he proposes several modal operators. The complexity of this 
proposal gives the impression of representational power, but it seems that 
Frederiksen is still primarily oriented toward the repreqentation of 
1 
indiyidual propositians. Although he shows the representation of simple time 
and causally ordered actions sequences, he has not yet demonstrated the 
adequacy of his system for text, in gerlerdl. And, he has utilized a number of 
elements of quantified modal logic without demonstrating their usefulness for 
modelling human understanding of d'iscoursc. His discussions do raise a number 
of interesting questions about representation, and certainly deserve 
Page 10 
consideration in the development of any set of semantic types and relat~ons. 
113 Kintsch - 
Kintsch [74] C reviewed by van Di jk [75bJ ) attempts to set forth a fairly 
complete theory of langudge understanding. Kintsch's basic rep resentational 
unit is the proposition, and he includes discus&ion of the usual difficult 
problems of definiteness (including generic versus specific distinct ions) of 
noun phrases, quantification, modality, implicetian and presupposition, 
location, time and tense. He proposes a text base which underlies discourses, 
but his revised and elaborated thought on this is discussed in Section 1.2. 
,Me, describes a model of discourse processing in which he makes a distinction 
between episodic and semantic memory, argues that the processing operations 
must be well defined (he suggests pattern matching and completion, abstraction 
and generation) and argues that recall is essentially a different pmcess from 
recognition in that recall requires input organization while recognition does 
not, 
Kintsch reports a number of experiments undertaken to test aspects of h;i$ 
theory. One conclusion that he reaches is that the semantic representation is 
partially independent of the actual input sentences. One kind of experiment 
performed to test this was the performance of a cammsn task by subjects who 
had read substantially the same material but in forms ot varying comnlexity. 
In other experiments he notes that sentences containing multiple propositions 
are much less likely to be recalled (completnly) than are single propositibn 
sentences. This also supports the idea of an undctlying representaticn. A 
Page 11 
related conclusion is that the agent of a proposition is the most likely 
case 
to be recalled. However, Kintsch did find that in passive sentences, the 
subject, rather than the agent, was the most likely to be kecalled. A final 
conclusion conce ms the presence of inferred in£ ormation in memory. Kintsch 
found that immediate recall showed some difference between implicit and 
explicit propositions, but that after twenty minutes or more the difference 
had disappeared and the two were indistinguishable to the subjeet. An 
interesting side result was the distinct difference in reading and recognition 
times for argumentative versus descriptive discourse. The argumentative 
discourse required significantly more time in both tasks.. Kintsch also 
includes reaction time experiments to study proposition retrieval, 
deterruin-ation of the truth or falsity of general propositions and processing 
of complex lexical items but these will not be discussed. 
1.1.4 Thorndyke - 
Additional investigationb on the role of inference in understanding text 
have been carried out by Thorndyke [76]. He uses compound sentences asserting 
a causal connection which is not familiar or obvious. For example: 
The hamburger chain owner was a£ raid that his love fox f rench fries would 
ruin his marriage. 
Sentences like these are imbedded in a meaningful text, and are followed later 
in the teFt by a continuation sentence which references the previously 
mentioned relationship. 
This continuation sentence might be neutral like 
Page 12 
He decided to see a marriage counselor in order to save his marriage. 
or it might encourage one particular eqlanatcrry inference like 
He decided to join weight watchers in order to save his marri ~ge. 
which strongly suggests the inference 
He is fat from eating Erench fries. 
Thorndyke uses recognition tests after the presentation of the story to 
compare inferences that have been reinforced by a continuation sentence with 
neutral and inapp rop riate inferences. He found the reinforced inferences much 
more likely to be recognized as part of She text than the neutral inferences, 
while recognition of inappropriate inf erenc~s was very unlikely. 
Thorndyke suggests that although this evidence indicates that inferences 
are made and stored as part of the understanding of a text, a more important 
implicaOion exrsts. This is that the role of inferencl ng is to aid in the 
4ntegretiot.r of new information into tble larger framewi3tk of the understanding 
o,f a text when no appropriate understanding could be obtained from only the 
explicitly stated information. 
1.2 The Structure Of Discourse Content 
Attempts to characterize the content of text necessarily lead to a 
structuring of that content. Simple one-dimensional representations in which 
propositions are connected only by co-reference or by time or causal ordering 
Page 13 
are 
adequate for only a restricted class of texts. 
The following studier all 
p ropose some type of multi-level representation of content. 
It seems appropriate to begin a discussion of memory structure with 
Bartlett [32] since he is often referenced as the originator of several 
cu rrently popular hypotheses. It should be remembered that Bart1 et t was 
concerned with memory in general and the phenomena associated with recall, and 
did, not restrict himself to the study of disc~urse. His pxinciple conclusions 
were that memories are not stored in isolated, static units, and that exact 
recall is very rare. \In fact, he often found cases of gross distortions in 
the recall of his subjects. He suggests that instead, memory is composed of a 
number of active, organized masses of past reactions and experiences, which he 
designates schemata, and a small amount of outstanding detail- Remembering is 
seen as a constructive process strongly affected by memories other than the 
one being retrieved. His ideas of organization beyond the sentence, of 
storage of something other than actual input sentences, and of active 
processes that modify text prior to its reproduction in recall are all 
currently en joying wjde acceptance. 
Page 14 
li).2.2 Crothers 
Crotbhers 1721 is concerned with the recall of short, expository 
paragraphs containin!:. material not likely to be familiar to his subjects, He 
presents results obtained wi tll paragraphs about hebulae. He proposes the 
existence of a semantic representation underlying any particular discourse 
which contains the rncal~lng of the discourge, but, does not reflect the details 
of the surface form of the teyt. Thus, a single semantic represent~tion might 
underlie numerbus actual discourses. His semantic representation assumes a 
conceptual taxonomy showing the relationship of each known concept to its 
superordinate concept (i.e. the familiar semantic hierarchy), but he does not 
provide details on this. He also proposes an additkonal set of hierarchies 
showing the relationships of the concepts in a particular discourse. For 
example, the following: 
means that a nebula is either seen or it is not seen, where all of the 
concepts are defined on the conceptual graph. Crothers does not carefully 
define his notation, so it is ~ot clear exactly what may or may not occur at a 
node, Primarily, concepts (words) or connectives are used (e.g. NEBULA. IS 
AND, OR, WHY). Since hc deals only with expository, descriptive material he 
finds no need to represent actions or time, Crothers performed recall 
experiments on two different versions of the same material, and arrives at 
three conclusibns. First, he concludes that the surface paragraph *s not a 
Page 15 
-fieant factor in recall, thus arguing for a surface-independent 
semantic 
repnsentation for the paragraphs. His- last two conclusions are negative 
xejecting hypotheses he had previously suggested. 
The first of these is that 
superordinate nodes in the meaning of the paragraph will be recalled more 
frequently than subordinate nodes. The second rejected hypothesis is that 
iaforaation not directly connected to the most superordinate pode of the 
paragraph I& lese likely to be recalled than information that is connected. 
Since his results do not support either of these hypotheses, Crothers suggests 
tbat other variablea, such as frequency of concepts, are probably also of 
irportance. 
1.2.3 Grimes And Meyer - 
Grimes [75] has proposed an extended case grammr supplemented with what 
he calls rhetorical predicates, which are higher-order predicates that take 
other propositions as their arguments. He subdivides rhetorical predicates 
into three groups: paratactic, hypotactic and neutral, Parataet ic predicates 
always take arguments of equal weight (i. e. no argument is subordinate to any 
other argument). Hypotactic predicates have one dominant argument, to which 
the others are ,subordinate. Neutral predicates may be used as either 
paratactic or hypotactic predicates. The following is a list of some of the 
rhetorical predicates proposed by Grimes, and their basic meaning: 
Paratactic predicrtes 
ALTERNATIVE o 
RESPONSE - 
Hypotactic predicates 
ATTRIBUTION o 
EQUIVALENT 
SPECIFIC 
EXPLANATION 
ANALOGY - 
Neutral predicates 
COLLECTION - 
Options, or 
Question and answer, problem and 
solution 
Gives qualities of the dominant 
proposition 
Gives restatement of the domiriafit 
proposition 
Gives more specific infomation 
about a general dominant 
proposition 
Gives an abstract ekplanation for 
the specific dominant 
proposition 
Gives an analogy to support the 
dominant proposition 
List of elements related in some 
unspecified way 
Causality, an antecedent and 
consequent 
Meyer [75] uses Grimes' rhetorical predicates and most of his case- 
grammar to create analyses of paragraphs she uses in recall experiments. For 
example: 
Page 17 
RE SP ON-S E 
PROBLEM 
COLLECTION 
need to gene rate elect ri c power 
p t~tect envi ronment 
rational utilization of natural resources 
SOLUTION 
breeder reactors 
would be the essentials of an analysis of a paragraph which stated that the 
need to generate electric power while protecting the environment and 
rationally utilizing natural resources is a problem which breeder reactors 
solves. Meyer indicates the hierarchical structure through the use of 
indentation. Since the rhetorical predicates in the above example were 911 
paratactic, nb indentation was used. However, the following example: 
finite reserves of natural resources 
SPECIFIC 
COLLECTION 
coal 
oil 
gas 
shows the subordination of the specific set of resources to the generaJ idea. 
 eyer's 
analyses of expository paragraphs using this scheme always follow the 
author's organizational st rvcture and always result in a purely 
hierarchical 
structure like those depicted above Meyer also recognizes class of 
sentences that can occur in expository paragraphs, but do not contribute any 
Page 18 
content to the passage. She calls these> signalling because their function is. 
to explicktly indicate the structure of the passage. She notes four types of 
signallins 
1. ~~ecificatibn of the structure (e. g. "Two options exist.'') 
2. Prematurely revealed information abstracted from the remainder of the 
passage (em go "The alternatives are solar energy, nuclear energy, 
geothermal energy aad laser fusion energy. ") 
3. Summary statements 
t t 
4. Pointer words (e.g. unfortunately", "an impdrtant point is") 
Meyer conducted recall experiments varying the position of certain material in 
the content hierarchy of passages, and including or omitting certain 
s ~gnalling in£ omat ion. Her principle conclusion is that material 
structurally higher in the content hierarchy is remembered substantially 
better than identical information placed lower in the hierarchy in another 
passage. She also notes that passages with identical structure but totally 
different content exhibit very similar patterns of recall at the higher 
leve$s. However, at the lower levels, the pattern of recall varied, 
indicating content dependence. In regard to signalling, she concludes that it 
has very little' effect at the top level, but seems beneficial at the middle 
levels. 
Page 19 
In regard to a theoretical explanation of why the highest ideas are 
recalled best, Meyer suggests three proposals and points out the weaknesses of 
each. First, lower propositions might be subsumed by hieher ones with the 
passage of time. Immediate recall experiments shw that the phenomenon occurs 
even without a time lapse, however. Secondly, at is possible that all 
propositions are stored but that retrieval is easier at the higher levels. 
She criticizes this proposal because even cued recall experiments were unable 
to retrieve the lower propositions. Finally, perhaps only higher ideas are 
ever stored. But her results show that the lose of lower level propositions 
begins immediately, but continues with time at a more rapid rate than for the 
higher level propositions. Meyer suggests that probably some combination of 
these processes is occurring, as well as other processes sensitive to the 
structure of the passage as a whole. 
1.24 Kintsch And Van Dijk - 
Kintsch and van Dl jk (van Di jk [74,7Sa,76], van Di jk and Kintsch 
[forthcoming] and Kintsch and van Di jk [76]) present a model for the 
organization of discourse as whole, and a number of (sometimes informal) 
experiments attempting to validate the model. Beginning at the lowest level, 
a discourse representation consists of a set of propositions. A distinction 
is made between two different types of representation, callCd text bases, as 
to whether all implied information is made explicit or whether it is left 
implicit. The 
notion of coherence is introduced, which is the property that 
distinguishes a discourse from a random set of sentences, Referential 
Page 20 
identity has been auggqsted as a major test of coherence, but it ie not an 
adequate definition (see discuesion of coherence and Bellert's proposals in 
Section 3). Kintsch and van Di jk argue that coherence is more accurately- 
captured by the requirement that each proposition of the text base be 
connected with one or more preceding propositions. Propositions are connected 
if one is a condition for the other, with the strength of the connectfon 
ranging from posaible to necessary. Thus, an explicit text base is one 
containing all of the propositions necessary for coherence, while an implicit 
text base is one with some of these propositinns deleted. The deleted 
propositions are those that can be assumed known or which normally would be 
inf~rred by the unde rs tande r. Different types of discourse would have 
different rules governing the deletion of propositions. For example, casual 
conve rgation would allow more deletion thari careful argumentation. 
The text base is organized hierarchically under macro-s t ructures, which 
are higher brde r propositions. Macro-st ructures may be related to their 
propositional arguments by a number of macro-rules. Four of these rules are 
suggested. =The first, information reduction, is a rule of generalization 
which would explain the existence of the macro-structure 
John is ill, 
which had as its arguments propositions like 
John has a feve r. 
John has the flu. 
The second rule, deletion, would explain the relationship between 
Peter saw a ball. 
and its subordinate propositions 
Peter saw a ball. 
The ball wSas blue. 
The third rule, integration, combines a central event with its normal 
pre-conditions, results and component actions. Thus 
John went to Paris. 
might have as its arguments 
John took a cab TO the station. 
He bought tickets. 
He went to Paris. 
The fourth rule, construction, explains the relationship between a complex 
fact and its component parts. For example, the macro-structure 
Peter built a house, 
might have as its arguments the event sequence 
Peter laid a foundation. 
Peter built walls. 
Peter built a roof. 
In general, two conditions hold fur all macro-structures. First, the 
Page 22 
macro-structure must be semant lcally implied by its micro-structure (i.e. its 
propositional arguments). The exact meaning of the term implication is not 
clearly stated by Kintsch and van Dljk, They sometimes refer to it as 
entailment and treat it as a formal logical relation, but at other times say 
that it is not logical in the strict sense. In any case, it is clear that 
there are semantic rules or structures that allow creation of macro-~tructures 
such as those above. Thc second coqdition is that the sequence of 
