UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Topics in discourse analysis Davidson, James Edward 1976

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-UBC_1976_A6_7 D37.pdf [ 7.88MB ]
Metadata
JSON: 831-1.0051779.json
JSON-LD: 831-1.0051779-ld.json
RDF/XML (Pretty): 831-1.0051779-rdf.xml
RDF/JSON: 831-1.0051779-rdf.json
Turtle: 831-1.0051779-turtle.txt
N-Triples: 831-1.0051779-rdf-ntriples.txt
Original Record: 831-1.0051779-source.json
Full Text
831-1.0051779-fulltext.txt
Citation
831-1.0051779.ris

Full Text

TOPICS IN DISCOURSE ANALYSIS by James Edward Davidson B.Sc., University of B r i t i s h Columbia, 1974 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in the Department of Computer Science We accept th i s thesis as conforming to the required standard The University of B r i t i s h Columbia September, 1976 (o) James Edward Davidson In present ing th i s thes is in p a r t i a l f u l f i lment of the requirements for an advanced degree at the Un ivers i ty of B r i t i s h Columbia, I agree that the L ibrary sha l l make it f ree l y ava i l ab le for reference and study. I f u r ther agree that permission for extens ive copying o f th i s thes i s for s cho la r l y purposes may be granted by the Head of my Department or by h i s representat ives . It is understood that copying or pub l i ca t i on of th i s thes i s fo r f i nanc i a l gain sha l l not be allowed without my wr i t ten permiss ion. Department of tC-*w /*»e/-f'£. ^ £c-s The Un ivers i ty of B r i t i s h Co 1umbia 2075 Wesbrook P l a c e Vancouver, Canada V6T 1WS Date *&es*Z? A* f 7 7* C Abstract This thesis deals with the theory and analysis of connected English discourse. The abstract theory of discourse, and i t s distinguishing c h a r a c t e r i s t i c s , are discussed. Some problems in computer analysis of • discourse - are' delineated; • a method of analysis, based on a-modified system of predictions^ i s introduced, and i l l u s t r a t e d -with 'examples from simple storias. A program embodying these concepts i s described. F i n a l l y , p o s s i b i l i t i e s for discourse, and i t s place in computational l i n g u i s t i c s , are discussed, and directions for further work indicated. i i T a b l e of C o n t e n t s 0. I n t r o d u c t i o n ..............• i ........... 1 1. D i s c o u r s e v I - . . . . . . . . . . . . . . . . . . . . 4 1. 1 D i s c o u r s e i n G e n e r a l . .. .... ... .. . • .•.. ............ . . . .4 I. 2 N a r r a t i v e D i s c o u r s e 10 I I . R e l e v a n t F e a t u r e s o f D i s c o u r s e .. 17 I I . 1 - S t a g i n g *.••; ;•.-. .•... ..... ....... ... 19 I I »• 2 • Cob ss i o n . . . ' . • . • • v • .•...'». ••••• ... * ..... •..... • • • * • • • • * 21 • I i ; 2 / 1 - I n f o r m a t i o n B l o c k s ;....... .• .v.. .. 21 - -'II; 2.2 R e f e r e n c e -...... ................. 22 I I ; 3 C o l l a t e r a l i v . . . . . . . . . . . . . . 24 I i ; 4 Text s t r u c t u r e ...... ........ ;. 24 II.-5 "Context ' - . v - . ...... ... .. 26 I i ; 5 . 1 F o r e g r o u n d i n g .. . . .... . 27 XX* 5 • 2 Px? & in s s • » • • • • • * • • • • • * # * •*»•••• » • • •«•*•••• 2*7 I I . 6 ' R e a l - w o r l d ' Knowledge .. 28 I I I . P s y c h o l o g y and N a t u r a l Language .33 I I I . 1 I n f o r m a t i o n ~ Theory . i . . . . . . . .. •. .-,... ..... .... . , .. .... .34 I I I . 2 Memory f o r Sentences •. ..... . ... , . ......... ... ..... 37 3 "Memory f o r •  D i s c o u r s e ... . . ; . . . . v . i . . .. ............. 40 I I I . 4 O r g a n i z a t i o n a l Schemata 44 III.. ;5 'Other* Work"-; v.............. ..... . 46 IV. C o m p u t a t i o n a l - P r e r e q u i s i t e s 48 IV. 1 <The- R e p r e s e n t a t i o n .. .. .-.•.................. 49 * IV. 1'; 1 E x t e n s i o n s t o c o n c e p t u a l Dependency ............ 54 IV; 2 An-1 A n a l y s i s system .. , ^.................... . 59 IV. 2.1 Comparison of Systems 60 ' • I V i 2;2 ,The P r e d i c t i o n System ........ ................ 70 V. A Model , ; f or" D i s c o u r s e - A n a l y s i s ....... ..................... 84 V. 1 E x t e n d i n g " the - P r e d i c t i o n System .. . 85 -v: Ti*14 The B a s i c ' c o n t r o l - S t r u c t u r e ..................... 86 Vi1 1 2 I n t e r a c t i o n o f P r e d i c t i o n s ...................... 87 V. 1.3- L e v e l s o f P r e d i c t i o n s 93 Vi-U'4 "Comparison w i t h P r o d u c t i o n Systems- 99 V. 2 •• Use of I n f or ma t i on Sources ,.. .... ............. • • • • • • - 103 V. 2.' 1 P r e l i m i n a r y - Requirements .....,.................. 103 •••72 2.'2 I n f o r m a t i o n Sources ........•............ 106 V, 3 A D e t a i l e d Example ,119 V.4 The Imp l e m e n t a t i o n .129 Co n t e n t s i i i VI. Conclusions f : .* •. ..'. • . • •.•••«• • • • • • • • • • • • • • • • • • • • • • 131 VI»• 1 ''Review''• • • • «»••••'• • • • •. ••• • »••«••••••,•• • • * • • * 131 VI. 2 Future Work.. 132 VI. 3 The' Future 1 of Discourse : Analysis ...-; ................ 134 VII. Bi bliocjr aphy ... •. .... ««,«••». •« • »••••..••••'•••'• • * • •«...« • • • • • • • • • • 135 Contents Iv Table of Figures Figure 11 - • A Schema for Discourse .-. .......... ............. .. ... 8 Figure 2 - ? A Sample Conceptual'Dependency Diagram .51 Figure-3 - Another'Conceptual^ Dependency Diagram ...... .....,.53 Figure 4 - Representation for a Short Paragraph .............. 58 Figure 5 (a)- --"-ft "Simple "A'TN grammar . . . ..-••* . v '... 64 Figure 5 '-(b)1' -*"An- : Extend3d"Grammar i .'*•*:>-•;'.'.., . .64 Figure 6 (a) - ! An ATN for • a Linear Sequence-. ................. 74 Figure 6 (b) - A n ATN * for' an"Exclusive -Or ... i . . . ............ . 75 Figure '6' (c) - 'An ATN"for'an ;Inclusive O r i . . . . . . . . . 75 Figure 7 (a) ' -Flow of'Control in Parsing a Sentence ......... 82 Figure 7 (b) - Resulting Parse . i ;..... .... ........ . .... 82 Figure '8 - -Control' Structure- of the• System- . i . .86 Figure ••-9 - Modified "Control Structure . v . . • * -. . • * *.............. 92 Figure 10 --Set"of'Inferences'for the Verb •Chased'-.,....... 126 Figure 11 - A Causal Chain ......................127 Figure 12 - Fin a l Representation of 'Rabbit' Paragraph ......128 Contents V l^knowle d;2§ment I would • l i k e to 'thank my supervisor, DC, Richard Rosenberg, for his patience and inspiration,- and for his work in helping to make t h i s "thesis a reality.* I would also l i k e to thank Dr. Alan Mackworth,; for his c r i t i c i s m s of preliminary drafts of t h i s ' t h e s i s . The-financial assistance of the National Research Council i s g r a t e f u l l y acknowledged. 1 0« I n t r o d u c t i o n Natural language analysis, l i k e most sciences, has proven amenable to : the-'divide and-conquer • philosophy: specify a small domain; and deal thoroughly with problems within that domain, while ignoring issues outside the domain. This thesis follows a ; s i m i l a r approach, except that i n t h i s case a"different; <and somewhat larger, target has been c h o s e n — the analysis of connected discourse. Work with single sentences i n i s o l a t i o n - i s recognized to be somewhat a r t i f i c i a l , ignoring as i t does the fact that language i s just hot processed out of context. It i s the intent of t h i s thesis to deal with some-of the-problems of connected discourse, in an e f f o r t to delineate exactly what 'context' i s . The f i r s t chapter i s b a s i c a l l y a survey of the previous work on discourse;-"The work- of various l i n g u i s t s w i l l be covered, and d i f f e r e n t theories of discourse considered. In the-second' chapter;*- t h i s .work w i l l • be covered again, t h i s - time - with - *;a "more pragmatic motivation: to discover the c h a r a c t e r i s t i c s ' of 'discourse-which make i t s analysis d i f f e r e n t from that of-single sentences. A number of such c h a r a c t e r i s t i c s w i l l be-,-described, together- with th e i r possible use in an analysis system.*- , . .... The' t h i r d " 'deals -"with- a somewhat•• peciphsral area, psychology. I • s h a l l 'review the relevant work-in the f i e l d , to provide a certain amount of validation for my approach. Introduction 2 In the'-'fourth - chapter, we return to the computational aspect of the problem. • Previous work in computational l i n g u i s t i c s •'is 'reviewed; and various issues of representation, control, etc; discussed. The f i f t h , and" most s i g n i f i c a n t , chapter i s a r e l a t i v e l y complete specif i c a t i o n of • a'model'for discourse a n a l y s i s . The various - pieces :• of > 'the ! • model are explained, and the ov e r a l l motivations" ( l i n g u i s t i c , psychological,- * and computational) d i s c u s s e d ; A -detailed ; example of a n a l y s i s ' i s presented, to i l l u s t r a t e the' points made." • The 1-final•chapter consists of conclusions, and indications of possible future directions. -Like previous-systems, the work here concerns a r e s t r i c t e d domain. Problems ' o f intention and b e l i e f are ignored completely;' 'Speech acts are not considered.' Questions of intension and extension-are - handled i n a s i m p l i s t i c manner. The domain of a n a l y s i s 1 i t s e l f i s - h e a v i l y r e s t r i c t e d , being confined to one area: written narrative discourse. Why, then, study discourse? I f e e l t h a t - i t o f f e r s - p o t e n t i a l , i n terms of e f f s c t s which are simply - not seen at-the - sentence l e v e l . Exactly what these effects are' ;will "become'clearer as the thesis develops. No 'attempt'"has"been* made here to explain In d e t a i l previous work in the f i e l d ; Good;reviews of computational l i n g u i s t i c s can be found'in'Borgida'(1975) and Gercone (1975) . It w i l l be obvious that the work presented here draws Introduction 3 heavily from-' many - sources. In pa r t i c u l a r , Riesback's (1974; 1975)'thesis was the•inspiration for many of the ideas developed here; his influence w i l l be obvious. -Also, work by Schank, Winograd, and Marcus has• contributed many insights. In the f i e l d of l i n g u i s t i c s , 'the most s i g n i f i c a n t impact has come from the work of Fillmore' (1968), Chafe (1970), Grimes ( 1975) and Halliday - (1967) . • "Other authors are referenced, where appropriate, throughout the paper. In many -ways, ' t h i s • thesis deals with an approach to discourse, " rather than a-complete system. To-quote a popular phrase, 'this paper raises-more questions than-it"answers'. The system• 'described• here has - been p a r t i a l l y implemented, but much remains to bs 'done; • It i s my hope that the reader w i l l , at the end, have a clearer grasp of the issues involved. I. Discourse This chapter i s an e f f o r t to characterize the general notion of discourse, i n order to provide a basis for the work which follows; • ? I t ' i s divided'into two sections; the f i r s t presents a broad overview of what we commonly c a l l discourse, and the second focusses on one s p e c i f i c aspect of t h i s -- narrative discourse. 1.1 Discourse i n General •"Discourse" can be described as "connected communication of thought"^. In fact, i t is- * more than t h i s ; - a discourse i s somehow more than'just'a'string of connected sentences, and t h i s fact must be taken'into account for e f f e c t i v e analysis. Discourse 'has appeared under a number of names i n previous work: •rhetoric';'composition',•communication', and •discourse• are a l l aspects of the same topic, which I s h a l l simply c a l l 'discourse•. • The primary motivation•for discourse study i s l i n g u i s t i c . Some l i n g u i s t s c u r r e n t l y recognize that language cannot be studied effectively-without going beyond the sentence l e v a l . In e a r l i e r l i n g u i s t i c s , 'from Bloomfield to Chomsky, the sentence was held to be the largest coherent unit of l i n g u i s t i c *Funk and Wagnall's Standard College Dictionary, Toronto, Fitzhenry and Whiteside, 1974, p. 380 Discourse in General Discourse 5 structure 2; - 'As Chomsky ' (1965, P. 3) stated: '[This study] w i l l be concerned with ... tha rules that specify the :well-formed strings'of minimal s y n t a c t i c a l l y functioning units...* [where 'minimal s y n t a c t i c a l l y functioning units' were l a t e r stated to ba sentences]. The-major' reason for t h i s was complexity; the t r a d i t i o n a l l i n g u i s t s saw discourse- as* a large, uncontrollable mass, not amenable to the precise 'mathematical formalization which they could apply so e f f e c t i v e l y to sentences. The<emphasis i s now changing. As Sanders (1970, p. 73) remarked: •' " • • • 'It can thus be concluded that the only possible natural domain' for a s c i e n t i f i c theory about any language i s the i n f i n i t e set of a l l possible discourses of that • language;* - •;- -• • The main reason for-the'switch-is that i t i s recognized (by some l i n g u i s t s , at least) that there are l i n g u i s t i c constructs which can only be' ~. seen * • at the ' supra-sentential - l e v e l . That i s , language 'is '"used " f o r - t h e ' purpose of communication, and any attempt-to :ignore t h i s by imposing a r t i f i c i a l constraints i s doomed to f a i l u r e . Interestingly, a number of uses •of discourse study can easily-be -seen; " F i r s t , there-is the p o s s i b i l i t y of learning how to' produce- 'effective' discourse. A common flaw in l i t e r a t u r e (both f i c t i o n a l * a n d non-fictional) i s the f a i l u r e to recognize 2 F o r example, Bloomfield' (Language, 1933, p^ 170) defined the sentence as ''an ' independent "form, not included in any larger (complex)*" form";'- Beneviste' (P^obiemes de linguistigugs general, 1966; p.: 128) stated "On phrase ne peut done pas s e r v i r d'integrant a un autre type d'unite"[A sentence cannot serve as an integrant of another type ,of unity] Discourse i n General Discourse 6 and make1-use1" of inherent st r u c t u r a l features which increase r e a d a b i l i t y ^ In ' f a c t , the journal College Composition and Communication1 conducted a- symposium on the paragraph (Becker, 1966; Christensen, 1966; -Rodgers, 1966) addressed to precisely t h i s problem; Obviously, a theory of discourse w i l l not provide an immediate solution, but i t w i l l constitute a step. Seconds-translation might be more easily-dons. One of the obvious desiderata-to a l l who watched the- -Machine Translation programs - of e a r l i e r years was the f a i l u r e to take into account-more global 'context 1 mechanisms. It seems that machines and people* alike • would benefit from the a b i l i t y to a f f e c t i v e l y characterize'a discourse. • Related " t o " t h i s i s the work on abstraction and content analysis. (see; for example, Hols t i (1969), or Carney (1972)). For obvious reasons, i t 1 i s desirable to be able to extract the •central thread••from a-text> to store i t for-whatever purpose. Again; t h i s seems d i f f i c u l t without a coherent theory of what a text i s . • ' ' ' Given the • possible uses for discourse analysis, what work has been done so far? Surprisingly (to me at l e a s t ) , there i s a f a i r amount; although much of i t f a l l s into f i e l d s peripheral to l i n g u i s t i c s . ' •'<•• ' One • <of- the- :standard f i e l d s of study, dating back to A r i s t o t l e , ? , i s - r h e t o r i c . It i s s t i l l taught today, as a method of enabling- the''writer to 'express himself more powerfully. Perhaps-it should b e - c r i t i c i z e d for being prescriptive rather than descriptive, but i t i s nonetheless an important f i e l d (note Discourse i n General Discourse 7 that-It'comes-close t o • s a t i s f y i n g the f i r s t advantage mentioned e a r l i e r , of production of e f f e c t i v e discourses). Literary c r i t i c i s m i s another •field--which has taken on aspects sof discourse-analysis; In order-to successfully review a work, the c r i t i c must have' in his own mind a fundamental idea of; what" i t ^ays; ; " and- • more* importantly; whether i t says i t e f f e c t i v e l y . Although there i s l i t t l e here i n the way of written" guidelines; the f i e l d has produced some int e r e s t i n g examples of-'discourse structure. The-temainder-of the work on discourse has come from within the • f i e l d s of l i n g u i s t i c s , -although even here i t seems scattered; f-and out' of'the-^mainstraam. " • -One such example i s Longacre (1970) , who worked under Pike in the -field"of"tagmemic grammars. The work-is too complex to be discussed'here; but he did - at least -provide a reasonable taxonomy of types: narrative; procedural, expository, hortatory, dramatic, '• a c t i v i t y ; -'and' epistolary. • . . . . Kinneavy (1971) 'provided what Is probably the single most comprehensive basis f o r 'work-in discourse.-• • He:: i d e n t i f i e d the categories '-of-'reference; • l i t e r a t u r e , persuasion, and expression (similar to Longacre•s- taxonomy); but : more importantly, presented - 'a'"reasonably^coherent theory of a discourse. He saw the encoder (speaker)-;"the: decoder (hearer) --and- r e a l i t y , as providing 1' ^ t h e ^ ' v e r t i c e s - of 'a triangle,' of which, the signal (discourse) i s the inside. Discourse i n General Discourse 8 encoder decoder r e a l i t y Figure 1 - A Schema f o r Discourse Thus, the function of the discourse i s formalized: to communicate from the encoder to the decoder some s i g n a l , which bears an undetermined r e l a t i o n s h i p to the r e a l world. Kinneavy developed i n d e t a i l each of the four points mentioned, and i l l u s t r a t e d d i f f e r e n t examples of the discourse s i t u a t i o n ; space precludes my saying more here. Within t h i s one framework, we can see how to deal with some of- the more advanced notions of current l i n g u i s t i c s : speaker/hearer r e l a t i o n s h i p , speech acts, purpose, communication s i t u a t i o n , etc. Inte r e s t i n g l y , various peripheral d i s c i p l i n e s have emphasized d i f f e r e n t aspects of t h i s structure. For example. Shannon; • in his work on information theory, paid less attention to the possible r e l a t i o n s h i p s to r e a l i t y , but added a new component, noise, to the s i g n a l . Morris (1945), a sign t h e o r i s t ; distinguished the actual r e a l i t y (denotatum) from what • i t means to the interpretant ( s i g n i f i c a t u a ) . More d e t a i l on t h i s matter i s given i n Kinneavy's o r i g i n a l work (pp. 18-26). Hausenblas (1964) of the Czech school, developed another method o f " c l a s s i f y i n g discourse. He i d e n t i f i e d three dimensions along which a discourse can vary: simple/complex; dependent/-independent (of s i t u a t i o n ) ; and continuous/discontinuous. These Discourse in General Discourse 9 components' seem- to - -characterize the style of tha discourse, rather -than its-content {cf. 1 Longacre) . Thus, 'we'•'have --been able to identify the function (Kinneavy) , content' (Longacre) , and style (Hausanblas), as potential'descrlptions-of'discourse. Within-the genera l ' f i e l d of discourse,' two l i n g u i s t s H a l l i d a y (1967; 1970; Halliday and Hasan, 1976), and Grimes (1971; • 1972;: 1975) — have completed a tremendous amount of work. Their influence w i l l be seen throughout th i s paper. Discourse in 1 -its ' entirety i s too broad a f i e l d to be covered'in a study such as t h i s . Since we want ultimately to design"- "a functioning model for discourse analysis, a smaller domain'must be selected, with more s p e c i f i c rules of structure and function. In determining what "kinds of discourse to study, one guestion concerns the medium:•is the discourse to ba spoken or written?" We w i l l ' d e a l with the l a t t e r because i t i s more 'complete1', i n the sense-- that everything i s insi d e the discourse. Spoken discourse ( i . e . conversation) makas extensive use ]of • •-extraiiRguistic" i n f o r m a t i o n : t h e speaker/hearer relationship, 'objects in'the environment, etc., as wall as such medium-dependent' features 'as intonation. - Such i s not, in general; 'the"case in"written discourse (although thare can be exceptions); " A l s o ; written- discourse -generally adheres more c l o s e l y to= standard' notions of grammatically; there i s less use made of sentence fragments, ejaculations, e l l i p s e s , e l i s i o n s , Discourse in General Discourse 10 etc. Accordingly, the study w i l l be confined to written discourse.