macro-st ructures representing a coherent text base must itself be coherent. 
Recalling the definition of koherence, this implies that no macro-structure 
may delete information contained in its micro-structure which fs a c~ndition 
for another macro-structure An important consequence of this is that the 
macro-structures 06. a ttext, taken by themselves, form a coherent summary of 
that text. 
A final component of a tept representation is the specification of the 
£om or structure of the discourse, but discussion of this proposal will be 
postponed until Section 2. Their investigations are primartly with narrative 
text, and they call the narrative structure a schema. 
Kintsch and van4i jk shggcst that the concept of -narrative st-ructure 
coabined with the idea of macro-structures leads them to the follawing 
hypotheses regarding co~np rehension and recall of neriative- Ffrs t 
macro-structures are primarily what is stored when a text is understood. 
Recall uses macro-structures as a starting point in retrieval, and summaries 
directly reflect the nlacro-st ructures. Secondly, since macro-st ructures are 
essential to comprehension, they must be constructed at the time of reading. 
Page 23 
Finally, a narrative schema is necessary for the organization of the text 
representation. Kintscll and van Dijk have done a number of summary, recall 
and other experiments to test these hypotheses. Most use a 1600 word text 
from - The ------..-- Decamcrot~. In recall experiments, the propositions corresponding to 
the macro-st ructures were found to be recalled most often, and were very 
unlikely to disappear in delayed recall (nine days later). Xn contrast, many 
other propositioi~s recalled immediately were omitted in the delayed recall, 
The propositions most likely to be recalled have the following functions: 
1, Introduce main characters 
2, Give major goals of characters 
3. Describe actions leading to these goals 
4. Describe events occurring to these characters leading to or from 
their goals 
while propositions having the following functions are likely to be forgotten: 
1, Sctting description 
2. Prepnratoq act ions 
3. Mental actions 
4. Comp oncnt act ions 
Page 24 
5. Probable consequences of an act ion or an event 
6. Meta-narrative statements by the author 
Summarizing experiments showed thbt summaries of the story *wet.e very much 
like delayed recall. When a summary was written after presentation of each 
successive part of the story, followed immediately by a complete summary, 
propositions were included Fn the partial summaries that were omitted in the 
final summary. A flnal group of recall experiments compared recall of a 70 
word paragraph in isolation, to cued recall of the same paragraph imbedded in 
ran 85O word text. Surprisingly, the recalls were almost identical But when 
the entire text was recalled, the reproduction of the particular paragraph wa8 
much smaller and less accurate than in the first two cases. Their conclusions 
from these results axe that the recall of small amounts of text is a different 
process from the recall of a long text, that recall of a long text relies . on 
the macro-structure of the text as a means of organizing the text, that the 
micro-st ructure is forgotten much more easily and that summarizing is based on 
the macro-stcructuxe. 
Finally, some experiments were done in which incorrect summaries were 
presented prior to the presentation of the story. These were found to have 
practically no influence on the final understanding o'f the story, and the 
subjects were unable to accurately recall the incorrect summaries. 
Page 25 
1.3 
Some Conside rati on's Regarding Memory Experiments 
Before attemptiqg to describe an informal madel based upon these 
hypotheses and experimental resulta, some general criticisms of memory 
experiments exp ressed by Spiro [ 751 deserve consideration. spire's principle 
argument is that recall consists of active reconstruction processes, rather 
than passive processes which merely reproduce that which was stored st the 
time of comprehension. His general position is thus much like that of 
Bartlett, and he Observes that almo3t all experimentets since Bartlett have 
failed to replicate his findings of significant errors in recall. As a 
result, psychologists have tended to concent rate on the process of 
understanding, or construction, and have treated recall as somethihg of a 
simple retrieval process. Spiro argues that for reconstructive errors to 
occur, it is necessary that the input be understood in terms of some 
preexisting schema. Any text that results in the creation of a new achema 
would not be subject to the same kind of effects of previous knowledge. The 
interference caused by the pre-existence of schemata is even more pronounced 
if other uses of the schema are made between the time of comprehension and the 
time of recall. Spiro also points out that any experimental subject would be 
unlikely to integrate material read or heard in an experiment into his general 
knowledge since its truth and source are unknom. This coupled with the 
desire to perform well, could easily result id a considerably modified process 
of understanding text in experimental settings.. Spiro performed a number of 
recall experiments which support his hypothesis. He used text about 
interpersonal relationships and instructed t31e subjects that the experiment 
was 
concerned with their reactions to the incidents described, thus maximally 
Page 26 
involving their pre-existing struerurea in the understanding process. Spiro 
found substantial errors in the subjects recal,l~, which he explaine as the 
effect of interaction wi ti1 prc-cxisting structures for interpersonal 
relationships. (Kintsch and van Dijk [75] found similar results using stories 
from the Bible.) Spiro conclr~des that most text recall experiments have not 
shown these effects because they did not meet the required conditions, but 
frequently involved the prcscntation of material that was totally new and 
which was kept isolated by the subject. These criticisms not only are 
relevant to the question of construction versus reconstruction, but also to 
inferring discourse organf zation f rofn recall experiments, and indicate that 
one must be very cautious in generalizing from the results of text recall. 
experiments due to the many intewctions between the subject's knowledge, the 
experimental setting, the type and content of the text and the 
extra-experimental effects (or lack of them) in long term experiments. 
1.4 Discussion And Conclusions 
A great deal of agreement is seen in the preceding studies. The idea 
that the meaning of a text differs from the meaning of its individual 
sentences, in that prior knovledge is used to infer implicit information 
during comprehension, is generally accepted. The idea that this meaning must 
be well structured, or coherent, is also expressed directly or indirectly by 
most of the researchers. However, explicit disagreement exists on both the 
nature of content structure, and its relationship to recall. Crothers and 
Meyer disagree as to whether a neutral struccure reflecting pre-existing 
Page 27 
knowledge is built, or the author's structilre used. Kintsch and . van Dijk 
essentially use both approaches since the time ordering in narratjve is author 
defined (although the concept of time is pre-existent), but the 
macro-st ructures which hierarchically organize the text are pre-existent- 
Meyer, Kintsch and van Dijk all agree that propositions higher in the 
structure are more likely to be recalled, but Crorhers evidence did not 
support this conclusion. Clearly, this hypothesis, if true, could be 
confirmed only if it is tested against the appropriate content structure. 
Spiro's comments regarding the use (pr laah of use) of pre-existing structures 
are relevant, In the case of unfamiliar expository material, such as that 
used by Crothers and Meyer, it seems reasonable to suppose that the author s 
organization would be used to organize the semantic representation due to the 
lack (or n,on-use) of any appropriate pre-existing content structures. Thus 
Meyer's results suppart the importance of propositional level, and Crothera 
results do not- Ylowevev, a familiar structure, like an action sequence, is 
understpod in the. same way regardless of presentation setting. Thus, Kintsch 
and van Di jk found that the role of a proposition in the pre-existing 
structure (the macro-structures) was primary in determining its importance, 
It should be noted that even Meyer found variations in the recall of midclevel 
propositions that she could not explain by reference to the author's 
organization. It is plausible to assume that these variations were caused by 
individual differences in prior knowledge, resulting in differing content 
structures. 
Page 28 
In light of thts, it seems plausible to propose a general model that 
distinguishes the following component processes bf text understaading 
(realizing that like almost all distinctions, such as syntax versus semanticq, 
they are really ~nadequate, but useful, oversimplifications). First there is 
the process of comprehensioh of the surface structure which builds a 
representation that is independent of the surface form of the text, This 
representat$on contains as much inferred information as is necessary to meet 
some minimal level of understandability, or coherence. This representation is 
organized hierarchically by higher level content st ructures ghich summarize or 
generalize over a larger amount of detailed in£ onnation. These higherlevel 
structures would normally be pre-existing (i. e. known to the understander) 
although a text could result in the creation of new structures. Different 
types of texts could certainly differ in the complexity of each of these 
tasks, explaining the variation ,in di f f iculty of understanding. 
The process(es) explaining forgettfng, and the integration of new 
information into prior knowledge, operates on this text representat-ion. This 
process is poorly understood, and there are no well defined hypotheses about 
its operation. It appears to operate primarily on the content structure of a 
text. So, for example, Spiro's subjects integrated information from the 
stories they read into their general expectations about love and courtship, 
and lost the specifics of the tpxt, even recalling it with radical 
mods f i,ca t ions This pxocess would explain the almost total loss of detail 
information with time, as well as the interactions and distortions caused by 
previous knowledge. 
Page 29 
Finally, recall is a retrieval process which operates on the current form 
of the rep resentation. If that form is relatively unchanged since 
comprehension, recall may be almost distortion-f ree. But if significant 
changes have occurred, recall will be distorted. As suggested by Kintsch and 
van Dijk, summary is seen as recall to a controlled depth. Retrieval accesses 
the higher-level organization first, and uses these structures to access more 
detailed information. As retrieved information is generated as output text, a 
test of coherence would be made. If something has been forgotten or changed 
so that a necessqry explanation or link is missing, the retrieval process 
would supply a probable piece of information from the general pre-existing 
structures. This, along with modifications from integration into prior 
knowledge, accounts for the reconstructive aspect of recall- 
This general model does not account for the recall of very specific 
surface details (e.g. words and constructions used in the input text) but as 
suggested by Kintsch's work, it seems best to treat this as a process separate 
from the general understanding of text. One other inadequacy of the model is 
that it specifies complete comprehension of all input propositions, 
including 
assignment to an appropriate role in a higher level structure. Frederiksen 
suggests that overgeneralization is a loss of detail which reduces the 
processing load. The idea that partial processing is sometimes done on input 
information is also suggested by Bobrow and Norman [75]. They describe such 
processes as resource-limited. This kind of processing seems to be indicated 
by facts such as loss of detail in immediate recall and the subjective 
impression of partial comprehension in a number of situations (such as reading 
verg rapidly, reading while partially distracted or reading very complex 
Page 30 
material). Wowever, while the idea of discarding comp rellended in£ ormation* 
could be described computationally, selective comprehension ia a process that 
requites further investigation. 
One of the principle contributions of computational investigations with 
this sort of model will be specification of the processes of comprehen~ion, 
which can only be done by specifying the nature of pre-existing memory 
structures. Meyer's work is an example of the use of undefined basic 
stmctures. She criticizes Crothera for requiring two graphs to represent a 
text one of which ia his Eundame~ital hierarchy of kndwn concepts. She, 
instead, simply doesn't define any of her predicates first or higher order, 
apparently taking them as primitive. Assuming that one attempted to define 
even the argument roles f~r the rhetorical predicates, enormous difficulty 
would be encountered. Advocates of case systems have always had difficulty 
defining the exact role of cases, or delimiting the class of entities that 
could fill a case role. Consider how much more difficult it would be to 
define what pairs of things could exist in the RESPONSE oF COVARIANCE 
relations, or exactly what the charactexistics of these arguments are. 
Kintsch and van Di jk attempt to set out much more well defined structures, but 
even they have little to say about the set of structures that must exist, what 
information they must contain and how the camprehension algorithms use the 
text and the structures to build higher-level structures. The section on 
computational models will consider algorithmic attempts at text understanding, 
and should clearly indicate the complexity of these issues. 
Page 31 
2.0 
THE FORM OF CONNECTED DISCOURSE 
The general model of text understanding suggested in the last section 
describes the structured content of text representation. This secaan 
contains work investigating a similar, yet distinct, type of structure - the 
abstract, underlying form of a text. (Not everyone would accept this 
distiaction, but it seems analogous to the distinction between the syntactic 
structure of sentences and the semantic structure of their underlying 
meaning.) Much of this work has been done by cultural anthxopolopists or by 
those interested in literary analysis. Little of the traditional work ha8 
been done with cognitive or computational models in mind, and hence tends to 
rely on intuitive understanding of the terms. This less computational work 
will be presented quite briefly, since the presentation is intended to suggest 
the kinds of results that have been obtained, rather than the details of these 
conclusions. It is necessary to define these ideas in a computational way 
that can be integrated into the models of text understanding, and some recent 
work has begun this task. 
2.1 Structural Analysis Of Text 
Most modern work in the structural analysis of texts has roots in the 
work of the Russian structuralist, Propp [68]. Propp was concerned with the 
form of Russian folktales. and developed a method for describing the 
similarities he found in a corpus of 100 folktales. The major structural unit 
that he developed , he designated a function. These functions act like 
meta-actions in 
that they describe major events or event complexes that were 
Page 32 
repeatedly found in the tales. 
For example, Absentation is the function that 
describes the situation in which one ~f the members of the family absents 
himself from home, and Trickery is the function which describes a villain's 
attempt to deceive his victim in order to take possession of him or his 
belongings. Propp fourid thirty-one such functions to be adequate to describe 
the events in the tales that he studied. Importantly, the ordering of the 
functions was found to be fixed. Some functicrns are optional, but if they am 
present their position is fixed with respect to the other ftlnctions. All of 
the functions are defined in terns of actors that participate in the 
situations they describe. P.ropp did some analysis of the restrictions on the 
assignment of various characters to these functional roles. He notes that the 
aame character is likely to play a fixed set of roles, which he designates a 
sphere of action. These spheres of action include such familiar story roles 
as hero, villain and false hero. Finally, Propp notes that a single folktale 