-The question of ty_p_e (in Longacre's c l a s s i f i c a t i o n ) must also <be decided; I have-chosen narrative, for a number of reasons. F i r s t ; i t i s well structured; • -narratives have an inherent' organization, as has been observed by a number of researchers. Second, there already exists some work i n the f i e l d ; -tha same researchers mentioned above have investigated the problem thoroughly, and presented a few answers. Third, narrative-seems'to-'lend i t s e l f best to compactif i c a t i o n ; one can deal with very short s t o r i e s , and s t i l l not have the f e e l i n g of having removed the*'significant aspects. • This section has been an e f f o r t to characterize discourse, and indicate' the general-truths which hold about i t . The next section*deals with one p a r t i c u l a r aspect- of t h i s , narrative discourse. I.2 Narrative Discourse Throughout- t h i s section, 'discourse'^ unless otherwise s p e c i f i e d , w i l l refer to-narrative discourse. By 'narrative 1 i n t h i s case; ^ 1 ' - s h a l l " mean- -very - short (i*e. one-paragraph) st o r i e s ; In keeping with the • best t r a d i t i o n of Eugene Charniak and Roger Schank,'the examples I w i l l deal with later w i l l bs of a f a i r l y simple nature,*at the l e v e l of perhaps a 9-year old. One example might be: Mary was making dinner. She t o l d John to put the casserole i n the oven, and gave the salad to Narrative Discourse D i s c o u r s e 11 P e t e r . r T h e n ' s h e t o o k • t h e d e s s e r t o ut o f t h e f r i d g e . F i n a l l y , i t was r e a d y ; e v e r y b o d y s a t down a t t h e t a b l e . An e x a m p l e ' o f ' t h i s s o r t " a v o i d s most -of - the e s o t e r i c and complex s y n t a c t i c - c o n s t r u c t s w h i c h , w h i l e o f i n t e r e s t t o some l i n g u i s t s , a r e ' i r r e l e v a n t - t o o u r s t a t e d - g o a l o f d i s c o u r s e a n a l y s i s . At t h e same t i m e , i t p r e s e r v e s t h e e s s e n t i a l f e a t u r e s o f a d i s c o u r s e : t h e ' c o h e s i v e n e s s , t h e t e m p o r a l - c o n t i n u i t y , t h e use o f c h a r a c t e r s , • e t c . , - s o t h a t any p r i n c i p l e s d e r i v e d h e r e w i l l h o p e f u l l y be g e n e r a l i z e a b l e . G i v e n ' t h e s e examples o f d i s c o u r s e , w h a t ' k i n d s o f g e n e r a l s t a t e m e n t s can be made a b o u t them? P r o b a b l y t h e most i m p o r t a n t and n o t i c e a b l e " a s p e c t of- a n a r r a t i v e i s i t s t e m p o r a l c o n n e c t i v i t y . G e n e r a l l y , - t h e r e i s a c e n t r a l t h r e a d o f t i m e r u n n i n g t h r o u g h t h e s t o r y ; T h i s t e n d s t o s e r v e a s t h e b a c k g r o u n d , "upon ' w h i c h o t h e r e l e m e n t s are- hung. I n some l a n g u a g e s ; - ' i n - - f a c t ; t h e - s t y l e o f d i s c o u r s e i s s u f f i c e n t l y c o n s t r a i n e d t h a t ' t h e • t e l l i n g o f an e v e n t must t a k e t h e same o r d e r a s t h e event* I t s e l f . • - •—- - • •Grimes' (1971) i d e n t i f i e d s e v e r a l d i f f e r e n t components o f d i s c o u r s e . T h e • • f i r s t • a r e e v e n t s . ' T h e s e a r e ' t h e a c t i o n s o f a s t o r y u s u a l l y • ' c o n s t i t u t i n g ' i t s most i m p o r t a n t p a r t . They p r o v i d e t h e - s e q u e n c i n g " o f - a d i s c o u r s e , and a r e i n c o r p o r a t e d I n t o i t s c e n t r a l t i m e - l i n e . He-can d e s c r i b e t h e s e q u e n c i n g o f e v e n t s as b e i n g e i t h e r " t i g h t ( i . e . c o n t i g u o u s ' i n • time) o r l o o s e (i-.e. - c o n t a i n i n g - l a p s e s ) v ( I n t e r e s t i n g l y ; t h i s d i s t i n c t i o n can be t r a c e d back t o t h e a n c i e n t Greek, and t h e de v e l o p m e n t o f t h e N a r r a t i v e D i s c o u r s e Discourse 12 perfect tense.) '•Events tend to be clustered -- that i s , gathered* i n t o f - s m a l l Islands of contiguous groups, with gaps between -the islands.' -The next' 'class i s that of participant. I n t u i t i v e l y , a discourse contains a 'cast' of characters; t h i s may vary in size from one to thousands (although usually no more than 20 or so w i l l 1 b e - d e v e l o p e d ) ' P a r t i c i p a n t s : are related to events and other' participants. The' continued use of a r e s t r i c t e d set of participants provides a certain amount of cohesiveness in a discourse. ' "v : -: Most discourses w i l l have a setting, locating the events both " s p a t i a l l y "'"and-temporally. The description of the setting may be'brief," or " i t may be "omitted e n t i r e l y , with the reader l e f t to -make his'own'assumptions. There -is "generally -a f a i r amount of peripheral information which may loosely ' be * t i t l e d bjy:kqround. This i s usually explanatory - information, -designed to convey to the reader something he 1would not otherwise have known. Interestingly, the amount of b a c k g r o u n d ' i n f o r m a t i o n used can be an indication of the development of the narrative. If the author has kept the background information to a minimum, he i s probably i m p l i c i t l y asserting that -either: i) •the s i t u a t i o n being discussed i s quite simple; i i ) the' reader 1 i s already r e l a t i v e l y f a m i l i a r with the various parts. Grimes refers to another c l a s s , evaluations, as instances i n Narrative Discourse Discourse 13 which- the' author conveys his own feelings about an event or character; This is"sometimes done e x p l i c i t l y ; more often, i t i s done implicitly,' through the connotative• meaning of the various words chosen !( c°mparestubborn» and 'tenacious') . T h e * " l a s t , although s t i l l very important, class i s c o l l a t e r a l : a-catch-all'group containing information about the discourse i t s e l f " e s s e n t i a l l y , a meta-discourse. This can appear'in a number-of ways. "One i s foreshadowing, the hinting of events to come. Another i s alternative: the mention of events which did-not happen, used'to emphasize those that did. Yet another - (more common in expository than-narrative discourse) i s summary: a r e i t e r a t i o n of the contents -of the discourse. C o l l a t e r a l w i l l be discussed in'more d e t a i l i n the next section. This gives us a reasonable 'feeling for what a discourse i s . One part-not discussed yet is-coherency 'what i t i s that makes a discourse adhere together, rather than remaining a c o l l e c t i o n of isolated' units. This w i l l be dealt with more f u l l y in the next chapter, but a certain'amount can be said now. One of the-foremost cohesive methods i s r e f e r e n c e . T h i s I s related-the question'of<-participants, mentioned above; i f a previously • i d e n t i f i e d participant must be referred to again, an 'abbreviated' form-can be used." E s s e n t i a l l y , the problem of reference can-be*broken into two subgoals: i) establishing the-'identity of the referent; i i ) maintaining t h i s binding. The devices used to do t h i s (pronominalization, d e f i n i t e noun Narrative Discourse Discourse 14 phrases;'--demonstratives; etc.) are indicative of the s p e c i f i c i t y of ?the 1'reference, and, l i k e background, carry a l o t of i m p l i c i t information about the salience or newness of the referent. • • P h i l l i p s (1975) also develops some ideas • for coherency. Besides 5 anaphora, one method he sees i s causality; the fact of one event d i r e c t l y causing-another seems to unify the text. Schank? (1974a) * developed ^ t h i s further, defining a complete taxonomy of types of'causation, and showing how they contribute to o v e r a l l unity. •-•.>---.. • -i.-r •. Another method i s spatio-temporal connectivity. That i s , events must- appear to • be connected in space and time, and to follow naturally in"some'way. A • * last cohesive aspect, which w i l l be covered i n considerably more d e t a i l : i n the'next chapter, • i s theae-. This i s es s e n t i a l l y " the0- topic* -of the discourse, and i t may take a tr e e - l i k e organization (Lakoff, 1972). • These factors- a l l seem to apply to narrative discourse, but at" a •l o c a l * l e v e l . , r T h a t * i s , what we have been dealing with i s the microstructure of•the-discourse. In the next section, we w i l l l o o k "at- various e f f o r t s which - have been made to characterize the macrostructure (I.e., the global organization). I. 2.1 Text Grammars "' ' * • • Text 'grammar' 'is -a^loose term, used to'describe a series of e f f o r t s by-Propp ' (1 958), Kummer (1972) , van - Di jk (1972), and Rumelhart (1975). These have endeavoured to provide a 'grammar' Narrative Discourse Discourse 15 for s t o r i e s , -in"the'sense that a set of (often context-free) rules - can be"given,' with which*it Is t h e o r e t i c a l l y possible to generate the'stories"(or, i n our"case, to 'parse' them). Propp provided the 1 e a r l i e s t , and perhaps least structured, work; 'His conclusions were 4based on an analysis• of about one hundred * Russian- f o l k t a l e s . • Unfortunately, the - r e s u l t s were somewhat fuzzy; as Lakoff (1972) pointed out,-his categorization of events •as " belonging to s p e c i f i c classes i s highly suspect. Nonetheless, t h i s provided a st a r t on the problem Kummer- and van' Dijk, working within the transformational grammar -paradigm', produced - formal, well-specified systems. Unfortunately; -their models suffered, to a certain extent, from the same flaws as -other work i n transformatiorial grammar counterintuitiveness- and lack"•of perspicuity. That i s , the mathematical formulae may r e f l e c t s u p e r f i c i a l properties of the discourse, but" they do not correspond i n a n y easily understood way to our underlying i n t u i t i o n s about discourse; Thus, the models' "should' probably be dismissed as-a serious but somewhat I m p r a c t i c a l " e n d e a v o u r a l t h o u g h t h e i r study c o u l d l e a d t o certain'insights about'the nature of discourse. O f ' a l l the systems; that^by Rumelhart would seea to be (not surprisingly) ^the'one best suited'to.the a r t i f i c i a l i n t e l l i g e n c e approach/ 'Rumeihart developed a system that was based to a large degree on i n t u i t i o n — but which embodied most of the concepts associated with discourse. An example of his grammar i s : Narrative Discourse Discourse 16 story -->• setting + episode setting --> (state)* • episode '-->•• event + reaction The problem with" his system' i s , ' l i t e r a l l y - , that we are l e f t hanging.1 The*"'grammar "is guite imprecise, and the 'primitives'' (i.e.•terminals) are expressed at such a high l e v e l that i t i s d i f f i c u l t ' to• see- how they are represented in a c t u a l i t y . For example, ' s t a t e ' * i s - l e f t as an • unanalyzed ' (and unanalysable) primitive.'"'This "creates : problems in analysis of discourse, since presumably a 'state' could be manifest • in a number of di f f e r e n t ways,1 and the search f o r i t could prove explosive. •I feel;- •however>' that t h i s i s the right d i r e c t i o n to follow. Ther nature 1 of the system i s such that i t could hopefully be extended"to a more concrete l e v e l . In general, text grammars would seem to be a viable notion, i f developed i n the right way. This concludes the-chapter-on discourse. In the next, the problem of analysis of discourse i s re-examined. Narrative Discourse 17 II. Relevant Features of Discourse H i - 1"Staging l l . 2 Cohesion II. 2.1 information blocks II. 2. 2 reference II. 3 C o l l a t e r a l II. H Text structure II. 5 Context • II.5.1 foregrounding II.'-5* 2 frames I I . 6 Real-world knowledge In this'section we w i l l review the c h a r a c t e r i s t i c s of discourse mentioned e a r l i e r , t h i s time considering t h e i r a p p l i c a b i l i t y to a theory of analysis. In many*"cases, the' examples used w i l l be i n d i v i d u a l sentences; t h i s • i s done for-reasons of space. The properties described here are a l l extensible to and applicable at the discourse l e v e l . E s s e n t i a l l y , discourse i s , as mentioned, more than just a group of sentences, and our-model must r e f l e c t that. That i s , any system that analyzed a discourse by merely analyzing i t s component sentences would be a f a i l u r e , having missed the primary'characteristic;' •In- considering discourse analysis, two questions come to mind: Relevant Features of Discourse 18 i) what" features of discourse make i t easier (or perhaps harder) to handle than single sentences in i s o l a t i o n ? • i i ) what' form should the f i n a l result of the analysis take; i . e . how can the meaning of a discourse be represented? About the l a t t e r I s h a l l have l i t t l e to say, and even that w i l l be postponed u n t i l a l a t e r chapter. The - former 'can ' be 'discussed here. What we want to characterize i s 'the additional information- available in a discourse which' i s not found at the level-of single sentences. More importantly," we must show how t h i s can be used to aid in analysis; • .< ...... It has been "recognized " for some time ••in computational l i n g u i s t i c s work : t h a t in"order to be properly analyzed, a text must be, in some sense, - understood i . e . the analysis system must have f a i r l y extensive 1 knowledge, • about - both the subject domain and the* discourse-medium. For t h i s reason, the somewhat s i m p l i s t i c work - i n e a r l y computational l i n g u i s t i c s f a i l e d . What we need here i s a way to get a handle- on the information necesary to c o r r e c t l y analyze discourse; we know that syntax alone Is - not' enough.'" • s • "Several 1 p o t e n t i a l l y 'helpful features can be i d e n t i f i e d ; -I s h a l l "deal with' them one ; at a time. Some of these were described in the' previous chapter, some'have been mentioned in other'computational l i n g u i s t i c s work, and some are t o t a l l y new. For each feature, two things need be expressed: Relevant Features of Discourse 19 i) '.the' surface manifestation of the feature -- i . e . the •form i t takes in a discourse; i i ) the organizational "information; contained i n the feature 1 i . e . what i t t e l l s us about the structure of a discourse. * I • 1 Staging This i s described (Grimes, 1975) as the manner in which the speaker organizes the information for the hearer*s benefit. The most important' aspect'- '-of • t h i s • i s the theme-: the 'point of departure' for the'speaker; with respect to which he organizes the discourse"' (see ! H a l l i d a y , 1967, for a comprehensive discussion on t h i s ) . -More s p e c i f i c a l l y , theme i s the (abstract) concept i d e n t i f i e d ; t o ^ i c is'the surface form used to signal i t . Theme-occurs at" the •• sentence, paragraph,' and discourse le v e l s (see-Phillips', 'mentioned'earlier) . It-can most easily be i d e n t i f i e d at the sentence l e v e l , and that - onev w i l l be dealt with here, although sentences can easily ba related to the discourse l e v e l . -In -'surface structure, theme may be marked or'unmarked. If i t i s unmarked '(as-in most cases), the f i r s t nominal concept in the sentence constitutes the*theme. -' '~ The horse ran from the c o r r a l . In t h i s •case,-the speaker i s tre a t i n g 'the horse' as his point of departure.'' " . • The theme can be more precisely i d e n t i f i e d i f i t i s marked; Staging Relevant Features of Discourse 20 t h i s can ;be'done in a number of ways. • Fronting - " i s " the rearrangement of the sentence to put. the desired unit into • i n i t i a l ' : p o s i t i o n . ' The most common manifestation -of"-this i n 1 E n g l i s h i s the passive voice. • The c i t y was" razed by the earthquake. Here the rearrangement- has served to emphasize that the c i t y , and not the earthquake, i s what's being talked about. Other methods-of fronting-exist: " 'Home-is'where'I want1 to be. The e f f e c t s 'are' the "same. '2aEii^i23iS3'i s' a n 0'ther method of reordering - the sentence to highlight" the theme. One method of thi s i s extraposition (clef ting) : -'1 . . I t ' s ' B i l l who'took the car. Another form i s pseudo-cleft: ":"•What I want • for Christmas i s a new bike. Note that' in " the 1 f i r s t casey ' B i l l ' - would have been the theme even i n the unmarked1'form; the-marking merely adds emphasis. In the second example' the p s e u d o - c l e f t i n g emphasizes the uniqueness *of'"the theme -- i . e . , 'I want a bike (and nothing else w i l l - do) . '; -.•.-..<...-,-,• Embedding-can also serve - to emphasize a concept. Compare: John opened the'door with the key. John used the key to'open the door. In the second case; 'the'failure to use the standard instrumental case indicates that the key "is important. Theme can be marked i n a number of other ways; often the Staging Relevant Features of Discourse 21 m a r k i n g H i s ' h a r d - t o ' r e c o g n i z e . Some languages have mora r i g i d r u l e s ; -often 'the' theme w i l l have i n f l e c t i o n a l f e a t u r e s , and there'may b e ' d i f f e r e n t f e a t u r e s f o r the sentence theme and paragraph' theme;' The a b i l i t y to re c o g n i z e the theme, even i f i t i s only at the sentence " l e v e l , • i s q u i t e b e n e f i c i a l - i n a n a l y s i s . I t e s s e n t i a l l y conveys the o r i e n t a t i o n o f the -sentence: where the speaker i s • s t a r t i n g out 'from. T h i s can be used to c o n t r o l such processes as foregrounding, e t c . II.2 Cohesion H a l l i d a y - (1967; H a l l i d a y and Hasan, 1976) and Grimes (1975) i d e n t i f y t h i s as the manner i n which incoming information-r e l a t e s t o the i n f o r m a t i o n p r e v i o u s l y given. There are a number of aspects of cohesion, of which two are r e l e v a n t : i n f o r m a t i o n b l o c k i n g , and the s p e c i f i c a t i o n o f r e f e r e n c e . II.2.1 Information Blocks Information blocks are u n i f i e d chunks of i n f o r m a t i o n , which are t h r u s t upon the hearer one un i t at a time. T h e i r s i g n i f i c a n t aspect i s t h e i r 1 l e n g t h , which e s s e n t i a l l y i n d i c a t e s how much'information the hearer i s t o absorb, and thus how important* that -information i s ' ( t h i s has been r e f e r r e d to as the 'rate o f - i n f o r m a t i o n i n j e c t i o n ' (Grimes, 1975, pp. .274-275)-)-. Information 'blocks are' s i g n a l l e d i n speaking, through intonation.- ' In w r i t i n g , " punctuation ( e s p e c i a l l y commas) and c l a u s e s e p a r a t i o n are the u s u a l methods. Cohesion Relevant Features of Discourse 22 For -example, compare: The beaver, a timid animal, i s rarely seen i n i t s ' " natural environment. The timid beaver i s rarely seen in i t s natural • ' 1 ' " ••environment. In the f i r s t example; the use of the appositive clause creates a new information block, and serves to emphasize the beaver's ti m i d i t y . In 'the" "second example, the -adjective has been absorbed- d i r e c t l y into the longer block, thus down-playing i t s importance. ! . . . . Information blocks' • also capture' the 'given/new* d i s t i n c t i o n . " Some-blocks may contain only new information, others "both 'given and new.' In the latter'case, the given information generally-constitutes the f i r s t part of the block, with the new information following. 'Given 1 i s often confused with 'theme';"as Halliday explains i t , given information i s what has been" t a l k e d a b o u t 1 before,. theme i s what i s being talked about now (sea'Chafe, 1974,'for more comment on t h i s ) . The- discovery" of - an information block can ba valuable i n discourse analysis, -since i t generally i n d i c a t e s how s a l i e n t the information -contained i n - i t i s . In p a r t i c u l a r , longer blocks generally contain-less'important information than shorter ones — i . e . , t h e i r rate of information i n j e c t i o n i s lower. II. 2. 2 Reference- • •  • • As mentioned'previously, reference i s -one-of tha stronger cohesive elements (empirical v e r i f i c a t i o n of this w i l l appear in the next chapter). This fact i s often overlooked in Cohesion Relevant Features of Discourse 23 computational l i n g u i s t i c s , where the common practice i s to merely' establish- the (extensional) identity-'of - the referent. The important^aspect of reference i s - i t s specificity.. The various forms'of-referring (pronouns, demonstratives, inclusive nouns; ' "'definite • noun "phrases, e t c . ) , are a l l - s p e c i f i c to a greater^ or lesser degrees ••The" specif i c i t y indicates the extent to which"*the 1-speaker feels ' i t necessary to 'point out' the referent-—-i.-ei' r how '.-strongly 'he thinks i t i s present i n the hearer's memory. - • ---The form of reference we are dealing•- with here does not connect j ;very""closely witlr Charniak' s (1972) work on reference resolation;' -"The 'information carried in our - case does not usually-'help -in resolution (although i t - may) > but-serves rather to indicate the'salience of the referent. Compare: the l e f t arm of the chair the arm ' . '- •' - - ''it ' i- : . . . . . . . . as refer r i n g expressions;•* Each of these • has i m p l i c i t in It assumptions17 "about- the *'presence• of the referent in the memory of the-hearer. Again, see Chafe (1974) and HcDermott (1976) for comment on t h i s . Other >forms- of• cohesion' e x i s t , including: substitution, e l l i p s i s , l e x i c a l cohesion,'-and conjunction^ Space precludes dealing with' a l l - o f these here; they are covered more f u l l y i n Halliday and Hasan (1976). Cohesion Relevant Features of Discourse 24 11.3 C o l l a t e r a l This was covered in the l a s t chapter (pp. 12-13); i t i s , as the jname" implies, 'peripheral " information^ outside the actual content of t h e t e x t , whose purpose-is to emphasize parts of the text.' a-number of types were-mentioned, including foreshadowing, alternative, ^nd _ summary. • Of these, the use of alternatives i s probably the easiest to recognize. This i s indicated, usually, by the presence-of-a negated clause (or negating conjunction): We might have been k i l l e d , but the plane landed ' " safely. The e f f e c t 1 h e r e - i s to stress the positive aspect, i . e . 'we are s t i l l a l i v e ' . • ' Summary i s r a form rarely seen i n narrative discourse, but i t i s occasionally found in children's s t o r i e s : This i s a story about Goldilocks and the three " ' bears. Even- though-it' may appear'at the start- o f the story, t h i s i s an example of summary. - ,. The b e n e f i t s - o f ' c o l l a t e r a l are immense; e s s e n t i a l l y , i t provides - gratuitous information about how to organize the discourse. - The task of summarizing a story i s obviously greatly s i m p l i f i e d i f the author does i t for us: 11.4 Text structure As -mentioned• e a r l i e r (1.2.1), the best work on text structure seems to be that-of Rumelhart. In theory, we should Text structure Relevant Features of Discourse 25 be able to- develop a set of rules which would enable us to generate or parse a story. The problem, as indicated, l i e s with the gap between • the - top < and the-bottonr-- i . e. between the terminals of the-grammars'and the actual surface forms of the discourse. • •'• • • • *-••• • • The main'possibility would seem to l i e - with developing a r e s t r i c t e d grammar to handle certain sets of s t o r i e s . For example, one' rule might be: -setting --> temporal location | s p a t i a l location | (character)* | :' continuing event where- there is'an i m p l i c i t ' i n c l u s i v e or between the branches of the rule. Thus, a sentence l i k e B i l l y was playing-in his yard. - • f u l f i l s three of the f o u r ' p o s s i b i l i t i e s . Note' " that "-the 'inclusive or* cannot easily be handled with the standard-context-free'-rules. What i s needed i s a more expressive''control structure -- not"only context-sensitive but powerful and-intuitive; This p a r t i c u l a r problem w i l l not be discussed-here;- • ft grammar of 'the type described could probably be made to work, If the input set of s t o r i e s were s u f f i c i e n t l y constrained. The current 'uses -* of'•'frames' and 'scripts'- in a r t i f i c i a l i n t e l l i g e n c e would seem to f a l l into t h i s class. Text structure Relevant Features of Discourse 26 I I . 5 Context •'• 'Context' i s a'somewhat'nebulous term,-used to describe the 'surroundings' of the current input. The word derives from the Latin- for-'weave together'; and t h i s i s as good a d e f i n i t i o n as any.'Me -can i d e n t i f y several d i r e c t contextual e f f e c t s : 'i) change*of word-sense; i i ) change of - importance — i . e . what this component • "means-to the ov e r a l l discourse; i i i ) reference, e l l i p s i s , etc. Sentences in i s o l a t i o n have different meanings from those same'sentences in context, compare these examples from Reisbeck (197U) John and Mary were racing. They were a f r a i d of being- beaten; -John and Mary were running. They were af r a i d of being'beaten. The meaning of the second-sentence changes completely, depending on what' precedes i t -- even a change to a single word af f e c t s the c o n t e x t ; • - In :working - with context. It i s tempting to say that we need merely save a l l the • information i n a discourse, so that i t i s available when we need i t . This approach, however, quickly becomes "explosive;*what i s needed instead is a well-specified indication'-of'•which ' information to save, why to save i t , and how to use - i t l a t e r ' (see-McCalla, 1976, for a discussion on t h i s ) . Many • of* 'the- benefits of context are manifest in the theme (see 'staging', above), but two others are relevant here. Context Relevant Features of Discourse 27 II. 5. 1 "Foregrounding •' •': • The f i r s t - i s what Chafe (1970; 1974) has referred to as foregrounding.' -' - At'" any • given time, certain concepts are 'on stage','in that'they are i n that they are in the consciousness of the hearer;' Concepts are brought onstage by being referred to, and can-be"kept there via repeated reference. They seem to leave the'stage"simply fading away over time, although t h i s can be affected by various aspects-of the intervening discourse. An example might'be v I just found a book belonging to Peter. I wonder where'he's l i v i n g now. Here the ''he' c l e a r l y refers to 'Peter', as- the only suitable foregrounded* concept' ('book' i s also foregrounded, but cannot match the pronoun). : • Complex'-rules can' be* derived, both'for the foregrounding and"'unforegrounding' -(backgrounding?) of a concept. Chafe (1974) presents a comprehensive discussion. II. 5. 2 Frames'"''' The'other-use of>context which i s easily i d e n t i f i a b l e f a l l s under - the -notion- *of 'frames • (Minsky, 1974) - (here used in the sense-'-of- a—general" "situation, rather than a plot-oriented s c r i p t ) . "For-'-instance;-the:-text " John and Peter were playing baseball when the bat - ' cracked. i s perfectly coherent because we know that the baseball frame has a « s l o t ' *• for-'ba t. In some sense, t h i s i s merely another aspect of Context Relevant Features of Discourse 28 f o r e g r o u n d i n g b u t " ! f e e l i t ' i s better to i d e n t i f y i t i n Its own class. B a s i c a l l y , the function of frames (context c l u s t e r s , beta-structures, • schemata,...) i s to provide the understander with knowledge in -useful-sized'chunks — i.e. to p a r t i t i o n the memory;'-in ' an" e f f e c t i v e way. They also provide power for word-sense "-disambiguation^ and possible control of inferencing. Frames can become" a r b i t r a r i l y complex, almost to the point of text grammars,-or'arbitrarily simple, almost to the point of foregrounding. <•-.-,...•.-.. . <-,. in order to use the notion of frames-effectively, we need a reasonable-'--idea' • of how- and when frame are activated and deactivated. " This-'is a common problem in-current • A. I. work, and I have ••no' 1'easy''solution. ' I n the work here; I w i l l use a s i m p l i s t i c mechanism; in which a' frame i s activated by reference to the "appropriate•'-word 1 or concept, and deactivated (i.e. 'pushed out') when a replacement frame i s activated at the same l e v e l . There i s obviously much more to -context than what i s described here; for the moment, however, this w i l l have to do. II.6 Real-world Knowledge 'One-point that'has not been heavily stressed so f a r i s the '"of*! the discourse --' the actual content. Lest I give the impression that•form'and s t y l e are a l l that are needed, I w i l l ' present a' b r i e f summary here of the use of real-world knowledge. Real-world Knowledge Relevant Features of D scourse 29 Perhaps' the' best example of' this i s - Sieger's (1975) verb-driven - inference program,- which makes semantic deductions based on the 1input.' The-system makes forward inferences i n an uncontrolled'manner, trying to deduce a l l possible f a c t s from a given- input. The system i n - i t s o r i g i n a l form i s explosive, but with some e f f o r t 1 -we should - be able to provide enough coherent di r e c t i o n that the'process "will be more controlled. Of Rieger's sixteen inference - classes,' four ( s p e c i f i c a t i o n , function, cause, and result) are v a l u a b l e i n discourse. This r e f l e c t s our previously-stated notion that causality i s a major cohesive factor i n a-discourse; For example; given the sentence: our system might'make the following 'predictions': cause: •John was'angry at Mary. re s u l t : — . • Mary h i t John back. ^Mary started crying. s p e c i f i c a t i o n : ' • " • • •  John h i t Mary with his hand. ••-•-'John' h i t Mary with a hammer. function: -" '(none applicable) Thus, the semantic'^inference provides us with, i n some sense, the a b i l i t y to- 'understand' 'the story.' From t h i s , we can predict- what comes next, and the work involved w i l l be s i m p l i f i e d ••if- the event' does 'in fact- happen. One factor which seems to control the explosion of John hi t Mary. Real- world Knowledge Relevant Features of Discourse 30 inferences- i s 'an interesting 'windowing' e f f e c t : the inference mechanismJ'can use o n l y the current part of the discourse, and not previous input, as a basis. For example, i f after the 'John hi t Mary' sentence,we find' Mary hit'John back, the old set of^inferences (about how and why John h i t Mary) would -die o f f , ' s i n c e the sentence upon which they are based i s no longer 'in the window'. A' new set of inferences i s spawned instead, based on the concept of Mary h i t t i n g John. • "Another" method of* inference control was suggested by Rosenschein (1975)i * His" approach viewed inferencing as an operation dependent upon a whole set* of "facts. Under t h i s assumption, - -'the -system -was designed to find' the -' least possible pattern' ( i i e . , minimal extension to the set). The requirements for t h i s "inference'were"-that'it: ••• i) cover""" the- "Input'- 'set of facts ( i . e . make maximum possible use"of ; the•• given information); i i ) be independent-of !the current facts (i.e. not assert something already known); , , . I i i ) -be -minimal' (i.e.- make the fewest possible assertions). There are obvious flaws in t h i s approach, but i t merits inspection. Thus far,'we have i d e n t i f i e d half a dozen features of discourse which are available to a s s i s t i n analysis. There are Real-world Knowledge Relevant Features of Discourse 31 others, notably presupposition and d i c t i o n ( l e x i c a l s e l e c t i o n ) . I w i l l r e f e r - - t o • these as information sources (I.S.'s). Several c h a r a c t e r i s t i c s of% these sources can be i d e n t i f i e d : i) They are weak; the amount of information they carry i s not as strong as'the standard areas of syntax and ' semantics.- • •• • • • • • • i i ) They" may - have •nothing to say; the module for c o l l a t e r a l , for example, may l i e dormant f o r long ' •• periods.' i i i ) - They • are gratuitous; the -information i s there anyway, so a proper analysis system must use i t . Given - these conditions, how can we incorporate the I.S.'s into an analysis system to make the most e f f e c t i v e use of them. What we would-like t h e o r e t i c a l l y i s a clean modular system, with"-each 'module suggesting things whenever i t recognizes i t s own need: This"'brings us~-to the standard A.I. problem of interacting sources of knowledge. A possible control system w i l l be outlined in chapter v. ••«-••• r As an 'aside,'the ' features discussed here are s i g n i f i c a n t , because-' they 1 draw attention to the form/content dualism in language. That' i s , the - speaker (writer) , i n preparing his message, i s concerned with two points: i) what to say; ' i i ) how to say i t . where ( i i ) -describes-the-form-of the discourse (i.e . , the manner in which i t i s structured), and often contains a lot of Real-world Knowledge Relevant Features of Discourse 32 information relevant to the meaning of the discourse. Halliday (1970) recognized 1 the difference, referring to content as iiSS-ilSMi"'meaning-, 1 and to 1 form as intergersonal and textual. Unfortunately,'"the- trend in-computational l i n g u i s t i c s work has been to downgrade the importance of t h i s aspect. Thus we have outlined the t h e o r e t i c a l constraints of a discourse analysis system. The next section deals with a peripherally related area, psychology, but the following one (at last) presents a step towards a solution. Real-world Knowledge 33 I I I . Psyc hology and Natural Language This chapter^ represents a minor, but necessary, digression from our central'topic of discourse analysis. The intent here i s to present-some of the background work i n psychology, which has had such a pervasive^effect on A r t i f i c i a l Intelligence. It Is by-no means a complete survey, nor would I wish i t to be one. Rather, I hope to extract some general themes-- ground rules, as i t - were*-- which can be used f o r guidance in l a t e r implementation. In discussing work in-psychology, I w i l l not elaborate upon the methods • used in various experiments. Rather, the conclusions reached by each experimenter- w i l l bs presented, and the results interpreted i n l i g h t of the goal of language analysis. Through t h i s , I hope to build a simple model of the language understanding process. The work •in psychology which i s relevant to language analysis can be-effectively categorized into five classes: i) Information theory-and related work; i i ) memory for sentences; • iii)"s«e«ory*-f or'-discourse; iv) organizational schemata; v) other work. Each of -these w i l l be dealt with separately. Note that the form of memory in general-will not be discussed; t h i s i s too broad a topic, and a number of comprehensive references already e x i s t Psychology and Natural Language 34 (see; for example,- Tulving and Donaldson (1972) , Lindsay and Norman (1972);•'Anderson and Bower (1973), Norman and Rumelhart (1975), Cofer (1976) , and Norman (1976)) . Much-jof- the - ; e a r l y work in the psychology of language Involved l i s t - l e a r n i n g experiments, in the style established by Ebbinghaus (1 885) ; As such; It i s not generally relevant to our s p e c i f i c purposes. • In the more recent work, however, a number of interesting r e s u l t s have been produced. III.1 Information Theory One of - the • e a r l i e s t - works was M i l l e r ' s (1956a) seminal paper on coding-processes, 1 or 'chunking'. Miller's assertion was that units- of arbi t r a r y complexity could be retained in short-term'memory i f they could be 'chunked' - - i . e . converted to organized units.* He found that subjects could retain a r e l a t i v e l y 'constant number of chunks in • semantic memory, E§3§.I<Il'§s.s of -the-complexity of the chunks (see -Simon (1972) for some comment "on-this)-. f' The implications for language processing are obvious;- I n t u i t i v e l y , words entering into short term memory remain' there-as "individual- -units, u n t i l -chunked into larger groups • (e.g.- a phrase) ; these larger groups^ - in turn, may l a t e r be "combined into s t i l l l arger"units. The interesting e f f e c t , of course, comes" when" the- • upper l i m i t of short term memory (in Miller's-case," 7) is-approached. At this time, the input must be 'organized in some manner, or else information w i l l be l o s t from short term memory. Information Theory Psychology and Natural Language 35 Related to- t h i s are the 'perceptual strategies * of Bever (1970), who discovered: ^ i) psychologically, the clause i s the main element of the • sentence --• subjects tend to group smaller elements up to the-clause'level;- ••• i i ) subjects - tend to perceive clauses i n a basic S-V-0 order; -hence transformations such" as passivization 1 " delay'processing, since they force-a reanalysis; i i i ) ' subjects' treat t h e - f i r s t clause of-the sentence as the main'one; any s h i f t i n the ordering causes a delay in processing. Interestingly,-these e f f e c t s can be largely explained within the framework of the-chunking hypothesis; the strategies mentioned are part of an e f f o r t to prevent information*overload, by making as much use as possible of the syntactic structure of the input. - • we - are 1 thus " f a c e d with a model-' of - a hearer who i s processing "information'as rapidly as possible/ to•remain within the constraints of-his memory system. I n t u i t i v e l y , he does t h i s with the aid" of 'the' highly complex structural' features of natural language;-" "A' few were mentioned •in-the'last section, i n dealing with cohesion, staging, c o l l a t e r a l , etc.- A more qeneral approach'will be taken here. • •• ' A starting* point for such an i n v e s t i g a t i o n i s provided by the incredible redundancy, of natural language.- Shannon (1952), in his pioneering-paper on information theory, stated that the entropy of English, at the l e v e l of single l e t t e r s , i s roughly Information Theory Psychology and Natural Language 36 1*2 b i t s ; : r a t h e r than the 4 b i t s or so that might ba expected from random *-words;-This redundancy would seem to be a result of the form/content' dualism; i f content alone were the determining factor, redundancy' should be eliminated. Obviously, t h i s finding cannot be applied mechanically, but the p r i n c i p l e would seem to be'sound: at-any point i n the discourse, we can predict, to a- greater - or "lesser degree; what w i l l follow. Osing the inherent • 'structure of language, we are able to 'guess' i n t e l l i g e n t l y . " In A.I. terms, - t h i s " i s a case of the standard in t e r a c t i o n between top-down and bottom-up information (about which more w i l l Abe'*said ' l a t e r ) ; - Basically, analysis proceeds in a bottom-up mode; u n t i l some^guesses and predictions-can be. made, at which point i t s h i f t s to top-down. This interchange goes on at several lev e l s at "once;' ••! • Thus, the "characteristics would seem to be as follows: analysis can ; proceed - in-bottom-^up mode, as long as i t does not exceed 'the'limits of 'short term memory. At each point, however, we are"-- consciously • or unconsciously -- making predictions about-what i s to come"next. The problem of an e f f e c t i v e natural language- program; then;>would be to make these predict ions in a powerful and" non-explosive manner. - Several sources of information'are available for thi s purpose; the ones mentioned in the previous-chapter provide a start.-•• ' 1 - Interestingly;-a 'model1 for - language analysis embodying these- concepts already e x i s t s . Marcus (1974; 1975) has Implemented a system which s a t i s f i e s precisely these Information Theory Psychology and Natural Language 37 constraints.•<*His 'wait-and see' system i s allowed a limited lookahead (based on-predictions from the input so f a r ) , aft e r which time i t -"must'-make a decision. The -original'prediction system, from • which my work has been drawn, -was' that of Riesbeck (1974). His system, however, was designed"with d i f f e r e n t goals' i n mind. Further comparison w i l l be provided in the next chapter. III.2 Memory for Sentences Another very popular- (and relevant) domain of psychological work'concerns the'memory forms used to encode sentences, and the processes used-to construct•these forms. • The surge of transformational grammars i n the- aarly 1960s resulted 'in' ;' much e f f o r t , despite Chomsky • s pro t e s t s , to prove that a • sentence • was1, stored : i n t e r n a l l y - i n i t s base form ( i i e . deep structure)/ This'assumption, which came to be known as the'derivational theory 'of'complexity, has i m p l i c i t in i t the assumption ^that- a sentence requiring -more - transformations (i.e. - passive,* negative) 1 w i l l take longer to - comprehend than one requiring fewer"transformations (see Fodor and Garrett, 1966, or Ammon',- 1968)• i " * For" • a time,*-- t h i s outlook seemed v a l i d (for example Mi l l e r 'and McKean,•1964,-found that- passive and negative t r a n s f o r m a t i o n s h a v e an • -additive effect on cognitive complexity);However, "other- r e s u l t s soon began to contradict t h i s theory, and"it was'soon'abandoned. Interestingly, at about t h i s time, Chomsky (1971; Memory for Sentences Psychology and Natural Language 38 o r i g i n a l l y * published ' i n " 1968) retreated from his e a r l i e r position;•'and admitted that the surface structure of a sentence does have an effect on 1 meaning — i . e . , that the deep structure does not adequately 1represent'the meaning-of the sentence. One interesting r e s u l t produced in t h i s paradigm came from the work of Harks and K i l l e r : (1964), who-were investigating the effects -of syntax' and semantics on sentence-comprehension. The de r i v a t i o n a l ' theory of; complexity would-predict that the only c o n t r o l l i n g variable'would be-syntax, whereas some of the more modern" • theories'would 'put the - emphasis on semantics. Harks and M i l l e r -found' that-"either 'syntax or semantics alone worked about equally well (in enabling subjects to understand sentences), but that the two together' produced much better • results. This i n d i c a t e s 1 t h a t subjects use*whatever information i s available to aid them i n the'comprshension process. • The •question now becomes "what do people store as a re s u l t of understanding sentences?". A number of experiments were run to test this. • '•• - The f i r s t , and perhaps most important, of these was that by Sachs (1967) , whodiscovered that subjects were able to r e c a l l and recognize the meaning of'• sentences guite well, but were unable- to"1 detect changes in surface syntactic form. This would indicate that what i s •stored i s some sort of prepositional meaning' representation,* ' rather' : that:: a s y n t a c t i c a l l y oriented structure; - •'- ;" ?•-•»-'-<•?