may be composed of a sequence of basic tales. He designates the basic tale a 
I1 
move". Thus, a folktale ,may be a one-move or a multi-move tale, with each 
move conforming to the definition of a tale. These rules capture the 
similarity of the folktales which could differ widely in the details of 
characters, setting, specific actions, etc. It Is clear that Propp's rules 
could be expressed formally, and that is what Klein et a1 f?4j did by writing 
a computer program to generate the function sequences of a number of 
folktales, using Prupp's definitions. 
The ideas of Propp have been refined and generalized by Barthes, Bremond, 
Greimas and Todorov (their work is not generally available in English but is 
reviewed in van Di jk [72] and Weinolci (72) ). Their work has been directed 
Page 33 
toward the redefinition of Propp's functions as propositions, and of actor8 as 
case relations of these propositions. 
They have also attempted to 
generalize 
the 
functions by making them less specific, and to apply them to other forms 
df texts. However, Hendricks [72] demonst rates the flexibility of Propp's 
original functions by using them to analyze part of the Bradbury novel 
Something Wicked This Way Comes. 
Chafe [72, 751 discusses the process of "verbalizatiou", by which he 
means 
the translation of non-verbal knowledge into verbal output. One of the 
proposed component processes of verbalization is breaking the knowledge into 
smaller chunks according to patterns, which are called "s~hemata'~. These 
schemata are grammar like descriptions of possible verbal sequences. Thus, he 
notes that a certain class of stories will always be of the form: 
PLOT + MORAL 
Chafe also mentions such schemata as TRICK and VISIT as occurring in the 
stories he has analyzed. He explicitly states that, at this time, schematic 
analysis of a story is done by "imagination abd intuition". He discusses 
aspects of the schematic analysis of several simple tales, but these are 
mostly illustrative rather than complete, in any sense. 
Colby [73] analyzes Eskimo folktales, identifying "eidons", which are 
similar to Propp's functions, and producing a grammar capable of generating 
some of these tales. Colby observes that most native speakers are probably 
not very familiar with story grammars of this type (or the knowledge they 
represent), since only a few individuals are able to generate such tales 
Nonetheless, it seems reasonable to assume that the same type of structure can 
Page 34 
be found in more conventional stories 
Although most work in this area has involved analysis of some form of 
traditional narrative, Labov and Waletzky [67] present analyses of oral 
vprsions of personal experiences. Even in this informal type of narrative, 
they find five regular components: an orientation, a complication, an 
evaluation, the resolution and a coda or moral, The last two components are 
optional. A totally different type of material is examined by Becker [65] who 
discusses patterns in expository paragraphs. 
All of these analyses rely on the intuitive understanding of the analyzer 
to grasp the meaning of the structures from infomal descriptions and to find 
these structures in the text being analyzed. For this reason, the particular 
structures that have been suggested are of less interest than is the general 
result that quite regular high level patterns are found in texts of many 
types, which may be characterized independently of the specific textual 
content. 
2.2 Kintsch And Van Dijk 
In addition to the content macro-structures (discussed in section 1.2.4) 
Kintsch and van Di jk [75] and van Di jk [75a, 761 discuss even higher level 
"super-structures", used to organize the content of a text according to the 
type of the text, which are derived from the work in structural analysis of 
texts. They limit their work to narratives, which they define as a specific 
type of action discourse. Narrative categories such as Episode Setting, 
Page 35 
Complication, etc. 
are superstructures under which are organized either more 
narrative categories or the content macro-structures of the text. Van Dijk 
[75aJ presents a fairly complete example. It includes a 1600 word text from 
The Decameron, a propositional analysis of the story, the content 
macro-st ructures of the story and the narrative categories under which the 
macro-st NCtures are organized. The following abbreviated fragment should 
give an indication of the kind of analysis being proposed. 
Lllllllllll----lo---*~~ 
I I 
STORY MORAL 
I 
----1---11-1 
I 
I I 
EPISODE1 
I 
O-WIIUIIII-I-I-LI 
I 1 
SETTING HAPPENING 
I I 
,--I--------,-- 
I 1 
COMPLICATION RESOLUTION 
I 1 
I 
Landolfo loses his fortune 
i 
(detail propositions about this loss) 
Some expetiments were done to test the role of the understander's notion 
of narrative structure. An Apache myth and three Decameron stories, which 
were of comparable difficulty at the sentence level, were presented. They 
differed in that the Apache myth had an organization not familiar to the 
subjects. There was much greater variety in the propositions different people 
Page 36 
used in tllc recall' of the myth than in the other three stories. Another' 
experiment comparcd recalls of subjects who had read a normal story to those 
of subjects who had rend scrambled versions of the same story. The recalls 
were indistinguishable by judges. Both of the~e groups of expeiiments 
indicate the active role of one's expectations about the fom of a discourse 
in arganizir~g the representation of that discourse, When this is impossible, 
as in the Apache myth, recall is less organized and more random. 
2.3 Rumelhart 
Rumelhart (74, 751 proposes a story grammar that is capable of producing 
structures similar to those described by Kintsch and van Dijk (section 2.2) 
His grammar contains rules like the following: 
1. STORY -> SETTING + EPISODE 
2. EPISODE -> EVENT + REACTION 
3. REACTlON -> INTERNAL-RESPONSE + OVERT-RESPONSE 
Using these rules, he describes the syntactic structure of simple stories. 
The following fragment illustrstes the types of structures derived (ignoring 
the parenthc?sizcd predicates for now): 
STORY 
I 
-I------LILICILICIIIIII 
I I 
SETTING (AND) EPISODE (INITIATE) 
I I 
-I--------------------- 
I I 
EVENT (CAUSE) REACTION (MOTIVATE) 
I I 
(propositions describing ~argie's I 
balloon being popped ) I 
I 
L-1--------------11--------- 
I I 
XNTERNAL-RESPONSE OVERT-RESPONSE 
I I 
Margie is sad Margie cries 
Associated with most syntacrtc story rules are semantic rules which are used 
to generate a corresponding semantic structure for the story, The semantic 
rules for the syntactic rules given above are: 
1, ALLOW (SETTING, EPISODE) 
2, INITIATE (EVENT, REACTION) 
where the semantic predicates are intended to mean what is suggested by 
their 
names (e.g. the SETTING ALLOWS the EPISODE to occur). The figure given above 
contains the appropriate semantic predicate, in parentheses, at each node in 
the syntactic structure. (Rumelhart uses a completely separate semantic 
structure,) 
Page 38 
Rumelhart digcusses some elementary summarization rules that operate on 
the semantic structure of a story. Some illustrative rules are: 
1. MOTIVATE (thought, response) => response 
2. INITIATE (X, Y) => Y 
3. INITIATE (X, Y) => Y (because 1 when 1 after) X 
These rules could produce the either of the following summaries (for the 
EPISODE in the story fragment given above): 
1. Margie cried. 
2, Margie cried because her balloon had popped, 
Rume.lhart8s ideas are intended to be primarily illustrative ("a tentative 
beginning of a theoryu), and deal with very simplified stories. They do argue 
clearly for the notion of both a syntactic and a semantic structure for the 
representation of a story's meaning. The syntactic structure allows 
production of the semantic stucture, which is justified by its usefulness 
in 
p roducing summaries. 
2.4 Mandler And Johnson 
Mandler and Johnson [77] discuss an extended form of ~umelhart's story 
grammar, which they propose as a basis for text recall studies. The extended 
Page 39 
grammar allows multiple episodes and does not utilize the additional structure 
containing the semantic interpretation which Rumelhart suggests. Mandler and 
Johnson feel that this second structure is unwieldy and frequently redundaat., 
Conjunction, time-ordering and causality are included as categories in the 
grammar. They observe that restrictions imposed on a story by this type of 
grammar include allowing only a single protagonist per episode, and 
prohibiting the realization of higher level nodes in the actual text. They 
demonstrate the utility of their grammar on several stories, including "The 
War of the Ghosts". They also note that a transformational component remains 
to be developed which would provide for deletions and reorderings of the ideal 
story structure- They discuss a set of predictions for recall studies, which 
are suggested by this definition of story. structure. These include: 
1. Accuracy will be better as the story is complete and ordered as 
defined by the ideal structure- 
2, Elaborations will be poorly recalled. 
3. Optional nodes' realizations will be less well recalledm 
4. 
Causality will be better recalled than simple time-ordering. 
5. Inversions should be fewer as the story is closer to the ideal 
structure- 
6. Omissions and violations of the ideal structure will be reflected 
by 
additions and distort ions. 
Page 40 
Unfortunately, the only experimental work reported is a cowrison of recalls 
of students from first grade, fourth grade and university levels, on a set of 
stories which are very near the ideal structure. Hopefully, this is only the 
beginning stage of validation of the proposed model* An interesting result 
from this experinlent is the probability of recall of a proposition as a 
function of its role For university students, Settings (introduction of 
characters) and Beginnings (initiation of event sequences) were best 
remembered* Attempts and Outcomes (which together formed Goal Paths) were 
next best recalled* Reactions (mental events) and Endings (which were usually 
either very predictable or omitted in the stories used) were least well 
recalled. 'These results are in harmony with the results reported by Kintsch 
and van Dijk (which are discussed in Section 1.2.4). 
2.5 Thorndyke 
Thorndyke [771 attempts to validate the basic notion of underlyihg text 
form as a necessary part of a text understanding. Re applies a story 
structure grammar, which is nearly identical to that proposed by Rumelhart, to 
two stories, "Ci rcle Island" (analyzed by Frederiksen [75b] ) and "The Old 
Farmer" (given by Rumelhart [74J), each pf which results in a straightforward 
hierarchical structure. The initial rule of the grammar is 
STORY -> SETTING + THEME t PLOT + RESOLUTION 
Thorndyke wishes to test the effects of violating this rule. He does so by 
performing recall, summary and recognition experiments on four vari-ations of 
Page 41 
the stories. Thesc are a normal Form of the story, a story with the Theme 
moved to the end, a story with the Theme omitted and a descriptive version 
which omits causal and temporal continuity. using only stative or single 
action sentences. In comp rehensibili ty judgements and recall tests, the 
normal form story was best, the Theme-after form next, with the last two cases 
more dependent on the particular story. Recalls of the Theme-after passages 
alao showed a strong tendency to relocate the Theme to its normal position, 
near the beginning of the story. There was also a tendency in the more 
structured passages for higherlevel propositions to have a higher probability 
of recall. Summarizations showed that, of the recalled propositions, the 
higherlevel ones were much more likely to be included as part of the summary. 
Summaries of the descriptive presentations yielded a wider selection of 
proposi,tions. In recognition tests, the more structured passages produced 
more recognition errors when the test proposition was consistent with the 
story, while the less structured passages produced more accurate recognition. 
In attempts to separate the effects of content and structure, 
Thorndyke 
tested the four stories obtained by using each of the original plots with two 
character / object sets. He found that the Farmer plot was always more 
comprehensible. He also found that presentation of a second story using the 
same plot structure improved later recall of the first story. 
Thorndyke concludes that text structures are basic to comprehension, 
although some are inherently more complex than others. The distinct ion 
between structure and content is supported by the positive effect of repeated 
structure prior to recall. And finally, the structural position of a 
Page 42 
p roposition has significant in£ hence on both recall end summarization. 
2.6 Text Grammar 
"Text grammar" is the designation used for another currently active line 
of investigation into the structure of text. The fundamental ideas of text 
grammar are discussed by van Dijk [72], Ihwe [72], Kummer [72], Petofi [72] 
and Petofi and Rieser [73]. (Unfortunately, a great portion of the work is 
not available in English.) The principle concern is the formulation of 
generative descriptions of texts, rather than individual sentences, and is 
usually approached using a model similar to that of the generative semantics 
school of linguistics. Van Dijk suggests the following five components of a 
text grammar: 
1. Semantic formation rules for the meanings of texts, as a whole. 
2. Transformations which operate on this text meaning. 
3. Transformations which produce a sequence of sentential semantic 
rep resentations from the text meaning, 
4. Transformations which produce q sequence of sentential syntactic 
representations (including lexical i terns) from tho sequence of 
sentential semantic rep resent at ions. 
5. Rules pairing syntactic representations with morphonological 
representations. 
Page 43 
nc- cqonenta would operate in the order given to generate a text* He 
cqhasires this is only the outline of a theory, and that most of the hard 
dttdls rerain to be specified. A significant amount of work has been done on 
ccqoaents 1 and 2, formall,y defining represenmtions of text. This has 
in&ded numems examinations of the use of some extended predicate calculus 
u a basic representation (primarily addressing the meaning of individual 
satemces and objects) but has also led to psychologically motivated 
imestiwtioas 
(as 
discussed in Section 1.2.4) and to studies in the general 
stacture of textual types (as discussed in Section 2.2). Discussions of 
capmmnts 3 and 4 have been mostly descriptive, indicating phenomena that 
rst be accounted fo~ without actually defining them. Many of these phenomena 
seem best characterized as operatdons operating on two or more underlying 
propositions- These phenomena include anaphora (including pronominalization 
md article selection), sentence stress, contrast, use of clausal 
caajtanctions, and verb tense determination. Of course, certain literary 
styles ma alter the tt normal" rules, choosing repetition over 
prooariaanzation omission of causal connect ion, repet ition of certain 
#.tcnce forms, etc. Operations which occur on a single proposition include: 
.amtic transfonations, such as personification and the use of metaphor; 
wmctic traasformations, such as inversions and other stylistic operations; 
.rd pbaaological transformations, such as those producing rhyme, alliteration 
.od meter. The set of realization rules used to produce a particular text 
fm its generated underlying structure is assumed to be limited by the type 
of tart that being generated (e.g. mystery, narrative, etc.), which is 
Lpdicated by some underlying abstract type marker. The selected rules will 
Page 44 
also determine the textual characteristics usually referred to as style or 
literary merit. 
It seems that any complete text understanding model must have a set of 
rules relating the underlying structure of the text - individual and groups of 
propositions together with global aspects of the text - to its surface 
realization. Although using a generative model, text grammarians are 
concerned with identifying and characterizing these relationships, and the 
results of their investigations should be of definite value in describing text 
understanding, once their research has reached the point of clearly defining 
the nature of these surface phenomena in terms of the meaning of text. 
2.7 Discussion And Conclusions 
The principal hypotheses of these investigators is that there is are very 
high 
level structures which organize content structures in the representation 
\of a text. This section was begun with the warning that the distinction 