••-. - -Bransford and • Franks' (1971; Bransford, Barclay, and Franks, 1972) carried t h i s one - step further, and discovered that Memory for Sentences Psychology and Natural Language 39 subjects 8 were unable •• even' to distinguish the separation of i n d i v i d u a l s e n t e n c e s ; I n their experiment-, -the- subjects were given'•'-a ' series'-of' short sentences, and-then asked to -recognize various ' combinations of these;- In general; - there *was a marked tendency for -subjects to 'coalesce* 'the meaning, and to 'remember* more'holisfic units than had actually been observed. Independent and rigourous v e r i f i c a t i o n of this work was provided by Johnson-Laird (1970), Bock & • Brewer (1974), and Griggs (1974) v . » • • K i n t s c h e t a l (1976) developed t h i s to- an even greater degree; ; A f t e r - ' t e l l i n g the- subjects a short (one-paragraph) story, ' they- discovered that the- subjects were unable to distinguish--between • • - • ••• " • ••'•••<- ••• i) Information e x p l i c i t l y given in-the text; i i ) - information inf e r r e d from- the text; • • ••••• i i i ) information previously known by the subject. Thus; i t seems -that the information had become t o t a l l y i a i § a £ l i § i . into the subjects* memory, to the point that i t could no longer be-distinguished as a separable text. • There " i s other* evidence"in t h i s 'regard; space precludes a more exhaustive analysis. E s s e n t i a l l y , the various experiments mentioned- seem-to point'to one conclusion: what is-stored after the analysis 'of 'a 'sentence' Is not any sort" of syntactic structure, surface or deep, but rather some abstract semantic form, which i s f u l l y integrated into the subject's memory. Memory for Sentences Psychology and Natural Language 40 III.3 Memory for Discourse • With the•domain of the sentence having been, in some way, disposed of; i t - i s productive to turn our attention to the more complex area of psychology and discourse. As Frederiksen (1976, p. 1) points out, >-... 'Most of'the knowledge which humans acquire in a l i f e t i m e derives ... from organized information units which • possess a high degree of structure' • Thus; ' i t " i s imperative-that any psychological research i n language examine the"problem'of discourse - comprehension. A number of obvious-questions can be raised here: i) What - i s the 1 form of the memory structures resulting '"- from discourse? •• -ii) How are these structures b u i l t ? • i i i ) -Why are - certain parts of a discourse remembered better than others? These probably cannot be answered separately, i f indeed they can be answered-at a l l . "In -general', •* the* facts discovered in the previous section can serve as a star'ting point. If seems that subjects do not store'a discourse ( i . e . ; an expository or narrative paragraph) as •a -series .of is o l a t e d sentences, or even as a series of connected sentences, but rather as some t i g h t l y interconnected unit. '•• •• -"*•' ' •. -• Fillenbaum (1971) studied the eff e c t s of conjunction, and observed that subjects tend to retain the conjunction when i t i s salie n t to the story (for example^ 'and' has more cohesive power Memory for Discourse Psychology and Natural Language 41 when used i n a temporal rather than a conjunctive'sense). This, then, -is evidence for the coherence mentioned e a r l i e r . r • •• Clark - and-"Clark • (1968) also found that temporal ordering and causality play a ' s i g n i f i c a n t part in the 'memorability' of sentences. •" '• • • .. ,-, Lesgold (1972) performed s i m i l a r experiments on coherency, and found-that'Ergnominalization i s important in determining the u n i t y o f ' t h e " textv • ''Basically, "'use of pronouns "to refer to previously-mentioned objects 'resulted i n • better r e c a l l of stories than use of d e f i n i t e noun phrases or other referring devices i 1 ' ^ • .• •  • The variables- mentioned -so far deal with what we previously called' "the 'micros tructure- - of • the text - - " - t h e - manner i n which i n d i v i d u a l pieces-are connected together. The work to t h i s point has confirmed the assumptions made by l i n g u i s t s , such as P h i l l i p s ; •• 'Grimes; 1 and " Schank, concerning connectivity (pp. 12-14) . We w i l l deal next"with'the macrostructure — the o v e r a l l unity which makes a text cohesive. One of the-experiments in t h i s area "was' performed by de V i l l i e r s (1974) .-« Ha found that, when given a text, subjects recalled' i n d i v i d u a l sentences according -to • their r e l a t i v e §§ii§HS§" to•"•the text" (where 'salience'- was subjectively determined)." When given •the same sentences in i s o l a t i o n , subjects^'tended t o - • r e c a l l them based on their i n d i v i d u a l S2H£E§i§3.ssS' (i;e. -' how' much' of an image could • be created) . Interestingly, Sulin and ' Dooling (1974) and Meyer (1975) Memory for Discourse Psychology and Natural Language 42 produced the-same r e s u l t s in different ways. Perhaps t h i s can be-construed as t h e - f i r s t evidence for the hypothesis :that "a text is•more than the sum- of i t s component sentences." What-remains;• of course, i s to get a better grasp of the notion of 'salience'. --A : number of experimenters- have worked-on this i n recent years. Probably the' most precise was -Kintsch (1974, 1975, 1976) ; who ;-devised-a propositional system of representation for meaning. ' The propositions could be embedded; i . e . the ones at the top* level provided the central thread of the•stony, the next l e v e l down-described- the-top-level, and so-on. •••-Whether or not his •system- i s ' c o r r e c t - (and there i s much dispute), i t at least provides one methodJ of ' judging importance: the propositions higher in the'tree-are'judged to be the most important. - In his experiments; Kintsch foundthe expected- -results: higher-level propositions" are "remembered better. This-confirms the theory that importance a f f e c t s memorability. •* ... These -results were derived independently by Mayer (1975). Working•from"a"model proposed by Grimes (but vary s i m i l a r to Kintsch*s) ,= Meyer - produced- the same results: higher-level propositions a r e v r e c a l l e d better. - 'Schank - (1974b)- rdeveloped- an interesting model of the process of"paragraph-understanding, which supported the same conclusions,*although no experimental v e r i f i c a t i o n was provided. Frederiksen (1975;" 1976) devised - a complexsystem of text representation, 'based on Halliday's work. E s s e n t i a l l y , he i d e n t i f i e d six lev e l s of information: concept, r e l a t i o n a l Memory for Discourse Psychology and Natural Language 4 3 t r i p l e s ; ' -systems'—proposition, •• r e l a t i o n a l •• system, dependency system; The dependency system i s the highest • l e v e l , embodying l o g i c a l , - temporal; - and causal- r e l a t i o n s ; ; - i t is s i m i l a r to the top l e v e l s of both Beyer's and~Kintsch's -systems. Frederiksen again'• "found" that' information 1 at the-highest l e v e l i s re c a l l e d best. Interestingly;'he also derived two-other'results: i ) . t h a t subjects"perform'a certain-amount of semantic and i n f e r e n t i a l processing at input- time (rather than • -waiting'-until r e c a l l time) ; ••>.-*ii)--that* information already i n the - system a f f e c t s the - acquisition • -of- l a t e r information; Frederiksen i d e n t i f i e d three - a c q u i s i t i o n methods: (1) selective processing (2) s l o t f i l l i n g (3) superpropositional inferences. Much'evaluation of t h i s work remains to be done. •-Interestingly, ' the f i r s t experiment in the f i e l d was the only one'which-failed to -produce the expected results. Crothers (1972) - delineated an experimental method which has been used by a l l succeeding'researchers: • .•••.•••<.-. i) "formulate a l i n g u i s t i c description of the structure of - - prose;':'"' •-•-' '" '-* - * -i i ) conduct - r e c a l l experiments;--analyzing both the passage and•the r e c a l l s of i t according to the theory • developed in the f i r s t stage; - - • i i i ) -"derive the empirical r e l a t i o n between structure and r e c a l l ; Memory for Discourse Psychology and Natur a l Language 44 iv)--design'a-process model of memory t o account f o r the f e a t u r e s d i s c o v e r e d . Crothers'-" own experiments, however, produced r e s u l t s which contradicted"his-expectations --•primary subtrees ( i . e . the top leve l ) - were not'-recalled'any b e t t e r than other l e v e l s . C r o t h e r s of f ered"some' explanations ' f o r • h i s f i n d i n g s , and Mayer (1975) g i v e s : others;-' ''ftf any ' r a t e , i t : i s : s t i l l to-be accounted f o r . Perhaps this-part"can-now be summarized.— -From the evidence given;-''•-it-'-seems'*" t h a t -text - i s s t o r e d in-memory i n a u n i f i e d , h o l i s t i c manner:" Subjects are aware of t h i s u n i t y , and i n f a c t i t 1 has "'a' " g r e a t ' " d e a l ' 'of " e f f e c t on t h e i r - u n d e r s t a n d i n g . In p a r t i c u l a r ; " d i f f e r e n t • parts? of- a text-tend-'-to - - be r e c a l l e d t o d i f f e r e n t degrees, depending on t h e i r importance. I I I . 4 Organizational Schemata A r e l a t e d , - though - e a r l i e r , p i e c e of work, was B a r t l e t t ' s (1932) study; Remembering. B a r t l e t t used- a long process of • s e r i a l - r e p r o d u c t i o n • - •• ( i . e ^ , - having s t o r i e s t o l d and r e t o l d through a chain o f " p e o p l e ) , t o observe the - form that a t e x t s t r u c t u r e takes when -allowed to adapt -it s e l f . - - T h e -material he used-came from va " l i t t l e - k n o w n Eskimo f o l k t a l e , The gar of the Ghosts. - B a r t l e t t observed s e v e r a l e f f e c t s i) omission-' n •'•-•-••" - d e t a i l "which' was not s a l i e n t to the s t o r y disappeared - f a i r l y q u i c k l y ; r i i ) r a t i o n a l i z a t i o n O r g a n i z a t i o n a l Schemata Psychology and Natural Language 45 -the story, being'mythical, had a number of peculiar and inexplicable'occurrenees; subjects soon modified these, or invented"events to account for them; " i i i ) 'transformation'of d e t a i l ' ^where'-names; • locations, etc., were'unfamiliar, subjects changed'them int o something more recognizable; iv) reordering of events-- the -events tended to be rearranged, so that more important ones were given more prominence in the story; v) bias'toward the concrete - • -" * abstract" concepts tendsd to be replaced by more concrete ( i . e . more imagistic) ones. In general;"--'Bartlett - observed • that stories were r a d i c a l l y modified, undergoing vast changes i n the- t r a n s i t i o n between-hearing "and"telling. " Fro ar t h i s , " he concluded that memory i s an active' "rather'than ' passive ' process — " l i e ; ' i t continually reorganizes" 'i t s -own "contents; He borrowed the term 'schema', and described t h i s as: "' !" -"• •'-••• •- • 'an; active "organization of past reactions, or past experiences; which must always by supposed to be operating in any well-adapted organic response' (Bartlett, 1932, p. 201) - ~ • These schemata Influence "both ; our perceptionand r e c a l l , so that remembering;''rather ' than being mere• r e t r i e v a l , takes on a SSSStructivist' nature. •• ' • "Much-of * B a r t l e t t ' s work i s open to c r i t i c i s m , but the general tenor " i s ' probably"valid: memory-should be viewed as an active, self-organizing'system, which mediates and rearranges Organizational Schemata Psychology and Natural Language 46 i t s own contents-I l l . 5 Other Work There - ' S t i l l remain a few int e r e s t i n g r e s u l t s , which do not f i t into the^categories discussed so far. Hunt' a n d - P o l t r o c k (1974), working within the •information-processing' approach, provided" evidence f o r the existence ""of •-' separate 1 •••buffers' (memories) a These were described-as short-,'• intermediate-, and long-term-memory. This work served e s s e n t i a l l y to provide the*new paradigm with the same basis as'the-older,'associative approach. " Kintsch *-(1975) ' derived some results which have direct bearing on A.I. work. ;He•found that as the-number of nominal concepts' ( i ; e;* things) i n - a " ' t e x t increased, the rate -of processing"was'degraded. This corresponds to the standard A.I. notion of semantic memory search, wherein a greater number of nodes r e s u l t s " i n slower processing. In-also relates to the role of participants i n - a discourse; a small number of participants, used-over and over, Is i n t u i t i v e l y more cohesive than a larger cast. This chapter has been a review of psychological work, to extract soma'common themes'. ' In conclusion, we' can i d e n t i f y a number' of findings •' from psychology which are of relevance to work in natural language: i) People seem to process language with a 'chunking' Other Work Psychology and Natural Language 47 approach; the-input i s organized into coherent units "as' soon-as "enough information i s available. i i ) •Understanding i s - p r e d i c t i v e ; at each stage, we are expecting 'something to follow. i i i ) People-use whatever information i s available — be i t syntactic, semantic, or pragmatic -- to understand language.' -iv) What - i s : ' remembered from a discourse i s not any syntactic structure> surface or deep, but rather some highly-abstracted meaning. v) Certain features tend to unify a-discourse, including causal and temporal connectivity-." vi) Some parts of-a discourse — notably the most s a l i e n t "are'remembered-better than others; • v i i ) - Memory-is an'active process, rather than a passive receptacle. These* conclusions" should- be kept inmind as we deal with the proposed model for analysis, discussed in the next two chapters. Other Work 48 C o m p u t a t i o n a l P r e r e g u i s i t e s IV.1 The R e p r e s e n t a t i o n IV.'1.'1 E x t e n s i o n s - t o C o n c e p t u a l Dependency IV.2 A n ' A n a l y s i s system IV.2.1 Comparison of Systems IV.2.2'The P r e d i c t i o n System T h i s c h a p t e r - i s an e f f o r t - t o e s t a b l i s h the groundwork f o r a complete model f o r d i s c o u r s e a n a l y s i s , t o be p r e s e n t e d i n the next c h a p t e r . "To a c h i e v e - t h i s , I . s h a l l develop some o f the computational-^ r e q u i r e m e n t s o f a d i s c o u r s e system, and p r e s e n t some p o t e n t i a l - s o l u t i o n s . • The ch a p t e r ' i s d i v i d e d i n t o two p a r t s ; In t h e f i r s t , I d i s c u s s the q u e s t i o n of r e p r e s e n t a t i o n , the v a r i o u s p o s s i b i l i t i e s a v a i l a b l e , and t h e re a s o n s f o r r e q u i r i n g an e f f e c t i v e r e p r e s e n t a t i o n ; The second i s a comprehensive o v e r v i e w of p r e v i o u s major a n a l y s i s - s y s t e m s (at the l e v e l of s i n g l e s e n t e n c e s ) * These are compared w i t h ' r e s p e c t t o both good and bad f e a t u r e s , and the m o t i v a t i o n 'behind my - p a r t i c u l a r c h o i c e ( p r e d i c t i o n s ) i s e x p l a i n e d ; ' ;* The • p r e d i c t i o n system i s then d e s c r i b e d i n d e t a i l , and an example of i t s - o p e r a t i o n i s p r e s e n t e d . •-... - T h e - " i n t e n t ' -here' - i s t o r e v i e w the r e q u i s i t e s of d i s c o u r s e a n a l y s i s ; and p r o v i d e t h e ' c o m p u t a t i o n a l t o o l s n e c e s s a r y f o r the j o b . • - •• • • -Throughout t h i s c h a p t e r , I w i l l make r e f e r e n c e t o the C o m p u t a t i o n a l P r e r e q u i s i t e s 49 p s y c h o l o q i c a l •:' c r i t e r i a j u s t m e n t i o n e d , - i n d i c a t i n g , how t h e p r o p o s e d s y s t e m s a t i s f i e s (or f a i l s t o s a t i s f y ) them. TV.1 The R e p r e s e n t a t i o n A s ' m e n t i o n e d ; one o f t h e m a j o r q u e s t i o n s i n d i s c o u r s e a n a l y s i s " - (and • i n any : n a t u r a l l a n g u a g e work) i s t h a t o f r e p r e s e n t a t i o n - t h e 'form i n * w h i c h t h e i n p u t i s f i n a l l y s t o r e d . T o ' d e c i d a ' this,";a-' number o f q u e s t i o n s must be a n s w e r e d : i ) -What'is t h e - £ u r £ o s e - o f t h e r e p r e s e n t a t i o n -- i s i t f o r e x e c u t i o n ( W i n o g r a d ) , q u e r y (Woods, P e t r i c k , Simmons), o r j u s t s t o r a g e and e x a m i n a t i o n (Rumelhart and Norman, S c h a n k ) ? ' ~* •- • ••• •' • • i i ) How b r o a d a domain i s t o be d e a l t - w i t h b l o c k s w o r l d ' ( W i n o g r a d ) , c l o s e d d a t a base (Woods), m o t i v a t e d humans * (Schank)-, o r s o m e t h i n g more g e n e r a l ? i i i ) •"How deep. >is' t h e " r e p r e s e n t a t i o n -to go • — i s i t t o be s u r f a c e - o r i e n t e d ( K i n t s c h , F r e d e r i k s e n , Simmons), c o n c e p t u a l l y deep " ( S c h a n k ) , o r s o m e t h i n g i n between (Rumelhart and Norman, K i l l e r , W i l k s ) ? W ith r e s p e c t " t o ' a n a l y s i s o f • n a r r a t i v e d i s c o u r s e , - the q u e s t i o n s can be answered:""" •' " • •'•<•-••- • -i ) m a i n l y ' s t o r a g e " - - my i n t e n t h e r e i s j u s t t o show t h a t a meaning- can be r e p r e s e n t e d , w i t h o u t s p e c i f y i n g any p a r t i c u l a r ' u s e ; ' .. i i ) r e l a t i v e l y ' g e n e r a l -- e x c e p t t h a t the c h i l d r e n ' s s t o r i e s we w i l l d e a l w i t h u s u a l l y f e a t u r e T h e . R e p r e s e n t a t i o n Computational Prerequisites 50 predominantly concrete forms ( i . e . one rarely encounters' abstract topics) ;• i i i ) ' probably' deep •-• more w i l l said about this l a t e r , but there'are'very qood reasons for favouring a •canonical form 1, i f one can be found. Given these - c r i t e r i a , the' !system I settled on was Schank's (1973; 1975)' Conceptual Dependency. This e f f e c t i v e l y meets- the goals- -we have* established, and has a-number of s i g n i f i c a n t benef i t s r '•••-•• • <•*-'•••'.• -,••:•• Conceptual Dependency" (CD)'is an extension of the idea of dependency grammars- (Hays',- 1964) , but i s based upon the notion that-; --there"is J a' small' number of = 'primitive'> acts '(in t h i s case, about'12) , i n terms'of which; everything can-be expressed. This reduction to p r i m i t i ves greatly f a c i l i t a t e s ^certain aspects of language analysis;-" since any -two sentences-'which -are paraphrases of - each • other• - are guaranteed to- have - the-- same underlying representation-''(oneof Schank's-fundamental assertions). In addition to this -taxonomy of verbs, CD also features a r i g i d "•"Set'; ";6f''••-cases-;' -•There"1' • are • -four*-1" (object, • instrument, direction;' and mrecipienty; of which each verb must be associated with two or"threes(these* are 'deep' or conceptual-cases, not• t o be '• 'confusedly-with" 'Fillmore' ;s' ; (1-968) -more -^surface^oriented system)-.'Another'"restriction i s ' that the - * 'instrument '• case, rather- "than"'-beingia'''simple"nominal as in^most- systems, must be another "conceptualisation ••'(i.e.'/another- act)1.";" -In" -addition to the cases, various modifiers, describing tense, location, The Representation C o m p u t a t i o n a l P r e r e q u i s i t e s 5"! manner, PART-OF, e t c . , are used. One complete qroup ( i . e . a c t o r , a c t i o n , and a s s o c i a t e d cases) i s r e f e r r e d to as a SSHf^SSEtyaAization, and normally corresponds to on event. Perhaps an example would help here. 