between these two different types of structures might be difficult to define. 
It seems clear that although each proposal that has been considered claims to 
be concerned with this high level structure, some confusion exists. The 
interactions between the abstract structures and the content structures are 
not characterized at all. Thus, Kintsch and van Dijk do not describe the 
restrictions on what type of events may realize a CON.PLICATION category. 
Similarly, Rumelhart does not explain how the correct semantic rule would be 
selected when a choice is possible (e.g. Two sequential EVENTS may be related 
by either CAUSE or ALLOW). Mandler and Johnson, and to some extent Thorndyke, 
create additional confusion by mixing syntactic categories with categories 
that would usually be considered semantic. Thus, conjunct ion time ordcdng 
and causality may appear in the "syntactic" structure of a story. Although 
this difficulty in distinguishing between the two types of structure. 
suggest that there is no distinction. the evidence provided by the exirrtenca 
of describable claesea of texts seems to indioate that some type of high level 
structures doexist. It remains to clearly define the nature of thasa 
struckures, and to explain their relationship with content structures, without 
repeating the fallacy of creating a multitude of subcategories (e.g. wny 
different EVENT subtypes) which are alleged to be purely syntactic categories. 
These structures fit neatly into the model discussed in the conclusion of 
the content section (section 1.4). They are the highest level of structure, 
under which content structures are organized. This is consistent with the 
hypothesis of Kintsch and van Dijk (discussed in section 1-2*4) that the 
structural description of a text is one of the important aspects of its 
representation. They specifically suggest that these st mctures organize 
macro-structures, which provide the content organization of the text. This ia 
quite ~0nSiSteht with the results of structural analysis which, as Hendrickc 
[731 points out, ordinarily uses a summary or synopsis, not the actual text. 
as a basis for analysis. (Recall that macro-structures are proposed ae the 
basis for summary generatfon). Of course very simple or very navel texta 
might not have any structure of this type, scnce no appropriate structure 
would be pre-exis'tent. Of course, the problem of learning these structure is 
as difficult to explain (if not more digficult) as the learning of content 
struct-uree. Such acquisition is simply not well understood. The work of the 
Page 46 
text grammarians suggests the complexity and Alversity of linguistic phenomena 
that nsay be related to underlying textual representnt:ion. This work also 
ruggsats that many aspects of a text, other than its literal meaning and 
general type, may require explicit representation. Although a eound 
computational representatian of atyli~tic aspects of text0 may seem a distant 
goal, it is still important to retain it as part of the goal of understanding 
text, and to realize that the richness of human use of language will be only 
partially accounted for without this component 
Page 47 
3.0 
COMPUTATIONAL MODELS OF TEXT UNDERSTANDING 
This section will survey computational work on the process of text 
understanding. Computational 
models will be seen to be of value in at least 
two ways. FLrst, many of the features postulated in other models will be 
found in these computational models, frequently motivated primarily by 
computational concerns. This tends to support these hypotheses. Secondly, 
many points passed over very quickly by other workers are seen to offer 
formidable problems when one attempts to set out a full computational 
description. The discovery of inadequate description is essential in 
develeping sound models. "Computational" will be taken to mean any model that 
is actually programmable or that is formally defined. The work selected for 
discussion in this section is generally treated in more detail than that in 
previous sectibns. The reason for this is that if the complexity of 
rcomputationally specifying certain representations and processes is to be made 
clear, it is impossible to treat these matters cursorily. This, in turn, has 
necessitated greater selectivity on the part of the author in choosing work 
that seems to best convey certain types of problems. 
Bellert 1701 attempts to define the notion of the coherence of a text, 
and in doing so suggests some of the same conclusions reached by computational 
linguists in their text understanding work. Coherence refers to the property 
of a set of utterances which make it a connected discourse rather than a 
random collect ion of utterances. Referential identity has of ten been 
suggested 
as an indication of coherence, but it is clearly insufficient. 
For 
example: 
Page 48 
John drinks a lot of coffee. John married a blonde. 
Both sentences refer to John, but do not form a coherent discourse. 
Furthermore, referential identity is unnecessary, Consider the two sentences: 
There has been a drought. People are starving, 
These are coherent without any explicit referential identity. Bellert defihes 
a coherent text as one in which the semantic interpretation of each sentence 
is dependent on the semantic interpretation of the preceding sentences. The 
semantic interpretation of a sentence ia defined as the set of conclusions 
that can be inferred from that sentence. She suggests that there are two 
types of conclusions that may be drawn: 
1. those drawn only from knowledge of the language 
2. those drawn from knowledge of the world 
Both types of conclusions are absolutely necessary in understanding text and 
may be appropriately drawn when the coherence of the text requires. She 
concludes that "an utterance has meaning only in the entire context and 
through our knowledge of the world". The rest of the work discussed in this 
section wd 11 strongly reinforce these conclusions. 
The discussion is divided into two sections. The first contains earlier 
work that supports the claim that world knowledge is essential to the 
understanding of text. The closely related problem of making implicitly 
conveyed information explicit is also a principle concern. The second section 
Page 49 
describes later work in whlch the or~anization of world knowledge is 
recognized as a critical question for text understanding. 
3.1 The Necessity Of World Knowledge 
3.1.1 Charniak - 
Charniak's work [72, 74, 761 is probably the first attempt to set out in 
a well defined fashion the dimension of information processing that must be 
carried on in the understanding of a story. The tarn "understanding" is 
necessarily vague, but Charniak suggests an intuitive definition. Consider 
the following story fragment: 
Fred was goipg to the store. Today was Jack's birthday and Fred was going 
to get a present. 
It should be clear that if a human had read and understood this fragment that 
he would be able to answer such questions as 
1. Why is Fred going to the store? 
2. Who is Fred buying the present for? 
3. 
Why is Fred buying a present? 
Charniak claims that a semantic representation of this fragment (i.e. an 
understanding of it) should explicitly contain the answers to ordinary 
questions such as these. Note the important point that this type of 
Page 50 
understanding could be attained only through the use of general knowledge of 
the world, along with the explicit statements of the text. Knowledge required 
to answer the above questions would include such facts as: 
1. A person having a birthday is likely ta receive presents. 
2. Presents are of ten bought at stores. 
So Charniak's goal becomes outlibing an answer to the question of hw 
common 
sense knowledge may be incorporated into the process of understanding natural 
language. A closely related goal is the determination of how much knowledge 
of this type is required by the basic problems of natural language, such as 
the resolution of pronominal reference, 
Charniak breaks the problem of processing natural language into two 
parts. The fir~lt part is the translation of natural l'anguage into a farm that 
is convenient for use in making deductions. This internal representation is 
like the understanding that a person would be capable of obtaining without 
cqntext. The internal representation is a canonical propositional form, 
Thus, either of the sentences 
Jack caught a cold. 
Jack came down with a cold, 
might be represented by the proposition 
(BECOME-SICK-WITH JACK COLD) 
which represents the explicit meaning of eithcr of the sentences. The second 
Page 51 
part of the problem is what Charniak calls Deep Semantic Processing (DSP). 
This is the processing that makes explicit the implicit 
information conveyed 
by the story. 
Charniak elects to examine only DSP. 
The function of the system that Charniak proposes is to read a story 
which has already been translated into internal representation and output a 
data base of propositions that explicitly represent all of the information 
conveyed by the story. In order to do this, the system uses commdn sense 
knowledge to make implicit information explicit- 
This knowledge is coded into 
the system initially, and is not learned or modified, 
Charniak's goal of understanding stories leads directly to the questions 
of what kind of implicit information is conveyed, what kinds of common sense 
knowledge are required to explicate that information, how should this common 
sense knowledge be represented and when and how is it used. In answering the 
question about the types of implicit in£ ormation Charni ak discovered that 
resolution of pronominal reference frequently requi red common sense knowledge. 
Consider Charniak's most discussed example [Charniak, 741 : 
Today was Jack's birthday. Penny and Janet went to the store. They were 
going to get presents. Janet decided to get a top. "Don't do that" Penny 
said. "Jack has a top. He will make you take it bock." 
The "it" in the last line is normally understood to refer to the top that 
Penny would buy, but any purely syntactic procedures would probably select the 
top that Jack already has, on the basis of recency. The correct choice of 
referents seems to be based on the common sense knowledge that if a present is 
purchased and given, it may be returned or exchanged. The fact that if a 
person receives a duplicate present he may not wish to keep it is also 
Page 52 
relevant. This realization led Charniak to concentrate on the problem of 
pronominal reference. Instances of reference are clearly recognizable in the 
input, and the need to utilize common sense knowledge seems fully present. 
Charniak decided to represent common sense knowledge as inferences. That 
is, implicit knowledge is made e~plicit by having rules which infer the 
implicit fact upon presentation of the necessary explicit information. Por 
example, if one knows that it is raining, the inference that anyone outside 
will get wet is valid, The presentation of the sentence 
John went outside. 
is sufficient to trigger the common sense inference 
John got wet, 
in a situation in which it is known to be raining. Charniak's chaice of 
in£ erences as a represent ation was probably strongly influenced by the 
availability of MICRO-PLANNER. (For a discugsion of MICRO-PLANNER, see 
Winograd [74] and Charniak 1761) Before examining his inferences in more 
detail, one other question must be considered. Since an inference is drawn at 
some particular time, the question of when inferences should be made arises. 
Three possible answers to the question are suggested: 
1. Make no inferences until the story is accessed (e.g. to answer a 
question about it, summarize a part of it, etc,) 
Page 53 
2. Make inferences as the story is read, but only those that are 
necessary to solve some particular problem (e. g. to solve p roblems 
of ambiguity or reference) 
3. Make non-problem inferences as the story is read, making explicit 
aa 
much information as possible 
Charniak rejects the first possibility because if it is necessary to make 
dedu.cti0ns from all previous propositions when accessing some proposition, 
there is no theoretical difference from making them as the story is read, And 
he argues 
thqt there is no way to be sure that you have made the correct and 
necessary inferences without examining all previous propositions. An argument 
against both the first and second possibilities is that the meaning of any 
proposition may be context-sensitive Consider a story in which Janet wants 
to trade with Jack for his paints. The sentence 
"Those paints make your airplane look funny" Janet said. 
should probably be understood as part of a bargaining strategy and not an 
expression' of Janet's true feelings. Since no explicit problem is presented 
by this sentence, no inferencing would be done, and hence the correct 
understanding would be missed. So Charniak selects the third possib'ili ty , 
making inferences whenever pos~ible as the story is read. 
Charniak implements inferences essentially as MICRO-PLANNER antecedent 
theorems. He distjnguishes two types of common sense inferences which he 
designates base rout ines and demons. Base routines rep resent knowledge that 
Page 54 
should always be avatlable and does not need to be triggered by a specific 
context* For example, knowledge about t rsding should always be available ro 
that any statement about a trade that has occurred will cause the ibfarencer 
that the ownership relations have been reversed. Base routincar are 
implemented as antecedent theorems matched against each input proporition. 
Some inferences are not always appropriate, such as the previous example about 
getting wet in the rain. The inference would be incorrect if made in r 
non-raining situation. This type of inference, called a demon, is implementad 
as antecedent theorems that are not always active. Demons are activated by 
base routines. Thua, there would be a base routine about raining which could 
make immediate inferences and activate demons appropriate to the context, 
rain. The previous inference rule uould be a demon of this type. 
Charniak refers to the set of propositions which match a base routine's 
pattern as its topic concept. He notes that a topic concept may occur either 
before or after a proposition which would match one of the topic  concept'^ 
demons. Consider the two sequences: 
1, It was raining. Jack was outside. 
2. Jack was outside. It was raining. 
In the first sequence, the rain base routine is matched which activates the 
outside-implies-wet demon. The demon matches the next proposition causing the 
inference that Jack is wet. Charniak calls this looking fomard, and it is 
handled correctly by antecedent theorems. In the second sequence, however, 
the demon is not activated until after the proposition so the inference is 
Page 55 
mirsed. This is called looking backward. Charniak proposes an extenaion to 
antecedent theorems so that when a demon is activated, the data base is 
marched for matches. Then, in the second sequence, the inference would be 
made when the the demon was activated. 
A final observation about demons concerns deactivation. Obviously, if 
demons are context dependent, they must be deactivated when the context is no 
longer preeent. Thi~ is illustrated by 
It was raining. When it quit, Jack went outside. 
It should not be inferred that Jack got wet. Although thia simple example 
could probably be handled by the rain base routine, in general the problem is 
quite difficult and Charniak offers no real solutions. He assumes that 
deactivating demons after some fixed number of intermediate propositions would 
be a satisfactory first approximation. 
Charniak's model includes two other components. Be suggests a 
bookkeeping component to keep the data base updated and tonsistent. When 
inferences are made, they will frequently replace other assert ions previously 
true. For example if Jack ia inside, then goes outside, bookkeeping would 
mark the original fact as no longer true, but would keep it for historical 
purposes (such as answering the question, "Was Jack inside?"). The fourth 
component is made up of fact finders. These are necessary as a result of 
Charniak's decision to make all possible in£ erences which are exp resaed as 
demons as the story is read. 
Clearly, many possible inferences should not be 
made 
to avoid clogging the system with a huge number of assertions. 
To avoid 
Page 56 
this, these unnecessary inferences will not be realized as dcmons. Inferences 
such as the facts that John is in his house, his neighbor!~ood, ctc, if he is 
in his kitchen fall into this category. Some situations may require the 
availability of these fact so The sequence 
Jack was in the house. Later he was in the kitchen. 