'John a t e the i c e cream' would be r e p r e s e n t e d as the c o n c e p t u a l i z a t i o n ; John John <=> INGEST ice cream 4- J do to spoon • (Schank and Colby, 1973, p. 200> F i g u r e 2 - A Sample Conceptual. Dependency Diagram T h i s has been a c u r s o r y overview of CD. Those u n f a m i l i a r with Schank's work, or w i s h i n g more i n f o r m a t i o n , are i n v i t e d to read any o f h i s many w r i t i n g s on the s u b j e c t . CD, as a r e p r e s e n t a t i o n system, has both good and bad p o i n t s . Davidson (1976b) p r e s e n t s a comprehensive e v a l u a t i o n of the system. I s h a l l i n d i c a t e a few main p o i n t s h^re. F i r s t the advantages. Probably the most i a p o r t a n t (to Schank, at l e a s t ) i s t h a t CD i s (purportedly) l a n g u a g e - f r e e : t h a t i s , i t r e p r e s e n t s i n f o r m a t i o n a t a s u f f i c i e n t l y deep l e v e l t h a t i t can be used with any language. Another i s the f a c t t h a t the c o n c e p t u a l l y dee? nature of CD has a c e r t a i n amount of p s y c h o l o g i c a l v e r i f i c a t i o n . I t s a t i s f i e s p o i n t (iv) mentioned at the end of the p r e v i o u s s e c t i o n — t h a t the r e s u l t of understanding language i s some h i g h l y - a b s t r a c t e d semantic s t r u c t u r e . T h i r d , and p r o b a b l y most important f o r our purposes, i s the The R e p r e s e n t a t i o n Computational P rere g u i s i t e s 52 fact that fthe"system'is computationally v a l i d ; r hat i s , despite the"disputes regarding the"effectiveness of CD- fr om a l i n g u i s t i c point "of view, ' -computational'- l i n g u i s t i c s 5 (inc l u d i ng the work described here) remains'-at such an unsophisticat ed stage that these-subtle flaws are -not c r u c i a l . • « ; - -•• ' An -interesting feature "'of CD; i s i t s a b i l i t y to c haracterize the- * r e l a t i v e - salience" -of -different parts*1 of • a sto ry. Schank (197Ub> "' ppv« 26-27) - formulated • quite * s p e c i f i te- rules for recognizing- and representing- : the important poi nts of•a story. This f u l f i l l s ' psychological" c r i t e r i o n " ( v i ) - ••. . that the importance -of an event to- the- o v e r a l l story affs cts how well i t i s remembered v ' - ' 1 i---= ••• •-••>';•• o :. i . . . . . . . '• A f i n a l - b e n e f i t - o f 1 CD, related to the-guesti on o f canonical form, i s "that "it'-serves" to- capture • r e g u l a r i t i e s • of language. That i s , ( " 'equivalence • and similarities-of-mean ing become much more obvious when everything i s represented in a uniform formalism; ' • •There--are also some'disadvantages of -CD, t whic h must be mentioned;" ••" First;- there- i s the question - of ac eura s j ; much of the meaning-of language' (especially the more subtle nuances) seems to'disappear- in!,the> t r a n s i t i o n into primitives. »I s l i c e d the meat with the knif e ' surely says more than: The Representation Computational Prerequisites 53 I'USS-UY -> meat I 4=^ M O V E hand t t ftcONT in G R A S P f o back & forth knife flif—> slices meat # ' < whole knife ~~I POSS-BY I h a n d < - l l (Schank and Colby, 1973, p. 228) Figure 3 - Another Conceptual Dependency Diagram Schank himself admits to t h i s problem (1975a, p. 32): •one must be c a r e f u l not to lose information i n a conceptual analysis (that i s ' k i s s 1 i s more than just •HOVE l i p s towards*)' However, as mentioned, we require only computational adequacy ( i . e . , the a b i l i t y to function e f f e c t i v e l y within the r e s t r i c t e d sphere of current computational l i n g u i s t i c s work), and t h i s has been achieved (as evidenced by the fact that the MARGIE system functions e f f e c t i v e l y ) . Another "problem l i e s with Schank's insistence on a r e l e n t l e s s and immediate expansion into primitives. That i s , every sentence must be expressed i n terms of the primitive acts, with no opportunity to delay processing ani wait for disambiguation. This i s a f a i r l y serious flaw i n CD, and i t r e s u l t s i n some n o n - t r i v i a l d i f f i c u l t i e s . In our r e s t r i c t e d domain, however, we w i l l be able to avoid the problems occasioned by t h i s approach. Note that t h i s l a t t e r d i f f i c u l t y contrasts with the •wait-and-see' approach to analysis, mentioned e a r l i e r . I n t u i t i v e l y , humans do not rush headlong into i n t e r p r e t a t i o n , but make decisions only when they have s u f f i c i e n t information. The Representation Computational Prerequisites 54 This concludes thercommentary on CD in i t s standard form. For' purposes-'Of" discourse--analysis, I have made a number of minor modifications, mentioned next. IV:1.1 Extensions to^Conceptual-Dependency To' handle • the extended 'requirements of discourse, several minor changes- have'-been: made to CD. These are' a l l ad hoc, and no theoretical""'motivation i s claimedr- the changes are made s p e c i f i c a l l y for the goal of discourse analysis. a)- theme-'•• •;'"-,"-"-'''f"'""';- ---•••'-<:•---.-'.-,. • • -One * "-of -the 'basic flaws -in' CD i s i t s f a i l u r e to ret a i n any indication'-of the"theme -of•a sentence; as-with transformational grammar;'the -difference-between- passives and-actives i s ignored. This' lack* i s ' b a r e l y noticeable at-the level-of single sentences, but becomes c r u c i a l 'when -dealing with extended ^discourse. Since the theme-can'-'-be1"detected 'in ; a straightforward' manner (see ' staging • ,*-chapter-TI) !,'it should be recognized and retained. •' - '•-I'have wdealt" with'this- in--a s i m p l i s t i c manner; an extra case; -•> ^ marked*"— 'theme','••- - h a s - b e e n added to the: standard conceptualizations This can point to any of- the other elements of " the ^conceptualization-: i.^e. actor, action, object, instrument; r e c i p i e n t ; or di r e c t i o n . " For example;""-"-" •-- —>--:.-•John-gave Mary "the.book." -- The book was'given to Mary by John.' • Mary was given the book by John, would a l l be represented by: The Representation Computational Prerequisites 55 ' o j-> Mary John <=> ATRANS <— book <-| " " - ' * - ' • : i -< -except-that ;the' theme'would be the ACTOR, OBJECT, and RECIPIENT, respectively. ..-: < Interestingly; - ' Halliday (1970) d i f f e r e n t i a t e d between psychological subject, grammatical subject, and l o g i c a l : subject — corresponding roughly to theme, ACTOR, and surface subject. • Obviously," t h i s ad hoc solution does not completely solve the problem, but i t provides enough of- a basis f o r us to continue, b) 'then' "links Schank-was' one of "the - f i r s t researchers - to examine connected discourse,' and he did produce a coherent taxonomy of linkage between conceptualizations (Schank, ; 1974a). This set was based s on"causality, with nine different-types of causation (including • reason; enablement, and i n i t i a t i o n ) 1 i d e n t i f i e d . Unfortunately> he stated'later (Schank, 1974b, p; 16) that these types were the only connections between conceptualizaiton in a S t o r y i -: •-• - - - '- <;••!•••••.:••••• < - - -. . •:. - .. - ; -. ••. .. .. This'-seems-excessively r e s t r i c t i v e ; what i s lacking i s the simple* notion -of-' a •' 'then'-' l i n k -- -i.e; of one event just following another. 'Even'Riesbeck (1974) 'recognized t h i s , and P h i l l i p s * -(1975) arrived* at the same conclusion independently • •' This - problem ••-is another manifestation" of Schank's insistence upon expansion into primitives. That i s , even though The Representation Computational P r e r e q u i s i t e s 56 i t may be p o s s i b l e a t some very deep conceptual l e v e l t o i d e n t i f y c a u s a l l i n k s between a l l events, there ara f r e q u e n t l y times when we do not wish t o f o r c e t h i s degree o f s p e c i f i c i t y . Schank r e c o g n i z e d ' t h i s , and attempted to amend i t by p r o v i d i n g f o r an ' u n s p e c i f i e d c a u s a l c o n n e c t i o n • , which i s 'unaxpanded and u n d i r e c t i o n a l * (Schank, 1974b, p. 38). I n t u i t i v e l y , t h i s i s a •then' c o n n e c t i o n . The s o l u t i o n I propose here, as i n the p r e v i o u s case, i s s i m p l i s t i c ; I have a r b i t r a r i l y p o s i t e d a 'then* l i n k which can connect two c o n c e p t u a l i z a t i o n s , with the obvious meaning. T h i s w i l l be a s o r t of d e f a u l t l i n k , i n t h a t i t w i l l be a r b i t r a r i l y i n s e r t e d whenever no other connection can be found. For the kinds o f simple n a r r a t i v e s we s h a l l d e a l with, t h i s assumption i s probably v a l i d . As before; t h i s i s by no means a complete s o l u t i o n , but i t i s adequate f o r our purposes. Note that i t r e f l e c t s p s y c h o l o g i c a l c r i t e r i o n (v) — t h a t c e r t a i n l o c a l l i n k a g e s tend t o u n i f y a d i s c o u r s e . c) other l i n k s A n - e f f o r t of r e p r e s e n t a d i s c o u r s e runs i n t o problems when i t i s approached from a m i c r o s t r u c t u r a l d i r e c t i o n , as we have done so f a r . Hhat i s needed i s some more g l o b a l c h a r a c t e r i z a t i o n o f a d i s c o u r s e , which captures the o v e r a l l p i c t u r e . > The s i g n i f i c a n t point f o r our work l i e s i n the need t o be ab l e to repr e s e n t a 't e x t grammar' i n CD terms. For example, we Tha Representation Computational prerequisites 57 must decide how to l i n k the setting into the rest of the meaning structure. Is i t a causal link? A 'then* link? I n t u i t i v e l y , neither of these w i l l do, and additional- representational structures w i l l be needed. I w i l l say no more about t h i s here, preferring to develop things on an empirical basis, as needed. d) u n i f i c a t i o n of representation Schank (1974b , p. 16) remarks that '[after a paragraph has been input] the conceptual dependency representation of each sentence i s included [ i n the r e s u l t ]» However, as Bransford and Franks (1972) showed, subjects often unify the memory into a more"holistic mass, within which i t i s impossible to distinguish the o r i g i n a l sentence. I have taken t h i s finding as a guideline, and t r i e d to integrate the meaning representation as much as possible. To conclude the section on representation, t h i s i s an example of the meaning structure b u i l t from the short paragraph given i n 1 . 2 (pp. 1 0 - 1 1 ) : The Representation Computational P r e r e q u i s i t e s 58 tiary <=> DO A III c / II! / Sinner s e t t i n g / / "•• - " ' I . |>John o o |>ovan Hary <=>MTR&NS <-J <-- John <=>PTRANS < — c a s s s r o l e <- ! |< t< o !> Peter Mary <=> ftTRANS < — s a l a d <-1 !< A | then I o |> out-of (fridga) Hary <=> PTRANS <--dessert <-| j< i n ( f r i d g e ) A | then I |-> dinner <==j ready !-< / / denouement / !_ o |>down l o c everybody <=> PROPEL < — everybody <-| < t a b l e |< F i g u r e H - Representation f o r a Short Paragraph T h i s r e p r e s e n t a t i o n i s l i k e standard CD except t h a t the 'then 1 l i n k s are new, as are the ones l a b e l l e d ' s e t t i n g ' and 'denouement'. For the sake of s i m p l i c i t y the theme has not been shown here; i n t h i s example, i t i s unimportant. Thus we have a method of r e p r e s e n t i n g t e x t . Note t h a t the • s e t t i n g ' l i n k p o i n t s to the e n t i r e c e n t r e p a r t of the diagram. T h i s r e f l e c t s the f a c t t h a t the statement 'Mary was making The Representation C o m p u t a t i o n a l P r e r e q u i s i t e s 59 d i n n e r . ' h o l d s t r u e t h r o u g h o u t t h e s t o r y , and s e r v e s , i n a s e n s e , a s a ' g r o u n d i n g * f o r t h e e v e n t s m e n t i o n e d . To more e f f e c t i v e l y v e r i f y t h i s s t y l e o f r e p r e s e n t a t i o n , we would have t o show t h a t i t c a n be e f f e c t i v e l y u s e d f o r d i f f e r e n t p u r p o s e s : p a r a p h r a s e , summary, e t c . As m e n t i o n e d , t h i s i s not one o f my g o a l s . TV.2 An A n a l y s i s s y s t e m Now t h a t a r e p r e s e n t a t i o n has been o u t l i n e d , we c a n r e t u r n t o t h e q u e s t i o n o f a method o f a n a l y s i s . E s s e n t i a l l y , t h e c r i t e r i a f o r s u c h a method a r e : i ) t h e s y s t e m must•< be n o d u l a r , so t h a t t h e v a r i o u s i n f o r m a t i o n s o u r c e s c a n c o n t r i b u t e whan t h e y a r e r e l e v a n t , and l i e dormant o t h e r w i s e ; i i ) i t must r e f l e c t t h e g u i d e l i n e s s e t i n t h e p r e v i o u s c h a p t e r (on p s y c h o l o g y ) ; i i i ) i t must be f l e x i b l e , s o t h a t e x t e n s i o n s c a n be made e a s i l y ( t h i s i s r e l a t e d t o t h e f i r s t c r i t e r i o n ) . T h e r e a r e two o p t i o n s : e i t h e r t o t a k e an e x i s t i n g s e n t e n c e - l e v e l s y s t e m , and e x t e n d i t , o r t o w r i t e one f r o m s c r a t c h . I have c h o s e n t h e f o r m e r , f o r a c o u p l e o f r e a s o n s — f i r s t , t i m e c o n s t r a i n t s p r o h i b i t a c o m p l e t e d e s i g n e f f o r t , and s e c o n d , I b e l i e v e t h a t p r e v i o u s work i n c o m p u t a t i o n a l l i n g u i s t i c s has p r o d u c e d some n i c e r e s u l t s , f r o m w h i c h we-can b u i l d . T h us, I w i l l f i r s t p r e s e n t a b r i e f o v e r v i e w o f e x i s t i n g a n a l y s i s s y s t e m s , e x p l a i n i n g why I f a v o u r 1 one s u c h scheme An A n a l y s i s s y s t e m Computational P r e r e q u i s i t e s 60 ( p r e d i c t i o n s ) , then (in the next chapter) o u t l i n e the manner i n which i t c o u l d be extended t o handle d i s c o u r s e . IV.2.1 Comparison o f Systems Three systems w i l l be d e a l t with here, as being the most advanced work i n computational l i n g u i s t i c s to date: Woods's (1968; 1970) Augmented T r a n s i t i o n Nets (ATNs) , Winograd's (1971; 1972) PROGRAMMAR, and Riesbeck's (1974; 1975) p r e d i c t i o n - b a s e d system. woods's ATN system marked the beginning of what has been c a l l e d the ' f i r s t g e n e r a t i o n ' of computational l i n g u i s t i c s work. ATNs have a number of strong advantages as a method of language a n a l y s i s . F i r s t they are fo r m a l ; the grammar and i n t e r p r e t e r can be expressed'in'a r i g i d mathematical formalism, and the symbolic manipulations performed a re w e l l - s p e c i f i e d . Second, they are si m p l e ; the r e l a t i v e l y s m a l l set of o p e r a t i o n s makes design or i n t e r p r e t a t i o n o f a grammar q u i t e easy. L a s t l y , they are E§£§Ei2uous; the pa s s i v e nature of a grammar ( i t can be viewed as a s t a t e ' t r a n s i t i o n graph) f a c i l i t a t e s r a p i d understanding of the a c t i o n s and i n t e r a c t i o n s of a given grammar. ' ' There are, however, a number of drawbacks, many of which are p r i c e s p a i d " f o r the advantages g i v e n . F i r s t , the ATNs are s t r i c t l y s y n t a x - d r i v e n ; the semantics, i f any, are added on as a set of K a t z / F o d o r ~ s t y l e set o f r e s t r i c t i o n s . T h i s i s acc e p t a b l e to a c e r t a i n p o i n t , but c u r r e n t work i n computational l i n g u i s t i c s i n d i c a t e s t h a t s e r i o u s flaws e x i s t . An i n t e r e s t i n g An A n a l y s i s system Computational Prerequisites 61 example - i s J e r v i s ' (1974) implementation of an ATN; examination of her (large) grammar'reveals that semantic checking i s c l e a r l y divided from, and subservient to, the syntax, and also that such an approach leads to a somewhat unnatural structure. Another drawback of ATNs i s a certain degree of counterintuitiveness; these mathematical symbols and formal operations do not seem to r e f l e c t what we know about language comprehension. However, Kaplan (1972; 1975) has provided some interes t i n g evidence to support the claim that ATNs are a v a l i d model of human cognition and comprehension. The case i s s t i l l open; A th i r d drawback to ATNs i s t h e i r tremendous i n f l e x i b i l i t y . Arcs must be l a i d out beforehand, in a s t a t i c manner, and cannot be changed dynamically during processing* (Scarl (1976) has developed a system with arc-moving c a p a b i l i t i e s , but t h i s i s at best a patch.) That i s , there may be an arc which i s taken i n only one out of every hundred cases; t h i s arc s t i l l must be represented i n the grammar, although i t i s i n v a l i d 99 times out of 100 (the ordering of arcs at a node permits a cert a i n amount of 'tuning*; but t h i s does not change the fundamental s t a t i c nature of the system). The problems caused by t h i s s t a t i c organization w i l l become apparent l a t e r , when we return to the question of discourse analysis. The* l a s t flaw i n the system, one which i s currently out of favour i n natural language c i r c l e s , i s i t s use of automatic backup. The system, upon f a i l u r e , begins wildly undoing and revising i t s decisions, i n an e f f o r t to find a v a l i d parse. An Analysis system Computational P r e r e q u i s i t e s 62 S c a r l , i n h i s system, provides a number of b a s i c mechanisms to c o n t r o l the backup; but again the approach i s wrong; backup should not be automatic. Rather, i t should be designed by the programmer, so t h a t : i ) i t does not'occur i n a l l cases; i i ) when i t does occur, the system can make an i n t e l l i g e n t guess about where and why i t went wrong, and act a c c o r d i n g l y . In summary, ATNs would seem to be a good mechanism f o r g e t t i n g up a s m a l l grammar very q u i c k l y , and a l s o a qood p e d a g o g i c a l t o o l f o r an i n t r o d u c t i o n to computational l i n g u i s t i c s . They would not seem t o be s u i t a b l e f o r the k i n d s of l a r g e systems needed t o handle d i s c o u r s e . Winograd's PROGRAHHAR, designed at the same time as Woods's work, has a number of s i g n i f i c a n t advantages. F i r s t , and probably foremost, i s the f a c t t h a t the syntax/semantics/pragmatics d i s t i n c t i o n has been b l u r r e d (although not completely erased). Semantic checking i s done at each s t e p 'of the a n a l y s i s , and pragmatic r e s o l u t i o n ( i . e . data base'checking) i s performed each time a complete component i s b u i l t . T h i s approach, r e v o l u t i o n a r y f o r i t s time, has s i n c e gained p o p u l a r i t y i n computational l i n g u i s t i c s . (In f a c t , t h i s method c o u l d be i n c o r p o r a t e d i n t o an ATN system, but Woods's o r i g i n a l work d i d not i n c l u d e t h i s . ) Another' i n t e r e s t i n g p o i n t , which was true as w e l l of Woods's work, i s the f a c t t h at SHRDLO's grammar i s based on a An A n a l y s i s system Computational P r e r e q u i s i t e s 63 w e l l - s p e c i f i e d l i n g u i s t i c system -- i n t h i s case H a l l i d a y ' s (1967; 1970) systemic grammar. A minor advantage o f PROGRAMMAR, which i s s t i l l open t o debate, i s the f a c t that i t i s pr o c e d u r a l , r a t h e r than d e c l a r a t i v e . T h i s f a c t i n i t s e l f i s not s i g n i f i c a n t , but one can d e s c r i b e b e n e f i t s of both approaches (see Winograd, 1975, f o r a -thorough d i s c u s s i o n o f t h i s i s s u e ) . L a s t , and probably most s i g n i f i c a n t i n terms of advances over Woods's approach, i s the power. A PROGRAMMAR program can be m a d e • a r b i t r a r i l y complex, to do whatever a c t i o n s are d e s i r e d . True, an a r c o f an ATN can a l s o have these a c t i o n s added, but the p r o c e d u r a l nature of PROGRAMMAR provides more e x p r e s s i v e power. So much f o r the advantages. The primary disadvantage of the PROGRAMMAR formalism i s i t s u n r e a d a b i l i t y . SHRDLU's grammar i s so opague and incomprehensible t h a t Rubin (1973) found i t necessary to t r y t o f l o w c h a r t i t ; even t h i s was not easy. (Compare t h i s to the simple t r a n s i t i o n diagrams i n any of Woods's" papers.) This problem stems d i r e c t l y from the pr o c e d u r a l nature of the system -- nothing w i l l 'hold s t i l l ' l o n g enough t o p i n i t down. Another drawback i s the f a c t t h a t the system i s s t i l l s t a t i c ; the order and s t r u c t u r e of the grammar are s p e c i f i e d by the programmer duri n g the design phase, and cannot be modified d u r i n g e x ecution (the system b u i l d s Microplanner programs 'on the f l y ' i u n f o r t u n a t e l y , i t does not do the same with PROGRAMMAR programs). An A n a l y s i s system Computational Prerequisites 64 & flaw shared with Woods's system i s the heavy dependency on OElSE* This i s not merely the constraint that only grammatical sentences can be accepted, but also the fact that the i n t e r p r e t a t i o n of a sentence depends completely on the order of words i n that sentence. Thus, for any given 'meaning' (i . e . deep sentence), a l l possible surface manifestations of that meaning must be accounted f o r . Perhaps an example would help. I f we wish to parse the sentence: John ran down the s t r e e t . . the appropriate ATN grammar would look l i k e : Figure-5 (a) - A Simple ATN grammar (Note: the remarks about ATNs apply as well to PROGRAMMER grammars.) Now, suppose the adverb »guickly* were aided to the sentence. I t could appear i n any of four places, and s t i l l mean * e s s e n t i a l l y the same thing. But, to handle t h i s i n an ATN, we would need a grammar of the form: Figure 5 (b) - An Extended Grammar i . e . with four separate checks for the adverb. My contention i s that, since the adverb has the same meaning, and the same An Analysis system Computational Prerequisites 65 surface manifestation, i t should be recognized by the same piece of the system. This s i l l become clearer when we deal with predictions, below. To conclude the discussion'of PROGRAMMER, i t has one clea r advantage over Woods's system -- expressive power (used here i n the programming language sense of convenience of expression; the two systems are obviously formally equivalent) and one associated drawback — lack of perspicuity. Riesbeck's system of predictions i s a new, and somewhat badly described, method of analysis. B a s i c a l l y , a sentence i s analyzed via a set of predictions, spawned by the various words in the sentence. These predictions (or REQUESTS, as Riesbeck c a l l s them), consist of two parts: a test i . e . a l i n g u i s t i c construct to be scanned for i n the input) and an astion ( i . e . a set of functions to be performed i f i t i s found. In t h i s respect, the control structure i s s i m i l a r to that of demons (Charniak, 1972) or production systems (Newell, 1973). Again, readers wishing further d e t a i l s are directed to Riesbeck's thesis. Predictions, l i k e the other systems described, have t h e i r good and bad points. The primary advantage of the prediction system i s that i t i s not as dependent on order as the two previous' systems. For instance; to handle the 'John ran down the street' example, we need only one simple prediction; of the form ( (ADV) (RUMINATE MANNER) ) An Analysis system Computational P r e r e q u i s i t e s 66 (rouqhly t r a n s l a t e d , t h i s means * i f an adverb i s found, t r e a t i t as the BANNER of t h e sentence 1.) T h i s p r e d i c t i o n would remain a c t i v e throughout the e n t i r e sentence; thus (to r e t u r n t o our example), wherever i n the sentence the adverb appeared, i t would be picked up by t h i s one p r e d i c t i o n . This f e a t u r e w i l l be d e s c r i b e d i n more'-detail i n the next s e c t i o n . Another p o i n t , which 1 have been s t r e s s i n g , i s the f a c t t h a t p r e d i c t i o n s ' are dynamic. The grammar i s not s p e c i f i e d beforehand (indeed, i t i s never r e a l l y s p e c i f i e d at a l l ) ; r a t h e r ; p r e d i c t i o n s are added and d e l e t e d from the c e n t r a l p r e d i c t i o n l i s t - i n a continuous process, so that the flow of c o n t r o l i s never r i g i d l y e s t a b l i s h e d . The advantages of t h i s f l e x i b i l i t y w i l l become c l e a r l a t e r , when we d i s c u s s the a p p l i c a t i o n s of t h i s system t o d i s c o u r s e . A t h i r d advantage of the p r e d i c t i o n system i s t h a t , t o a c e r t a i n extent,' i t i s not as domain-dependant as the o t h e r s . That i s , both -ATNs and PROGRAHMAR were o r i g i n a l l y aimed at p a r t i c u l a r domains, but the very nature of Conceptual Dependency r e q u i r e d t h a t the p r e d i c t i o n s be as general as p o s s i b l e . Whether or not t h i s o b j e c t i v e was achieved i s q u e s t i o n a b l e ; at l e a s t the s p i r i t i s t h e r e . L i k e -Winograd's system, Riesbeck reduces the b a r r i e r between syntax and semantics. In t h i s case, i t i s almost e l i m i n a t e d ; - p r e d i c t i o n s are s e m a n t i c a l l y based, but can use s y n t a c t i c i n f o r m a t i o n whenever i t i s h e l p f u l . Thus, the two subdomains have rbeen more or l e s s merged. Riesbeck's system has a c e r t a i n amount of p s y c h o l o g i c a l An A n a l y s i s system Computational P r e r e q u i s i t e s 67 m o t i v a t i o n , " i n t h a t language comprehension i s t r e a t e d as a EESSSSS' ( t h i s f e a t u r e was a l s o p r esent, i m p l i c i t l y , i n Winograd's work); I t s a t i s f i e s p s y c h o l o g i c a l c r i t e r a ( i i ) --that a n a l y s i s ' b e p r e d i c t i v e — and ( i i i ) — that people use whatever- type o f i n f o r m a t i o n i s a v a i l a b l e to understand language. Compare t h i s t o the normal 'competence' approach i n l i n g u i s t i c s , which i s concerned with the speaker/hearer's i d e a l knowledge of h i s language, separate from any c o n s i d e r a t i o n s of memory, and from any attempt t o use the language (Chomsky, 1965, pp. 2-3) . A f i n a l advantage o f Riesbeck's system, which was a l s o present i n the other two, i s th a t i t b u i l d s a form to be retur n e d from the parse. In e a r l y computational l i n g u i s t i c s work, i t was assumed t h a t the r e s u l t of the parse would be a h i s t o r y of the s t e p s taken i n the p a r s i n g process. Woods changed t h i s ; " i n h i s ' system, o n l y the components t h a t are "BUILDQed" i n t o the form are re t u r n e d . In g e n e r a l , though, an ATN parse i s u s u a l l y s i m i l a r t o the s u r f a c e s t r u c t u r e . Riesbeck c a r r i e d t h i s • one step f u r t h e r ; i n the CD paradigm, the parse re t u r n e d r a r e l y , i f ever, bears any resemblance t o the o r i g i n a l sentence. (Note t h a t t h i s i s i n many ways a r e s u l t of the nature of Conceptual Dependency, r a t h e r than an i n h e r e n t f e a t u r e of the p r e d i c t i o n process;) Again, t h i s r e f l e c t s - p s y c h o l o g i c a l c r i t e r i o n ( i v ) , concerning a b s t r a c t i o n of meaning. T h i s concludes the l i s t of the good p o i n t s (in my opinion) of a p r e d i c t i v e p a r s i n g mechanism. There are a l s o some bad ones. An A n a l y s i s system Computational P r e r e q u i s i t e s 68 F i r s t ; as a r e s u l t of the magnificent and f l e x i b l e nature of the p r e d i c t i o n system, the grammar i s t o t a l l y Incomprehensible: That i s , s i n c e the system i s so f l u i d , and s i n c e the set of p r e d i c t i o n s a c t i v e at any one time i s not f i x e d i n any way, the grammar simply cannot be represented i n any convenient way; even the de s i g n e r of the system operates c l o s e t o Winograd*s 'complexity b a r r i e r ' (Winograd, 1975a) when t r y i n g to understand the p o s s i b l e i n t e r a c t i o n . . Another flaw i n - p r e d i c t i o n systems, which i s at the same time perhaps a v i r t u e , i s t h e i r heavy b i a s towards semantics. The problem here i s t h a t i t i s a c t u a l l y d i f f i c u l t t o express s y n t a c t i c i n f o r m a t i o n . There are ways around t h i s (these w i l l be developed 1 i n j t h e next s e c t i o n , when we look at p r e d i c t i o n s i n more d e t a i l ) , but i n general i t ' s a n o n - t r i v i a l - p r o b l e m . A t h i r d drawback i s the problem of c o n t r o l . As Charniak (1972) found out with h i s work on demons, the method of spawning p r e d i c t i o n s can o f t e n be e x p l o s i v e , r e s u l t i n g i n a l a r g e number of a c t i v e p r e d i c t i o n s . Thus, t h e r e are three major a n a l y s i s systems a v a i l a b l e f o r e x t e n s i o n to the d i s c o u r s e domain. I have chosen the l a s t of these ( p r e d i c t i o n s ) , f o r a number of reasons: i) i t i s more powerful; as j u s t d e s c r i b e d (and w i l l be shown i n the f o l l o w i n g s e c t i o n ) , i n f o r m a t i o n can i n ge n e r a l be expressed more e a s i l y i n a p r e d i c t i o n s t y l e system than i n any o t h e r ; i i ) i t corresponds most c l o s e l y to what we d i s c o v e r e d i n An A n a l y s i s system Computational Prerequisites 69 the last section about psychology; i n t e r e s t i n g l y , Hiesbeck*himself v e r i f i e d his theories only on the basis of 'introspection, but we have more concrete proof; i i i ) i t ' i s more f l e x i b l e ; as explained, the dynamic nature of the grammar permits a great degree of modularity, and powerful interactions between the predictions. Of the three, the l a s t one i s the major concern i n designing a model for discourse analysis. The dynamic f l e x i b i l i t y of the prediction system provides the control needed to handle a number of weak, but possibly relevant, I.S.»s. Perhaps an analogy i s useful here. Novice programmers, when learning- a new programming language, tend to s t a r t with a basic subset of that language which, while not powerful computationally, i s simple and e a s i l y understood. As they become more sophisticated, the programmers tend to make use of more es o t e r i c constructs which, while providing a great deal of power, increase the complexity of the program, making i t harder to understand. Thus with natural language analysis. An ATN provides a c l e a r , elegant formalism which i s e f f e c t i v e within a certain r e s t r i c t e d approach to' language. At the other extreme, a prediction-based system provides a great deal of power, at the cost of comprehensibility. More w i l l be said about the power of predictions in l a t e r sections. An Analysis system Computational P r e r e q u i s i t e s 70 IV.2.2 The P r e d i c t i o n System Having i d e n t i f i e d the p r e d i c t i o n system, perhaps I should d e s c r i b e i t i n d e t a i l . I see - t h e b a s i c mechanism of p r e d i c t i o n s as being language-independent, although the p r e d i c t i o n s themselves are o b v i o u s l y language-dependent* That i s , the c o n t r o l s t r u c t u r e d e s c r i b e d here could be a p p l i e d i n t a c t to any language, but the s u r f a c e l e v e l p r e d i c t i o n s would have t o change. E s s e n t i a l l y , a p r e d i c t i o n i s a s e r i e s of f i e l d s : ( TEST ACTION STRENGTH SOURCE CANCEL OBL? ) where: t e s t - i s a s p e c i f i c s u r f a c e - l e v e l c o n s t r u c t t h i s p r e d i c t i o n i s l o o k i n g f o r . example: (NP ANIMATE) says to look f o r an animate noun phrase. a c t i o n - i s the a c t i o n to be performed upon f i n d i n g the s p e c i f i e d component; t h e r e are three main types of i n s t r u c t i o n s : i ) what to do with the i n p u t j u s t found example: (RUMINATE OBJ) says to t r e a t the component j u s t found as the o b j e c t ; i i ) what to do with the p r e v i o u s l y - b u i l t meaning base example: (REPLACE :SENTMEAN: (OBJ (EXTRACT ACI3R))) says t o take the concept which was p r e v i o u s l y thouqht to be the a c t o r , and put i t i n t o the o b j e c t s l o t ( t h i s would occur, f o r example, i n the a n a l y s i s of a p a s s i v e s e ntence); An A n a l y s i s system Computational P r e r e q u i s i t e s 71 i i i ) what to look f o r next example: (PREDICT (PG TO) (RUMINATE RECIP)) i f i t appears i n the ' a c t i o n ' s l o t of a p r e d i c t i o n , t e l l s the system to spawn a new p r e d i c t i o n , l o o k i n g f o r a p r e p o s i t i o n group beginning with ' t o ' , and i f i t i s found, to p l a c e i t i n the ' r e c i p i e n t ' s l o t ( t h i s would occur, f o r example, a f t e r the verb ' g i v e ' was d i s c o v e r e d ) ; Thus, the a c t i o n s can encode the two types of i n f o r m a t i o n a v a i l a b l e from I.S.'s — what t o look f o r next, and what to do with what we've al r e a d y got. s t r e n g t h - i s an estimate of the 'importance* o f the p r e d i c t i o n (on a s c a l e 0-10) , assigned when i t i s c r e a t e d . The key here i s t h a t the p r e d i c t i o n s are kept i n an o r d e r e d - l i s t , so t h a t l e s s important ones can be e a s i l y ignored (to a s m a l l degree, t h i s corresponds t o the o r d e r i n g of a r c s l e a v i n g a node i n an ATN). I t should be mentioned here t h a t E i e s b e c k , i n h i s system; kept the p r e d i c t i o n s unordered, to simulate p a r a l l e l i s m ; ray g o a l i n p r o v i d i n g the o r d e r i n g was t o achieve an e x t r a degree of c o n t r o l , not otherwise a v a i l a b l e . source - i s the name of the module which spawned t h i s p a r t i c u l a r p r e d i c t i o n ; t h i s i n f o r m a t i o n can o f t e n be u s e f u l i n working with the p r e d i c t i o n . An A n a l y s i s system Computational P r e r e q u i s i t e s 72 c a n c e l - i s the - ' c a n c e l l a t i o n f a c t o r ' f o r t h i s p r e d i c t i o n , decided e i t h e r when the p r e d i c t i o n was spawned, or when the p r e d i c t i o n manager (q.v.) examined i t . The c a n c e l l a t i o n f a c t o r i n d i c a t e s when the p r e d i c t i o n should be de l e t e d ( i . e . d e a c t i v a t e d ) , and can have th r e e forms: i) NIL - t h i s p r e d i c t i o n w i l l never be c a n c e l l e d ; i i ) a GENSYMed atom i f t h i s p r e d i c t i o n i s s u c c e s s f u l , a l l p r e d i c t i o n s having the same value of CANCEL ( i n c l u d i n g o b v i o u s l y the p r e d i c t i o n i t s e l f ) are de l e t e d . •••>••••••••• example:'CAN 157 i i i ) an a r b i t r a r y LISP p r e d i c a t e i f t h i s p r e d i c t i o n i s s u c c e s s f u l , a l l p r e d i c t i o n s s a t i s f y i n g t h i s p r e d i c a t e ( i n c l u d i n g p o s s i b l y t h i s p r e d i c t i o n i t s e l f ) are del e t e d . example: (EQ (CAR (TEST PREDICTION)) »PG) w i l l r e s u l t i n d e l e t i n g a l l p r e d i c t i o n s f o r any s o r t of p r e p o s i t i o n a l group. o b i ? - i s the ' o b l i g a t o r y ' f l a g , set t o T or NIL t o i n d i c a t e whether or not t h i s p r e d i c t i o n must be f u l f i l l e d t o complete the c u r r e n t sentence. E s s e n t i a l l y , t h i s corresponds to the idea of o b l i g a t o r y and o p t i o n a l cases. The important p o i n t s are, o b v i o u s l y , t e s t and a c t i o n ; the others are l e s s s i g n i f i c a n t , and w i l l not be d e a l t with i n d e t a i l . The c o n t r o l s t r u c t u r e o f a p r e d i c t i o n systen i s simple, but powerful. The work i s done through the ordered l i s t of An A n a l y s i s system Computational P r e r e q u i s i t e s 73 p r e d i c t i o n s . At each step of the a n a l y s i s , the system passes through t h i s l i s t i n order. As soon as a p r e d i c t i o n i s found which succeeds ( i . e . has i t s t e s t s a t i s f i e d ) ; t h a t p r e d i c t i o n i s allowed t o execute i t s attached a c t i o n . Then, the c y c l e r e p e a t s , checking f o r other s a t i s f i e d p r e d i c t i o n s (note t h a t the a c t i o n s of the f i r s t p r e d i c t i o n may have changed some g l o b a l i n f o r m a t i o n , and caused l a t e r ones to succeed). When no f u r t h e r p r e d i c t i o n s can be s a t i s f i e d , the next word of the i n p u t i s accepted, and the process s t a r t s over. Perhaps more should be s a i d about the nature of p r e d i c t i o n s themselves. At the sentence l e v e l , I see p r e d i c t i o n s being spawned by i ) the words i n a sentence ( i . e . the c u r r e n t i n p u t ) ; i i ) the c o n c e p t u a l s t r u c t u r e b u i l t so f a r . The p r e d i c t i o n s can be about t h r e e d i f f e r e n t l e v e l s ( f o r the moment, at l e a s t ) : i) words t o be looked f o r i n the input i i ) SSS£12t§ ( i . e . a c t i o n s , noun groups) that might be •  r e f e r r e d ' to- • i i i ) c o n c e p t u a l i z a t i o n s ( i . e . f u l l sentences) t h a t might be seen y e t . (note t h a t at the concept or c o n c e p t u a l i z a t i o n l e v e l , a p r e d i c t i o n i s e s s e n t i a l l y o p e r a t i n g on the output of other p r e d i c t i o n s , i n a manner s i m i l a r to production r u l e s . ) The q u e s t i o n o f l e v e l s i s by no means solve d ; i t w i l l be f u r t h e r d i s c u s s e d l a t e r . An A n a l y s i s system Computational Prerequisites 74 I t i s important to r e a l i z e that judicious use of t h i s sort of mechanism permits extremely powerful manipulation of the analysis process. For example, there i s a function, EXPECT, which takes a l i s t of components and builds the appropriate predictions for them. EXPECT takes three basic sorts of arguments:• i) AND - sequence -causes a set of predictions to be set up to handle a l i n e a r seguence of the form indicated, example: (EXPECT (AND (NG) (VG) (PG))) would b u i l d a series of predictions i n turn to handle, successively, a noun group, a verb group, and a preposition group ( i . e . , a simple declarative sentence). The equivalent form of an AIN would be: Figure 6 (a) - An ATN f o r a Linear Seguence i i ) XOR - exclusive or -causes a set of predictions to be set up, so that i f one succeeds, the others are removed example: (EXPECT (XOR (NG ANIMATE) (PG TO ANIMATE))) would set up two expectations, one looking f o r an animate noun group, the other looking f o r a preposition group beginning with »to» and having an animate object of the preposition (this i s obviously looking f o r tne An Analysis system Computational Prerequisites 75 i n d i r e c t object of a verb l i k e ' g i v e 1 ) . In ATN formalism, t h i s would be: Figure 6 (b) — An ATN for an Exclusive Or (with the appropriate r e s t r i c t i o n s on the aces). Note, however^ that i n an ATN, these arcs would a c t u a l l y be on d i f f e r e n t nodes, and a c e r t a i n amount of r e g u l a r i t y would be hidden, i i i ) OR - i n c l u s i v e or -causes a set of predictions to be spawned, looking f o r any member of the set, but making no e f f o r t to k i l l off other predictions i n the set. example-: (EXPECT (PG LOCATIVE) (PG TEMPORAL)) would e s t a b l i s h predictions looking for either a temporal or l o c a t i v e preposition group, or. both (e s s e n t i a l l y a s e t t i n g ) . This could not be represented e a s i l y i n an ATN; the c l o s e s t one could come i s : Figure 6 (c) - An ATN f o r an Inclusive 3r but there would have to be complex use of f l a g s , i f the number of choices was large. Thus, a model using the prediction mechanism f o r i t s basic An Analysis system Computational P r e r e q u i s i t e s 76 c o n t r o l s t r u c t u r e i s a t l e a s t ' as powerful as the oth e r system, and has some a d d i t i o n a l e x p r e s s i v e c a p a b i l i t i e s . One very i n s t r u c t i v e e x e r c i s e i s t o t r y to model the va r i o u s flows of c o n t r o l allowed by a p r e d i c t i o n system, i n a pattern-matching language l i k e SNOBOL. J? h i s p r o v i d e s i n t e r e s t i n g proof o f the power. At t h i s p o i n t I would l i k e t o e x p l a i n the ge n e r a l o r i e n t a t i o n o f the system. E s s e n t i a l l y , I'm proceeding under the assumption t h a t , i n comprehending n a t u r a l language, humans r a r e l y back up; r a t h e r , at each stage they are e x p e c t i n g the next i n p u t , and know unambiguously what t o do with i t . T h i s point was i m p l i c i t i n Riesbeck's work; i t has been r e i f i e d by Marcus (1975, p. 11): 'the s t r u c t u r e of n a t u r a l language provides enough and the r i g h t i n f o r m a t i o n to determine e x a c t l y what t o do next at each p o i n t of a parse' T h i s w i l l be r e f e r r e d t o as the ' s u f f i c i e n t - i n f o r m a t i o n assumption'. My co n t e n t i o n i s t h a t t h i s i s a v a l i d model of n a t u r a l language comprehension. To - implement the s u f f i c i e n t s - i n f o r m a t i o n assumption, we must r e c o g n i z e t h r e e p o i n t s : i ) i n some cas e s , a l i m i t e d amount of lookahead w i l l be r e q u i r e d ; i i ) s i n c e ' d e c i s i o n s are postponed as lo n g as p o s s i b l e , t h e r e may be a l a r g e number of c h o i c e s a c t i v e i n the system at any given time; i i i ) t h e r e i s o b v i o u s l y a great d e a l of i n f o r m a t i o n An A n a l y s i s system Computational P r e r e q u i s i t e s 77 a v a i l a b l e 'from the s t r u c t u r e o f n a t u r a l language, and from the surrounding context. The f i r s t of these i s o b v i o u s l y a r e f l e c t i o n of M i l l a r ' s work on chunking. " I f we r e s t r i c t the amount of lookahead s u f f i c i e n t l y w e l l , our system w i l l w i t h i n h i s g u i d e l i n e s , s a t i s f y i n g p s y c h o l o g i c a l c r i t e r i o n ( i ) , concerning chunking (pp. 46-47). The second a s s e r t i o n , about the number of c h o i c e s , w i l l be recognized' as an i n s t a n c e o f B a r l e y ' s (1970) p a r s i n g a l g o r i t h m (with r e s p e c t t o which Marcus and Riesbeck have been somewhat l a x i n g i v i n g c r e d i t ) , as implemented by Fowler (1976). The e s s e n t i a l p o i n t here i s that d e c i s i o n s are postponed as long as p o s s i b l e ; hence backup i s r a r e l y r e q u i r e d . Opon c o n s i d e r a t i o n of t h i s p o i n t , i t seems i n t u i t i v e l y wrong (I say ' i n t u i t i v e l y ' because I have no p s y c h o l o g i c a l support i n t h i s r e g a r d ) . In p a r s i n g , humans do not seem to c a r r y a l a r g e number of p o t e n t i a l parses forward, e l i m i n a t i n g the i n c o r r e c t ones o n l y as they are proved wrong. Nor, however, do they back up f r e q u e n t l y . Rather, they seem to f a s t e n very q u i c k l y upon a p a r t i c u l a r c h o i c e , which i s almost i n v a r i a b l y the r i g h t one. -The answer to t h i s problem, i t seems, l i e s i n the t h i r d statement, co n c e r n i n g the ' s u f f i c i e n t i n f o r m a t i o n ' which i s p r o v i d e d by the c o n t e n t of n a t u r a l language. I t appears t h a t there i s enough" i n f o r m a t i o n a v a i l a b l e t o enable us to p r e d i c t with reasonable accuracy (as i n Shannon's 'minimum entropy' d i s c o v e r y ) ; the problem l i e s i n r e c o g n i z i n g i t . My g o a l . An A n a l y s i s system Computational Prerequisites 78 therefore, has been to try to characterize the kinds of information available from natural language and use them i n analysis. The i d e n t i f i c a t i o n of information sources mentioned in Chapter I I , was a f i r s t step i n t h i s d i r e c t i o n . A related feature, which arose i n the discussion on psychology, concerns the di r e c t i o n of the analysis. Various psychological experiments, and the fact that ws can comprehend spoken input, indicate that we do not wait for the end of a sentence before beginning processinq; rather,"we are processing the input as i t comes in. Following t h i s guideline, I intend that analysis should be: 'always-forward' i . e . that the analyzer should have to look back at previous'input as seldom as possible. 