should not be interpreted as a change in location, from the house to the 
kitchen. Fact finders are implemented as MICRO-PLANNER consctluent theorems, 
with patterns which are matched against desired goals. They thus make certain 
information available through deduction that is not sufficiently important to 
infer as soon as possible. Fact finders are always available, not activated 
and deactivated like demons. 
Charniak's complete model is depicted below. The model shows that the 
inferences are treated ;Like additional propositions and are thus subject to 
DSPo 
Incoming Apply Apply 
>Base-------- 
Apply 
Asgertions------- Xla t ching---- >Bookkeeping 
CC 
Demons Routines 
I I I 
I v v 
------- Inferences (new assertions) 
3.1.2 Rieger - 
In some ways Rieger's work [74. 751 seems closely related to Charniak's, 
but it is not clear that this is completely true. Ricger's system accepts 
sentences as input which are already analyzed into their semantic 
representations, and makes explicit additional information that he claims is 
Page 57 
iqlicitly in the input. 
Re designates these additional pieces of information 
inferences, and agrees with Charniak that they should be made whenever 
possible Rieger discusses the use of in£ erences in handling problems such as 
reference, but the bulk of his work is specification of the inferences, 
themselves, and it is here that the main value of his work 
is found. Rieger 
Sdentlfies sixteen classes of inferences which will be discussed in three 
brod categories (not Rieger's categories). 
31.2 Causal Connection - 
Pirst, there are inferences which are concerned with the causal 
camectiopg between states and acts (Rieger's treatment resembles the use in 
robotics wo* (e.8. Fikes and Nilsson [71] ) of preconditions - states that 
be true for an act to occur - and postconditions - states that result 
fror m act occurring). Given that a state or act is true or has occurred, 
rbrt inferences may be made? 
1. Causative inferences suggest the likely cause 
input: Mary has the diamond ring. 
inference: Someone must have given or sold 
the ring to Mary. 
2. Enablement inferences suggest states that were necessarily t me 
input: Mary gave John a book, 
inference: Mary had the book just before she 
gave it to John. 
3. Resultative inferences suggest results that f ollswed 
input: Mae gave John the book. 
inference: John has the book, 
4. Missing enablement inferences explain why something cannot occur 
input: Mary couldn't see the horses finish. 
inference: Something must be blocking her view. 
Page 58 
5. Intervention inferences explain how something may be stopped or 
p revent ed 
input: Mary was hicting John ljith a bat. 
inference: Taking the bat away from Mary 
would stop her. 
3,1.2.2 Missing Information - 
The second category of in£ e rences concerns supplying common knowladgo 
about familiar objects or actions that is not in the input. 
1, Specification inferences fill in missing parts 
input: John hit Mary. 
inference: John used his hand to hit Mary, 
2. Function inferences supply the normal role of objects 
input: John got a book. 
inference: John will read tbe book. 
3. Normatiwe inferences supply information about what f s normally true 
input: Pete is a, human. 
inference: Pete probably has a gall bladder. 
4, State duration inferences suggest how long some state will persist 
input: John started eating at 6:00. 
inference: fie is probably still eating at 6:15. 
5. Feature inferences connect features of objects with the objects 
input: Fred wagged his tail. 
Xnference: Fred is a non-human animal- 
6. situation inferences supply other likely aspects of a sftuation 
input: Mary is going to a masquerade. 
inference: Mary is probabky wearing a costume* 
3.1.2.3 Motivation Atid Rnwledse - 
The third category of lnfarsoces concerns human mrivationr and 
knowledge, and their relation to one's actions. 
1. Motivation inferences suggest reasons for an actor to do eo~vlthing 
input: John hit Mary. 
infetenee John wanted Mary to be hurt. 
2. Actioa-pwdiction infbmncao suggert a poseible courae of actian from 
a pamon's wants 
input: John wants some naila. 
inaereacer John is likely to go to a 
hardware store. 
3. Enablement-p rediction inferences suggest reasons for s pamon 
bringing about a certain etate 
input: Mary put on hot glasses. 
inferencd: Mary probably wants to 
look at emthing. 
4, Utterance-intent inferences supply information intended, but not 
actually ~tated 
input: Mary couldn't jump the fence. 
inference: Mary wanted to jump the fence. 
5. Knwledge propagation inferences predict whet else a patson would 
know if given that he knows certain things 
input: Bill knew that Mary hit. John with a bat. 
inference: Rill knew that John had been hurt. 
One point that ehould be clarified is that the use of the tern inference 
to designate the implicit information a sentence conveys ie partidlly 
misleedcng. What is implicit, and thus misht need to be made explicit, is 
directly a function of the seaant ic representation. Consider the previous 
examp le 
Mary gave John the book. 
Page 60 
If the fact that John possesses the book after being given it is an inference, 
It 1s not clear what meaning the representation of the sentence captures. 
Surely, the transfer df possessian ia the meaning of the sentence, not just an 
implicit addition. At the other extreme, the example 
John hit Mary. 
surely does qot necessarily mean that John wanted to hurt Mary. This is 
clearly conveyed implicitly, if at all. 
It is very difficult to evaluate the completeness or adequacy of such a 
large set of inference types, but Rieger certainly suggests how large the 
class of information is that can be implicitly conveyed by text and, 
therefore which might need to be made explicit during the processing of that 
text. Any proposed scheme for representing and understanding text should be 
examined to see how each of these types of information is stored and in what 
way the information is asserted to be part of the meaning of the input. 
3.1.3 Schank - 
Charniak's model of story comprehension primarily discgsses conne~ttot~a 
between propositions of a story that alter the sequence of events involved in 
understanding the ston. Far example, the understanding of some sentence of a 
story. may be correct because a demon had been previously activated, thus 
allowing the correct interpretation of that sentence. Charniak dges not say 
much about explicitly rep reseating the connection between the proposition 
responsible for activating the demon, and the correctly understood 
Page 61 
proposition. However, since he mainly investigates reference, obtaining the 
correct referent is an explicit representation of the understanding. 
Clearly, 
many kinds of connections might exist between propositions of a connected 
discourse. Schank ( [73b], Schank and Abelson [77] ) investigates causality as 
one of the primary discourse connectives. Schank's proposals are discussed in 
terms of his Conceptual Dependegcy theory [Schank, 73a], but are largely 
independent of it. The term conceptualization is used to refer to either a 
proposition (using one of the small set of primitive predicates or acts) or to 
a state (i.e. an object having some value for some attribute). Schank notes 
that the sentence 
John cried because Mary said he was ugly. 
asserts a connection between 
Mary said John was ugly. 
and 
John cried. 
But closer consideration reveals that some of the links in this connection are 
unspecified, John must have found out that Mary had said it and this 
knowledge must have made him unhappy, which was manifested in his crying. In 
order to be able to recognize instances of missing links and to be able to 
supply them, Schank develops a classification of types of causality and 
characterizes their form and meaning. (It should be clear that here, as in 
all Conceptual Dependency, there is no simple mapping from English to the 
Page, 62 
underlying representation, ) 
The first type of causation is a result in which an act may bring about a 
change of state, For example 
John went to New York, -RESULT-> John is in New York, 
The second type of causation is enablement This is the situation in which a 
change of state brings about the conditions necessary for some act to occur, 
John has a ring, -ENABLE-> John gave it to Mary. 
The third causal type, initiation, is the relationship between any 
conceptualization and the act bf someone thinking about that thing. 
John -INITIATE-> I think about Bill. 
(John reminds me of Bill. ) 
The fourth and final type is reason causation, This is the relationship that 
holds between the act of deciding to do something and actually doing that 
thing. 
John decided to leave town. -RFASON-> John left town, 
Schank also suggests the utility of s non-specific caugal connection, which he 
designates "lead-t 0". This would be used to rep resent causally connected 
events in situations where the specific type of causality or the specific 
chain of causes is unknown, 
Page 63 
Although Schnnk explores the purely syntactic expansion of English 
causals into his formal causals, he concludes that the syntactic approach is 
too limited. Consider the sentence 
The hurricane caused my depression. 
Treating this as an instance of initiation causation (the hurricane 
initiated 
depressed thinking) makes it causal-syntact ically correct, but there ie 
clearly something missing that explains what led to the depressian. Schank 
suggests that other knowledge is used to find the connection, and this 
suggestion seems to indicate a far more general approach to understanding 
English expressions of causality than any syntactic approach, 
Schank suggests that a person would, if possible, find a reasonable 
connection such as that the hurricane probably blew down the speaker's house 
which caused him to lose money which caused him to become depressed, In order 
to accomplish this elaboration a person would have to use a great deal of 
world knowledge. Schank proposes that a number of different kinds of 
knowledge are involved here. First, there are axioms about the way people 
feel, such as 
Al. Bad results can cause depression. 
Associated with an axiom would be a number of more idiosyncratic beliefs, such 
as 
IB1. Less money is a bad result. 
Finally, general knowledge of the world would also be necessary. 
This would 
necessarily include facts auch as 
WK1. Objects can have monetary value. 
WK2. Changes in an object's physical state can cause a 
change in its worth* 
A person making the above connections would essentially be engaged in problem 
solving. Given the final state, depressioo, he would find en axiom which 
explains it then find an idiosyncratic belief which would meet the exim's 
preconditions (bad results), then use world knowledge to establish that a 
negative change of state could be the real culprit, then notice that a 
hurricane is able to destroy objects. The real point of this discussion seems 
to be that much additional knowledge is required to establish conrlections that 
people are able to (and do) make when understanding connected sentences. This 
is an affirmation of Bellert's hypothesis, and con£ irms Charniak's 
conclusions, but is reached after work on a different text understanding 
p roblem, 
Schank [75b] attempts to use these types of causality as a basis for 
defining the semantic rep resentation of a paragraph. lie argues that the 
collected rep resentat ions of the individual sentences of the paragraph do not 
form a representation of the paragraph as a whole. Much addttinnal implicit 
information could be made explicit, but Schank suggests that undirected 
explication is not plausiblee He offers a solution to the problem of what 
information should be made explicit by defining a paragraph representation as 
a causally connected sequence. Inferences are limited to those items of 
information required t~ find the causal connections. 
Page 65 
The etatea that are usually true when an act occurs are called the 
neceasaty condition8 of that act. 
These states would be ~onnectcd to the act 
by an enabling causation. 
Schank divides necessary conditions into two types: 
absolutely 
necessary conditions and reasonable necessary conditions (ANCs and 
RNCs). The former must always be true when an act occurs. RNCs, however, tare 
normally true, but may be violated without creating an anomalous situation. 
For example 
John was working in his yard. 
has a8 an ANC (among others) that John has a yard. An RNC might be that it 
was nice weather, but this could be violated* It would only indicate 
something unusual, and perhaps significant. Schank illustrates the use of 
neceeeary conditions to establish causal chains by analyzing several stories. 
A brief description of one of these should indicate his methods. He considers 
the following paragraph: 
John began to mow his lawn. Suddenly his toe started bleeding, ... When 
he cleaned off his toe, he discovered that he had stepped in tomato sauce. 
Mowing the lawn has as ANCs such things as the existence of John, the lawn, 
a 
mower, etc. and as RNCs that it is good weather and the grass needs mowing. 
These conditions would be inferred when the first sentence was processed since 
they were not explicitly stated. If the first sentence had been preceded by 
an explicit statement of either or both of these RNC6, they would have been 
connected as enabling the mowing. (There is great difficulty in finding a 
general scheme which allows inferring of normal necessary conditions, but 
prohibits inferring unusual ones. In this case, it secms reasanable to infer 
Page 66 
all of them.) The second sentence presents a more difficult problem. Same 
sequence of inferences would have to be made which produces the chain: mowing 
involves turning blades, which could hit a toe, which could cause it to bleed. 
The final sentence of the story illustrates the point about normality 
considerations. To explain the tomato sauce, one must infer that John got it 
an his toe, but this seems like an abnoml inference to make with no other 
explanation. However, if the paragraph was preceded by the sentences: 
John was eating pizza outside on his lawn, He noticed that the grass was 
very long and and he got out his mower. 
Then the final sentence seems much more reasonable. One would infer that 
pizza has tomato sauce, and John dropped some on the lawn and later stepped in 
it, 
In summary, Schank states the following conditions for the representation 
of a paragraph: 
1. Each input sentence is represented. 
2. These rep resentat ions should be conceptually connected, primarily by 
causal chains, 
3. The necessary condl tfons for every conceptualization must be 
explicitly represented, and may originate as input sentences or may 
be inferred either from previous conceptualizations or because they 
are normal. 
4. A story is the joining of causal chains that crilminote in the "point" 
of the story. 
Other paths are of less interest. 
3.2 The Organization Of World Knowledge 
A characteristic common to all of the preceding work is that a great deal 
of world knowledge ia required, but there is no clear organization of this 
knowledge. It is clear that any system with large amounts of knowledge 
represented simply as demons or inference rulea would become bogged dawn 
searching for relevant knowledge and would quite probably draw incorrect 
in£ erences because knowledge would be applied in inappropriate contexts 
(recall Charniak's concern about the deactivation of demons). Charniak [ 75, 
76, 771 himself suggests that more organized knowledge would be superior to 
the demon approach. These realizations have led to a number of proposals 
for 
the organization of knowledge. Minsky [75] iatroduced the term "frame" for 
knowledge structures, but Winograd I751 Bobrow and Norman [75] (using tb 
term "schema") and Rumelhart and Ortbny (771 discuss closely related ideas. 
Recently, Bobrow and Norman 1771 have described a language for knowledge 
representation (Ec2RL) in which the frame concept plays a central role. These 
knowledge structures are intended to organize conventional or encyclopedic 
kncrwledge, rather than def initionsl features or characteristics, This 
knowledge is described in terms of roles or slots which participate in the 
situation. These are the variables of the frame. Frames must also have the 
ability to reference other frames. A detailed discussion ot the nature of 
Page 58 
frames is inappropriate here, primarily because it is outside the scope of 
this survey, but in part because many of the computational questions regarding 
frames are not completely answered. However, recent computational work haa 
begun the attempt to incorporate such organization of knowledge into text 
understanding systems. 
3.2.1 Schank And Abelson - 
3.2.1 1 Scripts - 
Schank ([75bl, Schank et a1 1751 and Cullingford [751) extends his ideas 
wfth the addition of large, pre-existing knowledge structures that he calls 
scripts. He says that it is necessary to have large amounts of specific 
knowledge about known situations, since otherwise it is difficult or 
impossible to recognize the causal relationships between events. A script is 
a sequence of actions that provide knowledge about the typical occurrence of 
some situation. The following is a partial description of the script for 
going to a restaurant: 
Script: Restaurant 
Track: Coffee shop 
Roles: Custome r, Waitress, Chef, CasKier 
Props: Tables, Menu, Food, Check, Money 
Reason: To get food to eat 
Entry conditions: Customer (is hungry; has money) 
Results: Customer (is not hungry; has less money; 
is pleased) 
Cashier (has more money) 
Scene 1: Entering 
Go to restaurant 
Find an ernp ty table 
Page 69 
Decide where to sit 
Go to table 
Sit down [MAINCON] 
Scene 2r Ordering 
Receive menu 
Read menu 
Decide what to eat 
Tell Waitress what is wanted [MAINCON] 
Scene 3: Eating 
@em 
Scepe 4: Exiting 
1 *. 
The scenes are the main episodes of the event'and each is defined in terms 
of 
a sequence of specific actions. Schank suggests that scripts would need to be 
divided at the top level into different tracks whicih distinguish the sequences 
for different types of restaurants. 
Th& usefulness of a script is seen by considering the sequence 
Johnwent to a restaurant. He ordered a hamburger from the waitress. He 
paid and left. 
Many details of this sequence have been omitted and connections between the 
actions would be impossible to establish wf thout knowledge of what constitutes 
going to a restaurant. Furthermore, the definite reference to the waitress is 
meaningful only because all participants in a script are automatically 
introduced by any reference to that script. The script also contains the 
conditions that must be true for successful execution of the script, as well 
as the results of successful execution. Each scene has one act that is 
considered the essential act of that part of the script. These acts are 
Page 70 
called the MAIN CONceptualizations of a scene. 
Scripts represent the knowledge about conventional situatiow that any 
reader would be assumed to know. But the separate question of how to 
represent a text which contains script usages must be answered. If the 
rcy resentat ion of the restaurant story above wee merely an elaborated causal 
chain containing many events in£ erred from the restaurant script, the unifying 
description that these events together constitute going to a restaurant worrld 
be lost. Schank. proposes that that the elaborated causal chain is only one 
level of the story's representation - the Conceptual Dependency (CD) level. 
There is also a second level of representation - the Knowledge Structure 
(KS) 
level. At the KS level the above story would have a representation like: 
ScriptsRestaurant 
Custome rcJohn 
Food=Hambu rge r 
Additionally, the script representation would be linked to each of the events 
at the CD level which was part of that script instance. Schank suggests that 
the causal chain at the CD level should contain all the events explicitly 
mentioned plus the MAINCONS of any scene that is mentioned. If any event is 
encountered which is not anderstandable in terms of the current script, the 
event will be represented at the CD level, but will also be placed on a Wierd 
list. For example Schank discusses the processing of a story in which John 
has his pocket picked while riding a subway. Later, he has no money with 
which to pay the bill after eating at a restaurant- The first sequence of 
events is understood in terms of the subway script, but the pocket picking 
event is placed on the Wierd list. When applying the restaurant Gcript, the 
&aability to pay is encountered in the text and is inexplicable from the 
rotm~rmt script. The script applier asks the monitoring program to look for 
cprrs~al occurrences that could result in John having no money. The 
dtor finds the pocket picking event, and returns the informati6n to the 
scrlpt qplier which uses it as the required explanation. 
Of course, the problems in general are much more difficult than this 
riqle example indicates. For example, an unusual event within a script could 
imterrupt the script's normal continuation. Obstacles, such as not being able 
to get wbat you want at a restaurant, may cause altering or abandoning the 
restaurant script. Distractions, sueh as a robbery in a testaurant may lead 
to suspension of the restaurant script for a sub-story, or to its abandonment, 
Furthermore, simultaneous,, independent scripts are possible. The situation of 
eating in a dining car involves both the train script and the restaurant 
script. Finally, the activation and termination of scripts is a very complex 
p roblem. For example, consider the sequence: 
John went to a restaurant, After eating lobster, he bought a watch. 
Does this describe a restaurant event followed by a purchasing event, 
or the 
more unusual case of buying a watch in a restaurant? Either is possible, and 
the story in which this s,equence is found would determine the most probable 
interpretation. 
Page 72 
3.2.1 2 Goals And Plans - 
Schank and Abelson (75, 771 (also see Meehan 1761 and Wilensky [76]) have 
recently suggested that a11 connected event sequences are not apprcrp riatoly 
represented by scrlpts. Consider the sequence: 
Johh wanted to become king- He went to get some arsenic. 
It seems that treating the poisoning-the-king situation as so conventioarl 
that a script for it exists is unrealistic, The solution lies in realixlag 
that while scripts handle well known situations, mechanisms rmst exist which 
are able to handle novel situations using general knowledge. "Plans" are 
suggested as the appropriate mechanism. A plan is a general course of actaon 
intended to realize some goal(s), A possible hierarchy of high-level goals is 
shown in Figure 3-l(a). Motivated actions are always associated with the 
satisfaction of some high-level goal- Frequently, more specific sub-goals 
exist which are motivated by a high-level goal. For example, someone may want 
to go to the train statton (instrumental goal) so he can go to New York (a 
1 I 
specified goal) so he can enjoy pleasure" (high-level goal). The purpose of 
clearly defining a goal hierarchy is to enable understanding of situations in 
which an actor faces goal conflict and elects $0 pursue the highest goal (e.gw 
"preserve health" rather than "en joy pleasure"), 
Page 73 
Achieve 
Sex Rest 
a. Goal Hierarchy 
------------------------------- 
GET (named plan) USE (named plan) 
INVOKE- THINE (planbox) 
ASK (planbox) 
D-KNOW (D-goal) 
C 
BARGAIN-OBJECT (planbox) 
'TELEPHONE-BOOK (script) 
b. Associated Knowledge 
Figure 3-1 
Page 74 
But general knowledge about Row goals can be achieved must also be 
available. The "D-goal" is proposed as the fundamental unit of organization 
for suhh information. A D-goal ie a point pf access to planning information 
for the realization of some goal. Foe example, D-KNOW is the D-goal for the 
goal of knowing something, D-CQNT for the goal of physically controZling 
something, and D-PROX for the goal of being in proximity to something. The 
D-goals' associated knowledge is an ordered list of "planboxee", each of which 
ptovides detail infoption on one method for achieving the goal. The 
ordering prwides thb sequence in which the methods are likely to be 
considered. For example, D-KNOW has an ASK planbox (as a highly likely 
method) which specifies the actual act of asking the appropriate question, the 
intended result (fee. getting the answer), and the pre-conditions necessary 
to successfujlly ask. Some of the pre-conditions are: communication is 
possi"b1e; the persbn being asked knows the answer; the person being asked is 
disposed to answer the question, When a planbbx is ext rernely specific, it 
becomes a script. Using the telephone book is a method of acquiring certain 
kinds of knowledge that is so conventional that it is a script, Certain 
'recurring sequences of B-goals are called "named plans". The named plan, 
USE(x), stands for: 
D-KNOW(the location of \x) 
D-PROX (x) 
D-CONT (x) 
perform p_reparatlons and do action appropriate to x 
Figure 3-l(b) sham the relationship of D-goals, planboxes, scripts and named 
plans to each other and to the goal hierarchy. 
The representation of the plans of actors in a text would be at the KS 
level (with scripts). The D-goals would be explicitly represented and would 
be connected to actual acts at the CD level which (attempted to) implement the 
plan For example, the sentence: 
John tried to find out who ate 
the candy. 
would be repreaentsd arro: 
Ks Level 
CD Level 
- 
plan 
D-KNOW(John,?.ate the candy)---implementati-on--> DO 
I 
Failure <--------------------- result------------ 
I 
That is, the plan of trying to know who ate the candy was implemented BS some 
unspecified act (DO) which failed to achieve the plan's goal. Acts (at the CD 
level) which meet pre-conditions of planboxes would be linked to the D-goal 
they enable. 
Additional information is suggested as being appropriate and necessary 
for the KS level. This includes, for each character, his goals, the current 
status of each goal, strategies which could be used to achieve each goal and 
facts (true information as opposed to actual occurrences) relevant to goal 
understanding. This in£ onnation i.s maintained on "Coal Fate Graphs", which 
also contain associations between characters and any "Themes" (large goal 
complexes such as BECOMING-RICH) in which he is participating. 
Page 76 
Schapk and Abeleon appear to be committed to the development of a very 
complex system for the organization of knarledga, and the representation of 
such knowledge when it bccurs in stories. It should be apparent that they rare 
attempting to model goal related human knowledge and actions, since such 
knowledge and actions are common in stories. Since this knowledge pn be 
required to underatand particular stories it 14 not clear at what point (if 
any) one ceases t o study text unde rs t anding and begins modelling personality . 
3.2.2 Phillips - 
Phillips*[75a,75b] presents probably the most comprehensive computational 
model of text, in that he is concerned with the representation of all types of 
knowledge and textual relatiorships. This breadth is informative, but 
necessqrily results in a lack of depth in some areas. He presents a 
representational scheme which he uses for the vari~us required types of 
knowledge, which include world knowledge, li%gulstir knowledge and the 
knowledge conveyed by the text. Phillips w.orld knowledge, "the 
encyclopedia", consists of both the static data structures and the dynamic 
processes which operate upon these structures. The static data structure is a 
fairly conventional but very well defined, semantic network (closely related 
to the suggestions of Hays [73] ) which use nodes for entities and events, and 
arcs for the relationships between nodes. The set of hierarchical 
relationships between entities (the taxonomic- structure) is called' the 
paradigmatic structure. The syntagmat ic structure is the see of relationships 
between events and the event particip ants (case or argument relationships). 
Page 77 
Phillip8 incroaucea a Modality node attached to every event which is always 
used to refer to the event and its participants as a whole. (In many systems 
this is represented only indirectly.) The discourse relationships of 
cauoali ty , time ordering and spatial relations are rep resented by arcs between 
the Modality nodes of events. Phillips defines another type of relationahip 
which he designates the Metalingual organization of knowledge. This allows 
him to represent a concept as a unit, and yet have a complete subnetwork 
elqborating the meaning of that concept. For example, the unitary concept 
"poison" would have as its composition "someone ingesting something which 
causes that person to be ill". This rep resentational technique provides the 
fundamental capabilities of frames or scripts - expansion of something into 
its parts, and knowledge about those parts. 
The processes that operate on the semantic network are divided into two 
classes, The first, path-t racing, involves only following pathts through the 
paradigmatic structure. This type of process would be used to find that "Mary 
gobbled caviar" was a more specific instance of "People eat food". The second 
type of process is pattern-matching, and involves constraints between the 
components, For example, determining that "John killed himselft'is suicide 
while "John killed ~111" is not, requires a coreference constraint in the 
definition of suicide. Phillips observes that path-tracing is computationally 
equivalent to finite state automata, while pattern-matching is necessarily 
more complex. 
Page 78 
Discourse, or text, is represented using the same representational 
scheme, but may be characterized by properties not applicable to discontinuous 
knawlddge. The first of these properties is connectedness. Two propositions 
are paradigmatically connected if each has an argument such that the two 
arguments have a common paradigmatic superordinate node, or if the first is 
the immediate superordinate of the the second. For example, "lions" and 
'It igers" both have "animal" as superordinate, so the following are connected 
paradigmatically: 
Lions are indigenous co Africa. Tigers have stripes. 
And the two propositions: 
Man is a hunting animal. Modern man hunts for sport. 
are also connected since the second is a more specific proposition ("modern 
man" is immediately subordinate to ''man") than the first. Two propositions 
are discursively connected when discursive relations (em g. causality ) exist 
between them. 
The second attribute of discourse is thematicity. A theme is a 
prescribed pattern (represented by a Metalingual construction) to which a 
discourse may conform. A theme may be ~'contentive" like "accidental drowning" 
t l 
which is represented as existing when a person is caused to be in the water 
and is unable to act, the combination of which causes him to drown". This is 
contentive siqce it specifies both the parts (events) of the theme and their 
interrelationships. A 'lnon-contentive" theme is one that provides only a 
structural pattern or the interrelationships between some unspecified 
entities "clue" is defined to be "an unobserved act that causes the 
existence of something". The exact events and results are unspecified. A 
discourse is thematic if its propoaitiona can be matched to a theme or 
hierarchy of themes (i.e. a theme which matches several propositions in a 
discourse, such as "accidental drowning", may in turn be a part of another 
theme, such as "tragedyft). Phillipa then defines a coherent discourse as one 
which is both connected (each proposition is connected to at least 
one other 
proposition) and thematic. 
Phillips points out that although there is no simple one-to-one mapping 
between his representations and the constructs used in text analyses like 
those of Grimes (section 1.2.3) and Propp (section 1. .2), his structures do 
provide for the representation of most of the proposed relationships, Some of 
Grimes rhetorical predicates, such as Attribution, Specific and Collection, 
correspond to paradigmatic connectedness, while others, such as Covariance, 
correspond to discursive copnectedness. Still other rhetorical predicates, 
such as Respqnse and Analogy and Propp's patterns of functions in a move (as 
well as the functions, themselves) correspond to thematic structure. Hence, 
Phillips claIms to have presented computational interpretations for the 
principle text phenomena. 
As a test of his proposed model of text, Phillips presents a model of 
text understanding embodied in a computer program. His program inputs a 
discourse in the form of parse trees of the sentences and builds the knowledge 
representation of the discourse. An interesting use of the Metalingual 
const mcti on is its function in replacing non-cognitive surface words 
by 
the 
Page 80 
appropriate cognitive structure. For example, the preposition "through" is 
replaced by the cognitive st tucture for "in-contact-wLth" in a sentence like: 
The Abominable Snwman walked through the snow. 
"fhus, the same representational mechanism is used for such divergent types of 
knowledge as syntact ically-related information and thematic structures. 
Once the parse tree has been converted into the knowledge representation 
of the proposition, the understanding process involves two major steps. The 
first fs begun by matching the input proposition (IP) against the encyclopedia 