'Instead, i t should check the information i t has abstracted from the previous input, and which i s currently providing guidance for the analysis process. Perhaps an example would serve here. Host case^-based systems extract the verb, as the c e n t r a l component of the sentence, then proceed to categorize the r e s t . More s p e c i f i c a l l y , i f a verb l i k e 'grow' were found, the system would form a set of -constraints on the subject (e.q. (HOST-BE ANIMATE)), then check back to see i f these can be s a t i s i f i e d . In my system, the subject (say, 'John') would be found f i r s t , and t h i s would put constraints on the verb (e.g. (MUST-BE SUITABLE-FOB' HUMAN)") '- which would then r e s t r i c t the class of verbs to be found ( i . e . , a verb l i k e 'exploded' would not be accepted). An Analysis system Computational P r e r e q u i s i t e s 79 T h i s i s a t r i v i a l example, used to emphasize what I co n s i d e r ' t o ' b e the-fundamental point of n a t u r a l language: a l l the i n f o r m a t i o n a v a i l a b l e at any po i n t i n the a n a l y s i s process must be used t o r e s t r i c t the c h o i c e s at l a t e r s t e p s . That i s , at any point i n a parse, we must have e x t r a c t e d the maximum i n f o r m a t i o n p o s s i b l e from the p r e v i o u s context, and t h i s must r e s t r i c t our present c h o i c e s . The idea o f e x t r a c t i n g the maximum i n f o r m a t i o n i s not new; what i s c r u c i a l here i s my c o n t e n t i o n t h a t there i s enough i n f o r m a t i o n a v a i l a b l e to g r e a t l y reduce our p o s s i b l e c h o i c e s . T h i s goes back t o Marcus* work and the s u f f i c i e n t - i n f o r m a t i o n assumption. In t h i s r e s p e c t my work d i f f e r s from t h a t of Riesbeck, who made a s l i g h t l y weaker a s s e r t i o n concerning the 'always-forward* d i r e c t i o n . An* example here w i l l serve to i n d i c a t e the p o t e n t i a l b e n e f i t s of t h i s approach: I w i l l compare the performance of my model t o another, s i m i l a r , system: T a y l o r * s ( T a y l o r and Rosenberg, 1975) c a s e - d r i v e n p a r s e r . . T a y l o r • s ' system was s e m a n t i c a l l y o r i e n t e d , and based on an extended taxonomy of cases. H is parser operated by f i n d i n g the main verb of a sentence (and sub-verbs, i f any), o b t a i n i n g the l i s t o f case s l o t s f o r th a t verb, then moving around the sentence s e l e c t i n g components t o f i l l these s l o t s . On c o r r e c t sentences, oar systems behave i n a somewhat s i m i l a r manner; the b a s i c d i f f e r e n c e s are not n o t i c e a b l e . The v a r i a t i o n s i n approaches can more e a s i l y be seen by observing t h e i r b ehavior on i n c o r r e c t sentences. An A n a l y s i s system Computational P r e r e q u i s i t e s 80 Take an example where a case i s missing: John put the b a l l . Here, the m i s s i n q c o m p o n e n t i s obv i o u s l y a l o c a t i o n s p e c i f i c a t i o n . T a y l o r ' s system and mine would d i s c o v e r the anomaly at the same time: when the end of the sentence was reached. In h i s s y s t e m , the e r r o r would be manifest as an u n f i l l e d case s l o t ( w h i c h was marked as being o b l i g a t o r y ) ; i n min e , ' i t would'show up as an u n s a t i s f i e d p r e d i c t i o n ( a l s o marked as o b l i g a t o r y ) . Thus, on t h i s sentence, the r e s u l t s would be s i m i l a r . Now, however, look a t an example with an e x t r a component: John came the book i n the morning. In t h i s sentence, 'the book' i s anomalous, and should be fla g g e d as such by the a n a l y z e r . In my system, the anomaly would be d i s c o v e r e d as soon as •the book' was encountered. That i s , t h e r e would be a number of a c t i v e p r e d i c t i o n s (many spawned by 'came'), a l l o f which would r e j e c t 'the book' as a component. T h i s would-cause an immediate e r r o r (and p o s s i b l y , although not l i k e l y , backup) . In T a y l o r ' s a n a l y z e r , the d i f f i c u l t y would be n o t i c e d only when a l l the- r e s t o f the sentence had been processed. The system would have f i l l e d a l l the s l o t s f o r 'came' (in t h i s case, o n l y a s u b j e c t and a temporal p h r a s e ) , when i t found t h a t one component* w a s " s t i l l not accounted f o r . I t would t r y t o f i t t h i s component i n t o a number o f d e f a u l t s l o t s , and only then s i g n a l an ' e r r o r . •••••-••>•-••••••'••••••• ... I f we i n t r o s p e c t f o r a moment, i t seems obvious t h a t people An A n a l y s i s system Computational P r e r e q u i s i t e s 81 do not use a complete top-down, s l o t - f i l l i n g approach. Rather, they absorb the sentence one p i e c e at a time, f o r c i n g each piece to f i t i n t o some place i n the p r o g r e s s i v e l y - d e v e l o p i n g p r o s p e c t i v e meaning s t r u c t u r e . Thus, i n a sentence l i k e the p r e v i o u s example, they would d i s c o v e r an e r r o r as soon as the anomalous component was heard, i n the same manner as my system d i d . • • • • • • • > T h i s minor example was but a v a r i a t i o n on a theme -- the theme o f 'always-forward' p a r s i n q . B y : f o l l o w i n g the ' s u f f i c i e n t - i n f o r m a t i o n assumption', and e x t r a c t i n g the maximum amount of i n f o r m a t i o n from the p r e v i o u s c o n t e x t , we should be able to keep the number of p r e d i c t i o n s a c t i v e at any one time down t o a minimum. O b v i o u s l y , much t e s t i n g i s needed here, t o provide an e m p i r i c a l e s timate o f the number. To conclude t h i s d e s c r i p t i o n of the b a s i c a n a l y s i s mechanism, I w i l l present a sample parse f o r a simple sentence. The sentence i s 'John was coming home from s c h o o l . ' , and i t would be analyzed i n the f o l l o w i n g manner ( s i m i l a r to Riesbeck, 1975, pp. 89-90) : An A n a l y s i s system Computational P r e r e q u i s i t e s 82 word found p r e d i c t i o n s a c t i v e p r e d i c t i o n s a c t i o n s u c c e s s f u l taken John was coming home from s c h o o l 1 1 2 2 3 4 5 5 5 (NP) (NP) (VG) (VG) (ADV IOC) (PG TO) (PG FROM) (PG FROM) (PG FROM) (RUMINATE ACTOR) (SETQ rCONCEPT: •((ACTOR actor) (ACTION PTRANS) (OBJECT actor) (DIR (TO ?) (FROM ?) ) ) (PREDICT #3) (PREDICT #4) (PREDICT #5) (RUMINATE (DIR TO) ) (KILL #4) (RUMINATE (DIR FROM) ) Figure 1 (a) - Flow of C o n t r o l i n Pa r s i n g a Sentence with the r e s u l t , i n Conceptual Dependency terms: \-> home John <=> PTRANS <-- John < — | j-< s c h o o l Fi g u r e 7 (b) - R e s u l t i n g Parse Note t h a t I have not bothered t o show the tense i n the diagram. Tense i s e a s i l y recognized' (both Woods and Winograd had s o p h i s t i c a t e d systems f o r d i s c o v e r i n g tense) and rep r e s e n t e d (Schank (1973, p. 206) presented a complete taxonomy of tense, f o r Conceptual Dependency), but much harder to usa» In the simple s t o r i e s with which we s h a l l be concerned, tense w i l l r a r e l y be a f a c t o r ; thus the tense markers w i l l be omitted most An A n a l y s i s system Computational Prerequisites 83 of the time. An Analysis system 84 V. A Model f o r Discourse A n a l y s i s 7.1 Extending the P r e d i c t i o n system 7.1.1 The B a s i c C o n t r o l S t r u c t u r e V.1.2 I n t e r a c t i o n o f P r e d i c t i o n s V.1.3 L e v e l of P r e d i c t i o n s V.1.4 Comparison with"Production Systems T.2 Use of Information Sources 7.2.1 P r e l i m i n a r y Requirements V.2.2 Information Sources V.3 A D e t a i l e d Example 7.4 The Implementation Next, I s h a l l r e t u r n to the c e n t r a l goal o f t h i s t h e s i s d i s c o u r s e a n a l y s i s . In t h i s chapter, I hope to u n i f y what has been s a i d p r e v i o u s l y , and b u i l d upon the work, of the l a s t chapter, t o present a coherent model f o r the a n a l y s i s o f d i s c o u r s e . I should s t a t e at the o u t s e t t h a t many o f the f e a t u r e s here are by no means f u l l y s p e c i f i e d , even i n my own mind; many of the i d e a s are at best inchoate n o t i o n s about p o s s i b l e f u t u r e d i r e c t i o n s . The i n t e n t here i s t o g i v e a good i n d i c a t i o n of the problems i n v o l v e d i n d i s c o u r s e a n a l y s i s . To t h a t end, I have presented the v a r i o u s stages of development, and the r a t i o n a l e behind each d e c i s i o n taken. The chapter c o n s i s t s of f o u r s e c t i o n s . In the f i r s t , the p r e d i c t i o n system i s re-examined, t h i s time with the i n t e n t o f extending i t t o handle connected A Hodel f o r Discoursa A n a l y s i s 85 d i s c o u r s e . V a r i o u s problems are d i s c u s s e d , i n c l u d i n g d i f f i c u l t i e s with top^-down/bottom-up i n t e r a c t i o n , c o n t r o l of i n t e r a c t i o n of p r e d i c t i o n s , the question of l e v e l s of p r e d i c t i o n s , some b a s i d bookkeeping work, and other more s u b t l e f e a t u r e s . > • Next, I r e t u r n to the s e t o f Information Sourcss presented i n chapter I I . T h i s time, I t r y t o show how they f i t i n t o a d i s c o u r s e a n a l y s i s system, and more i m p o r t a n t l y , what happens without 1 them. T h i r d , a extended example i s presented, d e t a i l i n g the s t e p s taken a t each stage of the a n a l y s i s p r o c e s s , and showing how the whole system f i t s t ogether. F i n a l l y , the implementation i t s e l f i s d i s c u s s e d . As mentioned, t h i s i s at an incomplete stage, and much remains to be done. I t r y t o i n d i c a t e what have been the most e n l i g h t e n i n g aspects of the implementation process, and where I f e e l the most d i f f i c u l t work remains. V.1 Extending the P r e d i c t i o n System There was a p r e l i m i n a r y e f f o r t by Riesbeck (1974, Pt. II) to extend h i s system f o r the purpose of d i s c o u r s a a n a l y s i s . T h i s i n v o l v e d a somewhat vague notion of context, which i n c l u d e d what I have c a l l e d 'context' (foregrounding and frames) and 'real-world-knowledge' ( i n f e r e n c i n g ) , and a b r i e f mention of s t y l e ; However, the treatment was at best s u p e r f i c i a l and ad hoc, and no implementation was provided. Extending the p r e d i c t i o n system A Model for Discourse Analysis 86 I would l i k e to outline a possible extension to the predic t i o n system, keeping i n mind the various information sources that we have i d e n t i f i e d . V.1.1 The Basic Control Structure The I.S.'s a v a i l a b l e were found to have three c h a r a c t e r i s t i c s : i ) weakness; i i ) p o t e n t i a l lack of information; i i i ) gratuitousness. To handle t h i s e f f e c t i v e l y , a modular system i s needed. One possible structure i s : Figure 8 - Control Structure of the System i . e . , a number of independent modules, interacting through a c e n t r a l control system. Extending the Prediction System A Model f o r Discourse a n a l y s i s 87 In l i g h t of the c h a r a c t e r i s t i c s of p r e d i c t i o n - b a s e d systems, perhaps' t h i s can be f o r m a l i z e d . I t seams t h a t the c e n t r a l u n i t = (here l a b e l l e d ' a n a l y s i s ) should be an autonomous system, able to perform a n a l y s i s by i t s e l f . This corresponds to the p r e d i c t i o n model o u t l i n e d i n the previous s e c t i o n — i . e . the system a c t u a l l y implemented by Riesbeck. T h i s module w i l l o perate, as i n the previous example, through the use of a p r e d i c t i o n l i s t . The o n l y d i f f e r e n c e i s that i n t h i s case the p r e d i c t i o n l i s t i s g l o b a l — i ; e . a c c e s s i b l e to other modules. The s i x p e r i p h e r a l p i e c e s would each be s e p a r a t e modules, abl e to g i v e i n f o r m a t i o n to the system on l y through the g l o b a l p r e d i c t i o n l i s t (compare t h i s t o Hearsay's 'blackboard' (Reddy and Newell, 1974; Reddy, Erman, and Neely, 1973)). Thus, these modules would remain s i l e n t u n t i l a s p e c i f i c form was recognized i n the i n p u t . at t h i s time, they would c o n t r i b u t e t h e i r knowledge by adding p r e d i c t i o n s to the l i s t (hence h e l p i n g to c o n t r o l the s t r u c t u r e - b u i l d i n g , as w e l l ) . Under t h i s scheme, some modules would be more a c t i v e than others; f o r example, a t e x t grammar might be i n t e r a c t i n g c o n t i n u a l l y , whereas a ' c o l l a t e r a l * module might run only once i n a while. V.1.2 I n t e r a c t i o n o f P r e d i c t i o n s ...... Examination of t h i s system, however, exposes a c o n c e p t u a l flaw. P r e d i c t i o n s cannot be allowed to be completely o b l i v i o u s of the o t h e r * p r e d i c t i o n s i n the l i s t (as i s now the c a s e ) , although the n o t i o n i s tempting. I f t h i s were so, we would e v e n t u a l l y reach the stage of having s e v e r a l p r e d i c t i o n s Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 88 competing f o r the same i n p u t , each p r e d i c t i o n with a d i f f e r e n t a c t i o n i t wishes t o perform. For example, i f we were f o l l o w i n g the s t o r y : John and Hary were p l a y i n g b a s e b a l l . John was angry at Mary. John ... At t h i s p o i n t , t h r e e o f the modules would be a p p r o p r i a t e : s t a g i n g (theme): would say t h a t we are d i s c u s s i n g John and Mary, hence t h a t 'John' i s a l l r i g h t as the s u b j e c t of the sentence; r e a l - w o r l d ( i n f e r e n c e s ) : would have seen that John was angry at Mary, and i n f e r r e d ( i n t e r a l i a ) t h a t John i s l i a b l e to do something t o Mary; context (frames): would s t i l l be i n the b a s e b a l l frame, and would t h e r e f o r e be look f o r some a c t i o n from John r e l e v a n t t o b a s e b a l l . Thus, 'John' can be i n t e r p r e t e d i n two ways, depending on whether the i n f e r e n c e module or the frame module i s q i v e n more a u t h o r i t y . Of course, the d e c i s i o n c o u l d be postponed u n t i l the next component i s found (and probably would be, i n t h i s c a s e ) , but the u n d e r l y i n g problem of competing p r e d i c t i o n s s t i l l e x i s t s . The example i s a t r i v i a l m a n i f e s t a t i o n of a much deeper problem. The flaw i s even more g l a r i n g i n the case of two c o i n c i d i n g p r e d i c t i o n s . Suppose, f o r i n s t a n c e , t h a t the i n f e r e n c e module'and the frame module are both e x p e c t i n g the same i n p u t ; and i n t e n d t o perform the same a c t i o n s with i t . The two p r e d i c t i o n s should i n t e r a c t , and, i n some sense, r e i n f o r c e Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 89 each other; r a t h e r than t r y i n g to handle the s i t u a t i o n independently. T h i s i s - a common A.I. problem: c o n t r o l l i n g the i n t e r a c t i o n of v a r i o u s sources (see, f o r example, Reddy and Newell (1974), Lenat (1975), Paxton and Robinson (1975), or Erman and Le s s e r (1975)). He want the I.S.'s to remain independent and l o c a l , and not r e q u i r e them t o be aware o f each o t h e r T (or indeed of the r e s t of the system). At the same time, we must c o n t r o l t h e i r i n t e r a c t i o n i n some way, f o r the reasons j u s t mentioned. T h e ' s o l u t i o n , i t seems, l i e s i n the uniform nature of the p r e d i c t i o n s ; - A l l work i n the system i s done through the g l o b a l p r e d i c t i o n l i s t , and a l l elements of t h i s l i s t have the same form. T h e r e f o r e ; I propose to c r e a t e a new, autonomous, module c a l l e d the p r e d i c t i o n MSMSE* The p r e d i c t i o n manager w i l l scan the- p r e d i c t i o n l i s t at each i t e r a t i o n , l o o k i n g f o r c o o p e r a t i n q and competing p r e d i c t i o n s , and other p o t e n t i a l sources of t r o u b l e . I t w i l l - have the power to make changes to the p r e d i c t i o n l i s t , by adding new p r e d i c t i o n s and d e l e t i n g or modifying c u r r e n t ' o n e s . The p r e d i c t i o n manager, as d e s c r i b e d here, w i l l o b v i o u s l y be a l a r g e module, with a l o t of i n f o r m a t i o n and a c o r r e s p o n d i n g amount of power; I t w i l l need t o know about a l l types of p r e d i c t i o n s , about t h e i r a s s o c i a t e d t e s t s and a c t i o n s , and about the types of i n t e r a c t i o n s which can occur between p r e d i c t i o n s . Much of t h i s i n f o r m a t i o n can o n l y be gathered e m p i r i c a l l y , by running the system- and o b s e r v i n g the kinds of d e c i s i o n s demanded of the p r e d i c t i o n manager. However, a couple of t a s k s Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 90 can be i d e n t i f i e d , based only on t h e o r e t i c a l requirements. Given two c o o p e r a t i n g p r e d i c t i o n s : the p r e d i c t i o n manager should check to see t h a t they r e a l l y are l o o k i n g f o r the same component, and f o r the same g e n e r a l reason. Then, i t should l e r g e the two p r e d i c t i o n s to produce a new one with the same t e s t , a higher s t r e n g t h , and a combined l i s t o f a c t i o n s , example: i f the theme module had posted a p r e d i c t i o n (note t h a t i r r e l e v a n t f i e l d s are o m i t t e d ) : ((NG) t S E T Q : THEME: : COMPONENT:) -7} ( i . e . f i n d a noun group, and a s s e r t that i t i s the theme), and the i n f e r e n c e r had posted: " ( (NG)' (RUMINATE ACTOR) 7) ( i . e . f i n d a noun group and assume t h a t i t i s the s u b j e c t of the c o n c e p t u a l i z a t i o n ) , the p r e d i c t i o n manager should merge these, and produce: ((NG) (PROG (RUMINATE ACTOR) (SETQ : THEME: : COMPONENT:) ) 9) ( i . e . a s t r o n g e r p r e d i c t i o n , but l o o k i n g f o r the same t h i n g , and performing the a c t i o n s that would have been done by both the previous p r e d i c t i o n s ) . T h i s r e f l e c t s the i n t u i t i v e n o t i o n t h a t i f two (or more) separate sources are expecting the same t h i n g , we should favour t h a t e x p e c t a t i o n . Given two competing p r e d i c t i o n s : the p r e d i c t i o n manager should attempt to d e f e r a d e c i s i o n ( i . e . keep both p r e d i c t i o n s a c t i v e ) ' u n t i l l a t e r i n f o r m a t i o n a r i s e s . I f that i s not p o s s i b l e ; i t ; w i l l probably have t o chose the one with the higher s t r e n g t h (remember, these s t r e n g t h s were assigned s u b j e c t i v e l y Extending the P r e d i c t i o n System A He-del f o r Discourse a n a l y s i s 91 by the v a r i o u s c r e a t i n g modules), and k i l l the o t h e r , making a note of t h i s f a c t , i n case backup s o u l d be r e q u i r e d , example: i f , say, the i n f e r e n c e r (a marvelously e r r a t i c and i n c o n s i s t e n t creature) had c r e a t e d two p r e d i c t i o n s : ( (NG) * (HOHINaTE aCTOS) 7) ( i . e . i f a noun group i s found; make i t the a c t o r ) , ((NG) (RUHINaTE OBJ) 5) ( i . e . i f a noun group i s found, make i t the o b j e c t ) , the p r e d i c t i o n manager should f i r s t check to see i f a noun group has been found; i f i t hasn't, no d e c i s i o n need be made* I f i t has, the p r e d i c t i o n manager should l e t the f i r s t p r e d i c t i o n succeed, ( i . e . l e t the noun group be recorded as the a c t o r ) , while d e l e t i n g the second one from the p r e d i c t i o n l i s t , and making a note to t h a t e f f e c t . Obviously, more s u b t l e kinds o f i n t e r a c t i o n s w i l l be t a k i n g p l a c e , and the p r e d i c t i o n manager w i l l have to be much more s o p h i s t i c a t e d . The scheme presented here i s j u s t an o u t l i n e . N o t e 1 t h a t t h i s s o r t of problem never arose i n Riesbeck's system; because o f i t s s i m p l i c i t y . He s t a t e d (Riesbeck, 1975, p. 103) t h a t , by f i a t , t h ere would be no c o i n c i d i n g or competing p r e d i c t i o n s . T h i s was q u i t e reasonable, s i n c e h i s system only had one source o f p r e d i c t i o n s — the s e n t e n c e - l e v e l syntax/semantics ones spawned by the various words i n the sentence. Thus, t h e r e was never any danger of unforeseen i n t e r a c t i o n s . In the'system d e s c r i b e d here, t h i n g s a r e n ' t so smooth. The c e n t r a l a n a l y s i s system ( i . e . the p a r t that Riesbeck included) Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 92 i s but one or s e v e r a l i n t e r a c t i n g sources. True, the c e n t r a l module i s the s t r o n g e s t , but a l l work must be done through the p r e d i c t i o n l i s t , and any module can w r i t e i n t o t h a t i f i t f e e l s so i n c l i n e d . One module (the i n f e r e n c e r ) can even cre a t e c o n t r a d i c t o r y p r e d i c t i o n s by i t s e l f ; the i n t e r a c t i o n between s e v e r a l modules i s t h e r e f o r e complex. T h i s i s a l s o the strong p o i n t of the system. Since the method of communication i s so uniform (and i s the same one used by the c e n t r a l module f o r i t s own i n t e r n a l work), the p e r i p h e r a l modules can i n t e r a c t with the c e n t r a l s e c t i o n i n a n a t u r a l way. The c o n t r o l s t r u c t u r e of the system, with the p r e d i c t i o n manager added, would probably now look l i k e : F i g u r e 9 - M o d i f i e d C o n t r o l S t r u c t u r e That i s , the a c t u a l e x e c u t i v e p a r t of the system, which runs the p r e d i c t i o n s and performs the a p p r o p r i a t e a c t i o n s , has been removed from the ' a n a l y s i s * module, and given i t s own autonomous l o c a t i o n . A l l t h e other modules communicate with i t v i a the Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 93 p r e d i c t i o n manager — i . e . , through the g l o b a l p r e d i c t i o n l i s t . Note t h a t ' the ' a n a l y s i s ' module ( i . e . the autonomous s e n t e n c e - l e v e l analyzer) i s now the same as a l l the o t h e r s ; the on l y d i f f e r e n c e l i e s i n the f a c t t h a t i t i s s t r o n g e r and more a c t i v e . T h i s concludes the study o f i n t e r a c t i o n s . S u f f i c e i t to say t h a t the p r e d i c t i o n manager i s an i l l - s p e c i f i e d p i e c e of the system, which probably won't s t a b i l i z e u n t i l a f t e r s e v e r a l i t e r a t i o n s of implementation. V . 