to find a corresponding generalized proposition (GP). Thus, the IP 
The boat contains Horatio Smith. 
is matched to the GP 
Things contain peaple. 
Notice that Phillips uses GPs to capture the same knowledge that features and 
selectional restrictions capture in many other systems. So the correct 
interpretation of a GP is that it is a plausible proposition, not a necessary 
truth. Once the IP has been matched to a GP, three additional checks are made 
starting from the matched GP. One is to determine if any of the terms of the 
proposition have Metalingual definitions. If so, new knowledge correspondin& 
to the definition is added to the discourse representation. For example, if 
the discourse included that "John was poisoned" the elaborated information 
that "John ingested something that caused him to be ill" would be added. A 
second check is to see if the matched GP is discursively connected to any 
Page 81 
other GPs. If so, new k~lowledge is added to the discourse which corresponds 
to the related GP and the d$.ecursive connection. If the rnatche'd GP were 
"People are injured" and it had a causal link to "People are unable to act", 
then the IP "John was injured" would result in the addition oE "John was 
unable to act" with a causal link from the original IP, It is by using these 
interrelationships of GPs that Phillips accomplishes inference. The third 
test is to determine if the matched GP is a part of any 'contentive theme. 
(The encyclopedia's GPs have pointers to all contentive themes which contain 
them.) If so the theme is matched against the discourse as a whole to see if 
all components are present and all constraints satisfied, If they are, the 
theme is addtid to the discourse rep resent ation. "Accidental drowning1' would 
be added to a discourse representation which had matched the GPs and 
connections I' ("People contact- water" and "~eople cannot act") cause ("People 
drown") " with a coreference constraint on "people", 
After processing Ips and adding the related structure to the discourse 
representation, the second major step is testing the discourse for coherence. 
This involve two tests. The first: determines if the discourse is connected. 
Then, thematicity, is tested by checking to see if a single undominated theme 
has been found. It should be noted that non-contenrive themes, since they 
have no component actions, and thus cannot be pointed to by GPs, must be 
tested for in a serial fashion. If the discourse passes both tests, then it 
is judged coherent. 
Page 82 
To illustrate the process, a very brief description of the understanding 
of a atory will be given. Given the story: 
(IP1) A boat contains Horatio Smith. 
(IP2) The boat overturns. 
(IP3) Horatio Smith drowns, 
the following events would occur: 
1, IP1 and IP2 would be matched to GPs which are parts of a complex 
event that has as a causal result, when instantiated, the added 
propositions: 
(AP1) Horatio Smith is in the water. (caused jointly by IPl and IP2) 
(AP2) Horatio Smith is injured, (caused jointly by IP1 and IP2) 
2. The GP corresponding to AP2 results in the addition of: 
(AP3) Horatio Smith cannot act- (caused by AP2) 
3. The GPs corresponding to AP1 and AP3 are part of a complex event that 
has the causal result "People drownt', But this is the GP which 
matches IP3. So a causal link is added from AP1 and AP3 to IP3. 
4. The discourse is tested for connectedness, and passes the test. 
5. The discourse matches the "accidental drowning" theme, which is the 
only theme it matches, so it is thematic, and thus coherent, 
Page 83 
Phillips principle goal was a computational model of text* His model of 
text understanding was intended to demonst rate that the proposed 
representations could actually be built from an input text. His text model 
does seem to present a reasonable representation for a number of text 
phenomena, several of which have not been considered by other computational 
models, Paradigmatic connectedness and some types of thematic structure seem 
particularly important. Iiowever, several object ions must be raised to his 
model of text underslanding. The first objection is that his testa for 
coherence are applied only to the complete discourse, and are not formulated 
in such a way as to suggest strategies to avoid incorrect interpretation of 
propositions pnd relationships, If this were done, the sequential processing 
of input p topositions would continually test for coherence, and incoherence 
would immediately suggest that some misinterpretation might have been made, or 
some inference omitted. A second objection i& that the definitions of 
connectedness and thematici ty are inadequate. A common superordinate node is 
simply not sufficient to explain paradigmatic connectedness. The example of 
sentences about lions and tigers would be connected only in some context that 
explained why these statements were being made (e, g. "All* I know about lions 
and tigers is . . ."). Thematicity is defined without respect to how much of 
the discourse the theme accounts for. A theme cannot account for all of the 
propositions in a discourse unless the discourse is very trivial, and yet if a 
theme matched only the first three propositionS of a one hundred proposition 
discourse, it could hardly be called the theme of the discourse. Finally, a 
number of aspects of the understanding processes are not convincingly shown to 
be computationally practical. The problem of avoiding incorrect additions to 
Page 84 
the discourse representation from a rich encyclopedia is ignored. Each 
exawle has exactly the right infamation available. Also, it ie mot clear 
that: the use of generalized propositions would work when there am ~averrrl 
.levels of generalization possible. For example, the g perelized proposition 
"People contact water" has associated knowledge, but the higher level 
proposition "People contact thingstt would also need to be present with its 
associated knowledge. And finally, the methods of accessing themes - by 
pointers to all occurrences of generalized propositions or by serial search - 
both seem comput at ionally unaccep table. Propositions like "~eople go" would 
result in, a combinatoric explosion of possibilities. For all of these 
reasons, Phillips' understanding model is useful more in suggesting the kinds 
of problems involved, than in providing an actual model of text understanding. 
3.3 Other Work 
Space limitations and the narrowly defined scope of this sunrey have 
combined to eliminate certain interesting work from detailed consideration. 
Wilks [75, 76, 771 has described a text understanding (and translation) system 
which uses a meaning representation called Preference Semantics. The system 
normally opercltes in a basic mode, dealing with sentences individually, but is 
capable of entering an extended mode when a reference problem occurs. An 
example is "it" in 
John drank the whiskey from the glass and - it felt warm in hi3 stoomch. 
Several processes are involved in the attempt to determine the correct 
referent. "Ext ractionl' is the addition of logically true propositions, only 
Page 85 
wtlicSt in the text, but available from the meanings of the unite. Thus, 
both 
"ltm was in John's stomach, 
The vhiskey is in a part of John- 
roipld ath be extracted. These two, pqopositions can be identified, thus 
resolving "it" as "the whiskey". 
*I 
Bwever, had the above example been . and it was good." the original 
pruposifions, as vell as relevant extractions, would then be subjected to 
m 
ccuncm sense inference rules" such as 
1. <1 drink 2) -> (1 judges 2) 
2, (1 is good) <-> (2 wants 1) 
rt3ich would be used to find the shortest inference chain identifying the 
referent. For the example, when "the whiskey'' is tested, the established 
(John drhk whiskey) -I-> (John judges whiskey) 
-subset-> (John wants whiskey) <-2- (Whiskey is good) 
where %ant" events are a subset of "judge" events. 
A recent addition to Wilks' system is large knowledge structures (like 
frames) called "pseudo-texts". They are suggested as containing other 
knavledge which might be required in solving reference problems (although they 
have other important use in understanding individual words). 
Page 86 
At the levei of text, Wilks differs from Rieger and Charniak primarily in 
his insistence upun entering an extended made only when a problem requires it. 
Reference p toblems are the only problems he discusses as triggering thia mode 
His use of pseudo-text B in establishing textual connections is still too 
briefly described to critically evaluate. 
Rieger ~([75b, 75c, 76a, 76b1, Rieger and Grinberg [771) has described a 
complex system for the organized representation of cause and effect knwledge, 
.a& plans utilizing this knowledge. A set of decision trees ("selection 
11 
networks") are postulated which. given a goal, select certain common sense 
algorithms" capable of realizing that goal. Rieger asserts that the same 
knowledge st-ructures used for planning, should be used for understanding the 
intentional acts of others, To do this in text, he requires knwledge which 
allows prediction of goals and actions likely to be made in response to the 
occurrence of some event or state, A subsequent sentence is tested to see if 
it confirms any prediction by being a step in an expected action or in an 
algorithm to achieve an expected goal, This testing requires that all 
algorithms be indexed by step. Thus, it must b~ possible to find all 
occurrences of (X GOT0 Y) in some set of algorithms, A confirmation is used 
to more- coslfidently predict the course of action being followed, Clearly, 
Rieger is dealing with the same problems as Schank and Abelson (section 
3,2.1.2), but ha.s not yet clearly demonstrated the utility or practicality of 
his approach to this aspect of text understanding. 
Page 87 
Hobbs 176, 771 has discussed analyses of various texts based on 
his own 
system of semantics. Possible fntersentential relations are described by 
pattern - action pairs which, when matched by input sentences modify the text 
representation tree apprap riately. These relations include causality, time 
ordering, paraphrase, examp 1.2, contrast, parallel const.ruction and vlolated 
expectation. These relationship8 indicate that Hobbs does not make a 
distinction between content and form (author imposed) relations. Thc 
complexity of Hobbs' system (in the large number of rules and their 
interactions) makes evaluation difficult until he has completed his computer 
implementation. 
Other work that should be mentioned includes Schmidt ([761, Schmidt and 
Sridharan [77]), who discusses the problem of recognitioh of plans from 
actions, as well as Novak [76] and Bobrow, et a1 [77] who both use frame like 
knowledge structures in specialized language understandi rag systems. 
3.4 Discussion And Conclusions 
What have the discrissed computational models added to the previously 
described model of text understanding? Charniak, Riegex and Wilks have 
demonstrated how-it is possible to infer information only implicit in a text. 
Schank has especially examined the richness and complexity of causal 
connections, which are often implicit. This capability raises many additional 
questions, however, including when, and how many, inferences should be made. 
Whi le Cllarniak and Rieger suggest making many infe rences whencvv r possible, 
Wilks argues for making them only as required. Neither approach has yet been 
Page 88 
applied to sufficient ly 1 argc trxt s and knwledge bases to convincingly 
demonstrate its validity. 
Pre-existcnt kn~wledgr structures have been suggested by Schaak and 
Philips to avoid the extrc:mc8ly difficult problem of making matiy-step inference 
chains to establish inlpli ci t con~~cct ions. Although use of these structures 
has demonstrated their ability to meet this goal, as well as their usefulness 
in generation of summaries and paraphrases, problems have appeared. Complex 
knowledge structures and real texts present many difficult.kea la matching an 
input proyosltlon with an element in the structure. When the match is 
imperfect, or- when many choices are (computationally) posgible, it is very 
difficult to perform the required matching correctly, 
In an attempt to deal with this difficulty (as well as others), and to 
recognize the fact that novel situations are also unde rstandable, Schank, 
Abelson, Rieger and Schmidt have been concerned with recognizing the 
motivations and intentions of actors. This aspect of understanding is clearly 
important, for the stated rt3asons, but the approaches described have almost 
certainly been too simplif icd for the understanding ,of actual texts. 
It 
Thus *the requi rcmcnt that an understar~ding of a text contains as much 
inferred inforniat~o~~ as h~!c~ss~iry to mqet some minimal level of ... 
coherence" (section 1.4) is wen to bc a computationally difficult problem. 
The orgat-lizing cor~tt~i~t SI 111, tt~rl~~; arc i~sc~ful in dealing with this problem, but 
create new problems. Knocll c-dgc) nrltl rr~p resr~nt at ion of plans and intentions is 
introduced tn hnndalr* sor.\c3 of thcsr problems. Tt should be noted that the 
previously dj sc~~r;sotl nrotlc~l tl~~ not spcrl f i c,illy distinguished this type of 
Page 89 
knowledge. Coherence has been defined in tllcse computational models as 
connectivity established by inferrencing, matching a knowledge structure or 
being a step in a plan. The exact distinctions between these three types of 
knowledge are still blurred, and it remains to be demonstrated that these 
distinctions are appropriate and adequate. 
Only Phillips and Hobbs have addressed representing the form of texts, 
bqt these suggestions have not beeh computationally adequate, nor has there 
been any clear distinction between form and content structures. A better 
defined notion of text form should certainly play a role in the ongoing 
determinatiion of textual coherence that occurs during comp rehension in the 
text understanding model. 
With regard to forgetting, little has been added, although higherlevel 
knowledge structures could be used to summarile their more detailed components 
(which could then be "forgotted'). Schank has also suggested that the least 
connected propositions in a representation are those most likely to be 
forgotten. However, these ideas have not been seriously investigated. In 
general, computational models understand texts perfectly (if at all), and do 
not contain any imperfect ret rieval processes. Permanent learning and 
integration with prior knowledge has not been investigated in these systems, 
nor has any explanation been offered for occasional recall of surface text, 
-- 
These systems generally do not keep the surface text at all, but could easily 
do sv. However, it would result in perfect recall of this information. 
Page 90 
The principal value of computational models has been the demonstration of 
the great difficulty in actually specifying comprehension algorithms. It is 
all too easy, when explaining hw a particular result is achieved, to ignore 
the problem of avoiding incorrect paths. The fact that higher-level content 
st ructures are computationally uaeful, and also psychologically indicated, is 
an important confirmation. The investigations of more, actual texts by 
systems having more knowledge available (not just the relevant knowledge) is 
an important step in validating the models being proposed. 
Page 91 
4.0 FINAL OBSERVATIONS 
The following points are accepted by most researchers concerned with text 
understanding: 
1. Much information implicit in texts is explicit in an understanding of 
that text. 
2. There is some type of hier; rchical or multi-level organization(s) 
of 
the understanding of a text. 
3. Organized, pre-exi5tent knowledge is requlred to achieve this type of 
understanding. 
The study of knowledge represcntation,and the nature of knowledge structures 
is a primary concern. Many computational problems have not been resolved. A 
particularly important question for text understanding concerns the necessity 
of redundantly copying general knowledge for specific instantiations. Fahlman 
(773 discusses this problem in detail, and offers some ways in which to avoid 
unnecessary redundancy. The relationship of the content and the form of texts 
needs additional clarification. Similarly, the relationship of general 
content structures to representations for plans and intentions needs study to 
see how distinct these are. And of course, the large areas of learning and 
forgetting are importaht, but missing, components of a text understanding 
model. In addition to these open problems, attempts to apply the combined 
insights of the diverse research perspectives to actual texts is a necessary 
step in evaluating the adequacy of current text understanding models. 
Page 92 

REFERENCES 
Anderson, R. C., Reynolds, R. E., Schallert, D. L. and Goetz, E. T. 
1976. Frameworks for Comprehending Discourse. Technical report no. 12. 
Laboratory for Cognitive Studies in Education, University of Illinois at 
U rbana-Champ aign. 

Anderson R. C., Spiro, R. J. and Montague, W. E. (eds.) 1977 (in press). 
3chooling -- and the Acquisition - of Knowledge. Hillsdale, N.J.: Lawrence 
Erlbaum Associatesl. 

Barclay, J. R. 1973. The Role of Compreh'ension in Remembering Sentences. 
Cpgnitive Psychology, 4, pp. 2291254. 

Rartlett F. C. 1932. Remembering: - A Study - in Experimental - and Social 
Psychology, London: Cambridge University Press. 

Becber, A. L. 1965. A Tagmemic Approach to Paragraph Analysis. College 
Composition - and Communication, XVI, 5, pp. 237-242. 

Bellert, I. 1970. On a Condition of the Coherence of Text. Semiotics, 11, 
4, pp. 335-363. 

Bobrow, D. G, and Collins, A. (eds.) 1975, Represent ation - and 
Understanding. New York: Academic Press. 

Bobxow, D. G., Kaplan, R. M., Kay, Ma, Norman, D. A@, Thompson. H. and 
Winograd, To .1977. GU~; A Frame-Driven Dialogue System. Artificial 
I.ntellip,ence, 8, 2, pp. 155-173. 

Bobrow, D. G. and Norman, D. A. 1975, Some Principles of Memory Schemata. 
in Bobrow and Collins [75]. 

Bobrow, D. G. and Winograd, T. 1977. An Overview of KRL, A mwledge 
Representatioxl Language. _Cognitive Science, 1, 1, pp. 3-46, 

Bransford, J. D,, Barclay, J. R. and Franks, J. J, 1972. Sentence 
Memory: A Constructive versus Interpretive Approach. Cognitive 
Psycho'Logy, - 3, pp. 193-207. 

Eransford, Jo D, and Franks, J. J. 1972, The Ahstraction of Linguistic 
Ideas: A Review. _- Cognition, . 1, 2, pp. 211-249. 

Bransford, J. T). and McCarrel1, No S. 1974. A Sketch of a Cognitive 
Approach to Comprehension. in Weimer, W. B. and Palermo, D. S, 
(cds. ), -- Co~ition - - and -- the Symbolic - Processes. Hillsdale. N. J. : 
Lawrence Erlbaum Associates. 

Chafe W. 1972. First Technical Report, Contrastive Linguistics Project. 
Department of Linguistics, University of California st Berkeley. 

Chafe, Wi 1975. Some Thoughts on Schemata. in Schank .,. and Nash-Webber [75] 

Charniak, E. 1972. Towards a Model of children's Story Comprehension. A1 
TR-266, Artificial Intelligence Laboratory, Massachusetts Institute of 
Technology. 

Charniak, E. 1974. "~e will make you take it back": A Study in the 
Pragmatics of Language. Technical report. Instituto per gli Studi 
Semantic1 e Cognitivi, Castagnola, Switzerland. 

Charniak, E. 1975. Organization and Inference in a Frame-like System of 
Common Sense Knowledge. in Schank an'd Nash-Webber [75]. 

Charniak, E. 1976. Inference and Know1etlge.a in Charniak and Wilks [76] 

Charniak, E. 1977. Ms. Maloprop, A Language Compreh.-nsion Program. in 
Proceedings -- of the Fifth - International Joint Conference on ~rtiflcial 
Intelligence. 

Charniak, E. and Wilks, Ye (eds. ). 1976. Computational Semantics. 
Amsterdam; North Holland. 

Colby, B. N. 1973. A Partial Grammar for Eskimo Folktales. American 
Anthropologist, 73, pp. 645-662, 

Crothers, E. J. 1972. Memory Structure and the Recall of Discourse. in 
Freedle, R.O. and Carrol, JOB. (eds.). Languafie Comprehension -- and the 
Acquisition - of Knowledge. New York: John Wiley and sons. 

Cullingford, R. 1975. An Approach to the Organizatidn of Mundane World 
Knowledge: The Generation and Management of Scripts. in Proceedings of 
the 13th Annual Meetinp, of the Association for Computational ~in~uisticx 

van Di jk, T. A. 1972. Some Aspects -- of Text Grammars. The Hague: Mouton. 

van Di jk, T. A. 1974. Models of Macro-stmctures. uapublished mimeograph. 

van Dijk, T. A. 1975a. Recalling and Summarizing Complex Discourse. 
unpublished mimeograph. 

van Di jk, T. A. 1975b. Discourse Meaning and Memory. to appear in Journd 
of - Reading Behavior. 

van Dijk, T, A. 1976. Macro-structures and Cognition. contributed to the 
Twelfth Annual Carnegie Symposium on Cognition. 

van Dijk, To A. and Kintsch, W. 1977 (in press). Cognitive Psychology and 
Discourse: Recalling and Summarizing Stories. in Drcssler, W.U. (ed.) 
Trends - in Text-Linguistics-. New York: de Gruyter. 

Fahlman, S. 1977. - A System - for Kepresenti~ - and -- Usinp; kq1,-World Know1edp;q. 
unpublished PhD dissertation. Massachtisetts Institute of Technology. 

Fikes, Re Eo and Nilsson, N. Jo 1971. STRIPS: A New Approach to the 
Application of Theorem Proving to Problem Solving. Artificial 
Intelligence, 2, pp. 189-208. 

Franks, J. J. and Bransford, J. D. 1974. Meaning for Syntactic Form as a 
Function of Semantic Context. Journal - of Psychology, 103, 
5, pp. 1037-1039. 

Frederiksen, C. Ho 1975a" Acquisition of Semantic Information from 
Discourse: Effects of Repeated Exposures. Journal - of Verbal Learning 
and - Verbal Behavior, 14, pp. 158-169. 

Frederiksen, C. H. 1975b. Ef fekt s of Context-induced Processing Operations 
on Semantic Information Acquired from Discourse. _Cognitive Psychology, 
I, pp. 139-166. 

Frederiksen, C. He 1975~. Representing Logical and Semantic Structure of 
Knowledge Acquired from Discourse. Co~nitive - Psychology, 7, PYe 
371-458. 

Grimes, J. 1975. - The Thread -- of Discourse. The Hague: Mouton. 

Hays, D. Go 1973. Types of Pvrocesses on Cognitive Networks. in Proceedings 
of the 1973 International Conference on Computati~nal Linguistics. 

Hendricks, W. 0. 1972. A Structural Study of Nsrrztio-: Sample ~xfalyses. 
Poetics, 3, pp. 100-123. 

Hendricks, W. 0. 1973. Methodology of Structural Narrative Analysis. 
Serniotica, VII, 2 pp. 161-184. 

Hobbs, Jo 1976. A Computational Approach - to - Discourse Analysis. Research 
report 76-2, Department of Computer Science, City College, CUNY. 

Hobbs, J. 1977. Coherence and Interpretation in English Texts. in 
Proceedings - of -- the Fifth - International Joint -- Conference on Artificial 
Int,elligencc. 

Ihwe, J. 1972. On the Foundations of a Genera1 Theory of Narrative 
Structure. Poetics, 3, pp. 5-14?. 

Kintsch, W. 1974. -- Representation -- of Meanini --- - in Memo Huntsdale, New 
Jersey: Lawrence Erlbtium. 

Kintsch, W. and van Dijk, T. A. 1975. Recalling and Summarizing Stories. 
unpublished English translation of "Comment on se' Rappelle et on Re'sume 
des Histoires". Langage, 40, pp. 98-116. 

Klein, S. 1974. Modelling Propp and Levi-Strauss in a Meta-Symbolic 
Simulation System. Technical report no. 226, Computer Science 
Department, Unive~sity of Wisconsin. 

Kummer, W. 1972. Outlines of a Model for a Grammar of Discourse. Poetics, 
3, pp. 29-55. 

Labov, W. and Waletzky, J. 1967. Narrative Analysis: Oral Versions of 
Personal Experiences. in Helm, 3. (ed. ). Essays on the Verbal and 
Visual Arts. Seattle: University of Washington Press* 

Handler, J. M. and Johnson, N. S. 1977. Remembrance of Things Parsed: 
Story Structure and Recall. Cognitive Psychology, 9, 1, pp. 111-151. 

Meehan, J. 1976. - The Metanovel: -- Writing Stories by Computer. unpublished 
PhD dissertation, Yale University. 

Meyer, B. J. 1975. -.- The Organization - of Prose and Its Effects on Recall, 
Amsterdam: North Holland. 

Minsky, M. 1975. A Framework for Representing Knowledge, in Winston, P. 
(ed. ) The Psycholo_sy of Computer Vision, New York: McGraw-Hill. 

Norman, D. A. and Rumelhart, D. E. 1975- Explorations Cognition. San 
Francisco: Freeman. 

Novak, G. 1976. Computer Understanding of Physics Problems Stated in Natural 
Language. American -- Journal - of Computational Lin~uistics, Microfiche 53. 

Petof i, J. 1972. Syntactico-Semantic Okganization of Text Structures. 
Poetics, 3, pp. 56-99. 

Petofi, J. and Rieser,H.- (eds.) 1973, Studies -- in Text Grammar. Dordrecht: 
"Reidel. 

Phillips, B. 1975a. Topic - Analysis. unpublished PhD dissertation, State 
University of New York at Buffalo. 

Phillips, B. 1975b. Judging the Coherence of Discourse. in Pro~eedin~s of 
the 13th Annual Mectin~ of the Association for Computational Ljnguistics. 

Prop~, V. 1968. Morphologycf - -- - theFolktale. - Austin: University of Texas 
Press. 

Rieger, C. J. 1974. Conccj - tual Memory. u~~published PhD dissertation, 
Stanford University. 

Rieger, C. J. 1975a. Conceptual Memory. in Schank [75a]. 

Rieger, C. J. 197Sb. Conceptual Overlays: A Mechanism for the 
Interpretation of Sentence Meaning in Context* in Advance Papers -- of the 
Fourth -- International Joint conference -- on Artificial Intelligence. 

Rieger, C. J- 1975~. The Commonsense Algorithm as a Basis for Computer 
Models of Human Memory, Inference, Belief and Contextual Language 
Comprehension. in Schank and Nash-Webber [75] . 

Rieger, C. Jo 1976a. An Organization of Knowledge for Problem Solving and 
Language Comp rehension. Artificial Intelligence 7, 2, pp. 89-127. 

Rieger, C. Jo 1976b. Spontaneous - Computation in Cognitive Models. 
Technical report no. 459, Department of ~omputer~cience University of 
Maryland. 

Rieger, C. J. and Grinberg, Me 1977. The Declarative Representation and 
P rocedu raf. Sinlulation a£ Causality in Physical Mechanisms. in 
Proceedings- -- of - the Fifth International Joint Conference - on Artificial 
Intelligence, 

Rumelhart, D. Em 1974. Notes on a Schema for Stories. unpublished 
manuscript. 

Rumelhart. Dm Eo 1975. Notes on a Schema for StorPes. in Bobrow and 
Collins [75]. 

Rumelhart, D. Eo and Ortony, A. 1977. The Representation of Knowledge in 
Memory. in Anderson, Spiro and Montague [77]. 

Scallert, Do L. 1976. Improving Memory for Pmse: The Relationship Between 
Depth of Processing and Context. Journal - of Verbal Learning and Verbal 
Behavior, 15, 6, pp, 621-632. 

Schank R. 1973a. Identification of Conceptualizations Underlying Natural 
Language. in Schank and Colby [73]. 

Schank, Ro 1973b. Causality and Reasoning. Technical report no. 1, 
Instituto per gli Studi Semantici e Cognitivi, Castagnola, Switzerland* 

Schank, R. d ) 1975a. Conceptual Information -- Processing- Amsterdam: 
North Holland. 

Schank, R. 1975b. The Structure of Episodes in Memory. in Bohrow and 
Collins [75] . 

Schank, Ro and Abelson, R, 1975. Scripts, Plans and Knowledge. in Advance 
Papers of the Fourth International --- Joint ---- Conference - on Artificial 
Intellip,ence. 

Schank, R. and Abclson, R. 1977. - Scripts. Plans, Goals and Understanding. 
Hillsdale N1 J. : 1,awrence' Erlbaum Associates. 

Schank, R. and CoJ by, K. M. (eds. ) 1973. Computer Models of Thou~ht 
Langua~e. San Francisco: Freeman. 

Schank, R. and ?ash-IJebber, B. (eds.) 1975. TheoreLical Issues in Natural 
Lan~ua~e Froc~essin~. Associ~tion for Cchputational Linguistics. 

Schank, Re and the Yale A. I. Project. 1975. SAM: A Story Understander. 
Research report- no- 43, Department; of Computer Science, Yale 'University. 

Schmidt, C. F, 1976. Understanding Human Action: Recognizing the Plans and 
Motives of Other Persons. in Carrol, J. and Payne, J. (eds.). 
Co~nition 2nd Social Behavior. Hillsdale, N. J, : Lawrence Erlbaum 
Associates. 

Schmidt C. F. and Sridharan, N. S. 1977, Plan Recognition Using a 
Hypothesize and Test Paradigm: An Example. ifi Proceedings of the Fifth 
International_ -- Joint -- Conference on Artificial: Intelligence. 

Spiro, R. J. 1975. Inferential Reconstruction in Memory for Connected 
Discourse. Technical report no. 2, Laboratory for Cognitive Studies in 
Education, Ul~iversi ty of Illinois at Urbana-Champaign. (to appear 5n 
Anderson, Spiro and Nontague 1771 ), 

Steinberg, D. Dm and Jakobovits, L. A. (eds.). 1971. Semantics* 
Cambridge: University Press. 

Thorndyke, P. W. 1976. The Role of Inference in Discourse Comprehension. 
Journal - of Verbal Learning -- and ,, Verbal Behavior, 15, 4, pp. 437-446, 

Thorndyke, P. W. 1977. Cognitive Structures In Comprehension and Memory of 
Narrative Discourse. Copnitive PsyChology, 9, 1, pp, 77-110. 

Weinold, G. 1972. On Deriving Models of Narrative Analysis f xom Models a£ 
Discourse Analysis. Poetics, 3, pp 15-28, 

Wilensky, R. 1976, Using Plans to Understand Natural Language. in 
Proceedings --- of - the --- Annual --- ACM Conference. 

Wilks, Y. 1975. Semantics for Natural Language Inference, Arttficial 
Intelligence, - 6, 1, pp. 53-74. 

Wilks, Y. 1976. Parsing English. in Charniak and Wilks [76]. 

Wilks, Y. 1977. Knowledge Structures and Language Boundaties. in 
Proceedings -- of the Fifth Internatienal Joint Conference 3 Artificial 
Intelligc11ce. 

Winograd, T. 1974. Five Lectures 23 Artificial Intelligence. A1 Memo no. 
246, Stanford Universi-ty. 

Winograd, T. 1975. Frame Representations and the Declarative / Procedural 
Controversy. in Bobrow and Collins [75]. 