1 .3 L e v e l s of P r e d i c t i o n s Once the i n t e r a c t i o n of p r e d i c t i o n s i s taken care o f , some more b a s i c q u e s t i o n s must be asked about the p r e d i c t i o n s themselves. The problem i n t h i s case concerns the l e v e l s of a p r e d i c t i o n — i . e . how w e l l - s p e c i f i e d should the t e s t part of a p r e d i c t i o n be? S e v e r a l l e v e l s can be i d e n t i f i e d , as mentioned p r e v i o u s l y : i ) words - - e . g . 'give','John'; i i ) concepts -* e.g. 'noun group','prep^group t o ' ; i i i ) c o n c e p t u a l i z a t i o n s — e.g. 'a h i t event','John doing something t o Mary'. I d e a l l y , we would l i k e t o s p e c i f y t h i n g s a t the word l e v e l i . e . , t o p r e d i c t s u f f i c i e n t l y w e l l from the context t h a t we can guess the next word. U n f o r t u n a t e l y , such i s seldom the case. As mentioned i n the s e c t i o n on t e x t grammars (although i t i s t r u e o f a l l I . S . ' s ) , e x p e c t a t i o n s g e n e r a l l y come at a much higher l e v e l — Extending the P r e d i c t i o n System A Model for Discourse Analysis 94 usually at the l e v e l of a conceptualization. Ona example, from the section on text grammars, i s 'expect a s e t t i n g * , where i t was shown that •setting* could be manifest in several ways at a more d e t a i l e d l e v e l . E s s e n t i a l l y , what we were doing at that point was to convert reguests from the conceptual l e v e l to the language l e v e l . I t i s tempting to think that t h i s can be done in general; that, f o r any high-level component, we merely specify a l l of i t s surface manifestations. This f a i l s , however, bacause the s i t u a t i o n i s explosive. In r e a l i t y , the predictions form a fan, in moving from higher l e v e l s to more s p e c i f i c forms. . i . conceptualization concept - word Thus, the prospect of 'pushing predictions forward' to the word l e v e l i s just not v i a b l e . An a l t e r n a t i v e approach would be to leave the predictions at a high l e v e l (say, the l e v e l of conceptualizations) and l e t * the analysis proceed bottom-up u n t i l i t reaches that l e v e l . For example, i n a text grammar* we might have an expectation 'look for an event• (where an event i s a s p e c i f i c type of a c t i o n ) . He could then l e t the analysis proceed bottom-up ( i . e . using only l o c a l , sentence-level information), u n t i l a conceptualization was b u i l t . Ihen, t h i s conceptualization i s checked to see whether i t i s , i n fact, an event. Unfortunately, t h i s approach i s also not f e a s i b l e . The Extending the Prediction System A Model for Discourse Analysis 95 problem i s obvious — afte r c a r e f u l l y extracting the maximum amount of information from the surrounding discourse, we ignore t h i s information, f a i l i n g to use i t i n parsing a sentence, and make use of i t only once the sentence has been completely parsed. This seems coun t e r i n t u i t i v e , and also reguires the a b i l i t y to back up — something we wanted to avoid as much as possible For once our i n t u i t i o n i s r i g h t . I f the analyzer i s turned loose •bottom-up* on a sentence, i t i s subject to a l l of the contradictions and ambiguities which plague natural language at the l o c a l l e v e l . Again, explosion occurs; t h i s time, i n the opposite d i r e c t i o n . conceptualization ^concept word Perhaps the degree of bushiness i s exaggerated. The point s t i l l remains: f a i l u r e to make use of discourse l e v e l constraints when processing s i n g l e sentences i n v a l i d a t e s our method. Humans do not perform bottom-up analysis of sentences; why should our system? The solution, i t seems, i s to pick a middle ground — a l e v e l at which neither tree i s too large. As the diagrams have indi c a t e d , the l e v e l I consider to be most e f f e c t i v e , and which I have used i n implementing the system, i s the co,nceB£. This i s e s s e n t i a l l y one element of a conceptualization (e.g. actor, action, object, e t c . ) , and can be manifest in saveral ways (e.g. noun group, verb group, preposition group, adverb, e t c . ) . Extending the Prediction System A Model f o r Discourse A n a l y s i s 96 As i n d i c a t e d , p r o c e s s i n g proceeds bottom-up ( i . e . without c o n s t r a i n t s from h i g h e r l e v e l s ) up to the concept l e v e l . To implement t h i s bottom-up work i n a c l e a n way, the b a s i c t r a n s i t i o n net (BTN) was used. TheBTN i s the same as Soods's ATN, except f o r two d i f f e r e n c e s : i ) i t i s not augmented; no use i s made of r e g i s t e r s and a r c t e s t s ; •" i i ) i t i s not r e c u r s i v e ; no PDSH or POP work i s allowed. These r e s t r i c t i o n s were made t o ensure t h a t the bottom-up work was l i m i t e d i n power, and t h a t no hidden p r o c e s s i n g was being done. T h i s approach i s not new. T a y l o r , i n h i s system, used a r e s t r i c t e d ATN to enable him to i d e n t i f y the components o f the sentence. Riesbeck, who claimed that no s y n t a c t i c work was being done, a c t u a l l y had a p r i m i t i v e noun phrase r e c o g n i z e r b u i l t i n t o h i s i n t e r p r e t e r . The'work done by the BTN i s not completely bottom-up; c o n s t r a i n t s ( i . e . Katz/Fodor type s e l e c t i o n a l r e s t r i c t i o n s ) can be passed down from the higher l e v e l whenever the BTN i s invoked. For example, to look f o r an animate noun group, we might use the form: " (PARSE NG ANIMATE) (where PARSE i s the t o p - l e v e l c a l l t o the BTN). T h i s would cause the f e a t u r e •animate* to be passed down, and i t would be checked a g a i n s t the noun-group component b u i l t , j u s t b e f o r e the Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 97 POP was executed. T h i s d e c i s i o n to a r b i t r a r i l y p o s i t the concept l e v e l as the p o i n t of i n t e r s e c t i o n between top-down and bottom-up i s s i m p l i s t i c , of course. I n r e a l i t y , t h i s s o r t of i n t e r a c t i o n i s going on at a l l times, on a l l l e v e l s (see Bobrow and Norman (1975),' Palmer (1975) , or Havens (1976), f o r i n t e r e s t i n g d i s c u s s i o n s on t h i s p o i n t ) . Given our d e c i s i o n to 'push* a l l p r e d i c t i o n s forward to at l e a s t the concept l e v e l , there i s s t i l l a s m a l l amount of c o m b i n a t o r i a l e x p l o s i o n to be d e a l t with. T h i s i s the p r i c e paid f o r the t r a n s i t i o n from the higher l e v e l s down t o the concept l e v e l . I t i s a l s o a fundamental c h a r a c t e r i s t i c of language — the number of s u r f a c e m a n i f e s t a t i o n s of a p a r t i c u l a r c o n s t r u c t i s o f t e n l a r g e . A l l of the work i s done by a l a r g e , complex, f u n c t i o n c a l l e d SURFACE," which takes an e x p e c t a t i o n at a h i g h e r l e v e l , and c o n v erts i t i n t o a s e r i e s of forms at the concept l e v e l , which can then be passed to the EXPECT f u n c t i o n , mentioned e a r l i e r . For example, a c a l l (SURFACE SENTENCE DECLARATIVE) would produce the l i n e a r seguence: (AND (NG) (VG) (XOR (PG) (ADJ) (NG) ) ) ( i . e . a somewhat s i m p l i s t i c s y n t a c t i c s t r u c t u r e f o r a sentence). Given (SURFACE RECIP) i t would produce the set of a p p r o p r i a t e formss f o r the Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 98 ' r e c i p i e n t ' : (XOR (PG TO) (NG ANIMATE) Thus; the o p e r a t i o n i s f a i r l y s t r a i g h t f o r w a r d . Problems a r i s e , however; when the arguments t o SURFACE come from higher l e v e l s . For i n s t a n c e (SURFACE EVENT) means t h a t a set of p r e d i c t i o n s c h a r a c t e r i z i n g an 'event' should be spawned. But i n doing t h i s , we run i n t o c o m b i n a t o r i a l e x p l o s i o n ; the d i s t a n c e between 'event' and the concept l e v e l i s too great.• • • • • What I am s a y i n g here i s t h a t I have no d e f i n i t e answer to the problem, j u s t an approach which works i n simple cases. An i n t e r e s t i n g i d e a here would be to use p r e d i c t i o n s t o convert from' h i g h e r " l e v e l s to lower ones. That i s , given the a b s t r a c t form 'event', t h e r e might be a p r e d i c t i o n which says •an event can be manifest as e i t h e r an a c t i o n or a response'. In t u r n , another p r e d i c t i o n might say 'an a c t i o n can be manifest by a b e n e f a c t i v e a c t or a d e s t r u c t i v e a c t ' , and so on. E s s e n t i a l l y , these t r e e s of p r e d i c t i o n s a r e performing the same work t h a t "SURFACE d i d , except t h a t they are now under the c o n t r o l and c o g n i t i o n of the system; thus the e x p l o s i v e nature can be c o n s t r a i n e d ; Again, t h i s aspect of the system i s remarkably s i m i l a r t o the production system method. T h i s approach, while not f u l l y developed, would seem to be the most promising. - • Riesbeck, i n h i s t h e s i s , o u t l i n e d a model of t h i s nature, but i t was r e s t r i c t e d to l i n e a r c h a i n s r a t h e r than t r e e s , and Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 99 was h e a v i l y dependent on some of the ' s l o t - a n d - f i l l e r ' a spects of conceptual dependency. At l e a s t i t ' s a s t a r t , though. V.-1.4 Comparison with "Production Systems I t i s i n t e r e s t i n g , at t h i s p o i n t , to re-examine the fundamental computational nature of p r e d i c t i o n - b a s e d systems. E s s e n t i a l l y , p r e d i c t i o n s c o n s i s t of a l e f t - h a n d - s i d e (TEST), and a r i g h t - h a n d - s i d e (ACTION). The c o n t r o l mechanism operates by c y c l i n g through the l i s t of p r e d i c t i o n s , s e a r c h i n g f o r one whose t e s t succeeds, and executing the attached a c t i o n . T h i s corresponds c l o s e l y t o the methodology of production systems. These were f i r s t proposed by Post (1943), as a g e n e r a l computational mechanism.• In i t s simplest form, a p r o d u c t i o n system (PS) c o n s i s t s of three components: a s e t of r u l e s , a data base, and an i n t e r p r e t e r . The r u l e s e t i s ordered, and the data base i s simply a c o l l e c t i o n of symbols. The i n t e r p r e t e r operates by s e a r c h i n g the r u l e s e t u n t i l one i s found whose l e f t - h a n d - s i d e (LHS) can be matched a g a i n s t the data base. When one i s found, the r i g h t - h a n d - s i d e (RHS) of the a p p r o p r i a t e r u l e i s i n s e r t e d - i n t o ' t h e data base i n p l a c e of the p a t t e r n which was matched ( i . e . , the LHS), and the c y c l e continues. In more recent work, the s t r u c t u r e of p r o d u c t i o n systems has been extended. The LHS can be an a r b i t r a r y form, which i s e v a l u a t e d , and the RHS i s an a c t i o n , which i s executed ( p o s s i b l y c a u s i n g s i d e e f f e c t s ) . The data base i s no l o n g e r a simple c o l l e c t i o n of symbols, but can now i n c l u d e such a s p e c t s as p e r c e p t i o n , although i t i s s t i l l completely g l o b a l . T h i s s t y l e Extending the P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 100 of PS w i l l be r e f e r r e d t o as pure production systems (Davis and King, 1975) . At t h i s p o i n t , production systems and p r e d i c t i o n - b a s e d systems would seem to be s i m i l a r (note that ' r u l e s * i n one system are ' p r e d i c t i o n s ' i n the o t h e r ) ; a comparison i s i n order. ' To provide t h i s we must f i r s t i d e n t i f y the c h a r a c t e r i s t i c s of 'pure' production systems, then v e r i f y whether they apply as w e l l t o our p r e d i c t i o n system. Pure PSs have the f o l l o w i n g f e a t u r e s (taken from Davis and King, 1975) : i) r e s t r i c t i o n s on i n t e r a c t i o n s between the r u l e s In'pure PSs, t h e - o n l y 1 i n t e r a c t i o n allowed between the v a r i o u s system modules i s v i a the g l o b a l data base. T h i s 'tends to preserve independence of knowledge sources, at the expense of e x p l i c i t n e s s i n the c o n t r o l s t r u c t u r e . i i ) c o n s t r a i n e d format of r u l e s The i r i s can be a Boolean combination of simple p r e d i c a t e s ; the RHS i s l i m i t e d to ' c o n c e p t u a l l y simple' o p e r a t i o n s . i i i ) m o d u l a r i t y T h i s i s a byproduct o f f e a t u r e ( i ) ; the c o n t r o l flow i s completely decoupled from the a c t u a l r u l e s . iv) o p a c i t y T h i s i s an unfortunate e f f e c t of the d e c o u p l i n g j u s t m e n t i o n e d t h e system i s d i f f i c u l t to understand. Extending the P r e d i c t i o n System A Model for Discourse Analysis 101 v) second-order understanding E s s e n t i a l l y , a pure PS, because of the r i g i d formalism into which* i t s rules have been-•forced, should be capable of 'introspecting', and examining i t s own rules. This i s a r e l a t i v e l y general description of a pure PS; i t i s obvious that the prediction system we have just outlined does not f i t completely into t h i s c l a s s . More important, however, i s the fact that most of the production systems actually implemented also d i f f e r in various ways from these guidelines. In discussing our prediction-based system, I s h a l l try to point out the p a r a l l e l s with current PSs. One of the major differences between our system and pure PSs i s the inte r a c t i o n of rules, I found i t necessary to create a prediction manager, with the a b i l i t y to examine and manipulate predict ions; to - handle possible interactions ( c o n f l i c t i n g and cooperating); Interestingly the DENDBAL system (Feigenbaum, 1971) also faced the problem of c o n f l i c t of rules; i t was dealt with, i n a manner sim i l a r t c my prediction manager, through a system of rule -precedence ( i ; e . , strength); In LISP70 (Tesler et a l , 1973); c o n f l i c t s were resolved by chosing the most s p e c i f i c r u l e . ' * Another difference concerns the constrained form of rules. In our system, the LHS * i s quite simple, as required (this i s not always the case; "ACT (Anderson, 1976) has a complex node-matching scheme, and DENDBAL permits not only complex Extending the Prediction System A Model f o r D i s c o u r s e A n a l y s i s 102 m a t c h e s , but a l s o s i d e e f f e c t s i n t h e m a t c h ) . The Has, however, i s more c o m p l e x ; t h e s t r u c t u r e - m a n i p u l a t i n g and p r e d i c t i o n - ' b u i l d i n g o p e r a t i o n s s p e c i f i e d c o u l d h a r d l y be c a l l e d ' c o n c e p t u a l l y s i m p l e * . A g a i n , t h e r e i s a p r e c e d e n t ; i n t h e PASII s y s t e m (Waterman, 1974), t h e RHS c a n s p e c i f y o p e r a t i o n s t o c o n s t r u c t new p r o d u c t i o n s . The •- q u e s t i o n o f s e c o n d - o r d e r u n d e r s t a n d i n g ( i . e . m e t a - p r e d i c t i o n s ) has n o t been t r e a t e d i n d e t a i l i n o u r s y s t e m , a l t h o u g h t h e p r e d i c t i o n manager o b v i o u s l y r e g u i r e s a c e r t a i n amount o f t h i s . The c h a r a c t e r i s t i c s o f m o d u l a r i t y and o p a c i t y a r e o b s e r v a b l e i n t h e p r e d i c t i o n s y s t e m , a s i n p u r e PSs. T h u s , i t seems t h a t a p r e d i c t i o n - b a s e d - s y s t e m , w h i l e n o t c o r r e s p o n d i n g c o m p l e t e l y t o p u r e P S s , i s s i m i l a r i n many ways t o t h e PSs a c t u a l l y i m p l e m e n t e d ( p e r h a p s t h e s e s h o u l d be c a l l e d ' e x t e n d e d P S s ' ) . One d i f f e r e n c e , however, r e m a i n s ; t h i s c o n c e r n s t h e g l o b a l i t y o f r u l e s . Onder t h e p u r e PS p a r a d i g m (and i n most o f t h e e x t e n d e d s y s t e m s ) , t h e r u l e s a r e e x p l i c i t l y g l o b a l . T h a t i s , t h e y a r e a c t i v e a t a l l t i m e s , and o n l y t h e o r d e r i n g ( t o g e t h e r , o f c o u r s e , w i t h t h e d a t a base match) d e t e r m i n e s which r u l e s may e x e c u t e . T h i s o f t e n l e a d s t o use o f c o m p l e x ' t a g s ' , s i m p l y t o b l o c k t h e e x e c u t i o n o f a p a r t i c u l a r r u l e . i n " t h e p r e d i c t i o n s y s t e m , t h e o p p o s i t e a p p r o a c h i s t a k e n . Much e f f o r t i s expended t o e n s u r e t h a t t h e s e t o f p r e d i c t i o n s a c t i v e •""•at any one t i m e i s ' r e l e v a n t ' — t h a t i s , o f i n t e r e s t t o t h e c u r r e n t s i t u a t i o n . The s y s t e m o f c a n c e l l a t i o n m a r k e r s E x t e n d i n g t h e P r e d i c t i o n System A Model f o r Discourse A n a l y s i s 103 ensures that p r e d i c t i o n s are d e l e t e d as soon as they are no l o n g e r v a l i d . T h i s approach has a l s o been used i n c e r t a i n p r o d u c t i o n systems. In p a r t i c u l a r , Moran (1973) and Rowat (1976) o u t l i n e d methods of 'grouping* r u l e s , so t h a t l a r g e r amounts of knowledge can be manipulated. MYCIN (Davis et a l , 1975) and DENDRAL have s i m i l a r mechanisms. In g e n e r a l , however, t h i s i s not emphasized i n production systems. In c o n c l u s i o n , I f e e l t h a t , while our system i s s i m i l a r to PSs, t h e r e i s a minor d i f f e r e n c e i n -approaches. In the p r e d i c t i o n * system, the data base i s r e l a t i v e l y s i m p l e , and the s i t u a t i o n a l knowledge r e s i d e s i n the set of p r e d i c t i o n s ; i n a PS, the converse i s t r u e . Thus, although the p r e d i c t i o n system i s a form of PS; and the g o a l i s the same" ( i . a . , dynamic c o n t r o l ) ; I s h a l l continue t o deal with i t as a d i f f e r e n t mechanism. V.2 Use of Information Sources In t h i s s e c t i o n , we w i l l d e a l again with the i n f o r m a t i o n sources developed i n Chapter I I , t h i s time with an eye t o i n c o r p o r a t i n g them i n t o a p r e d i c t i o n - b a s e d system. V.2.1 P r e l i m i n a r y Requirements In chapter I I , we d e s c r i b e d two t h i n q s about each I.S.: i) i t s s u r f a c e m a n i f e s t a t i o n ; i i ) the i n f o r m a t i o n i t t e l l s us. Use of Information Sources a Model f o r Discourse a n a l y s i s 104 The problem here s i l l be t o encode these i n t o p r e d i c t i o n s , so t h a t they w i l l f i t i n t o the system o u t l i n e d p r e v i o u s l y . Number (i) i s r e l a t i v e l y s t r a i g h t f o r w a r d , s i n c e r e c o g n i t i o n of s u r f a c e - s t r u c t u r e phenomena has already been f u l l y s p e c i f i e d , and i s a standard aspect of computational l i n g u i s t i c s . T h i s corresponds t o the ' t e s t ' p a r t of a p r e d i c t i o n . Number' ( i i ) might be harder. I n t u i t i v e l y , the i n f o r m a t i o n from an I.S. w i l l have to be encoded as the ' a c t i o n * part of p r e d i c t i o n s . Upon examination, I.S.'s are ssen t o c a r r y two t y p e s ' o f i n f o r m a t i o n : i ) what t o look f o r next; i i ) w h a t t o do with what's a l r e a d y been found ( i . e . , how to organize the r e p r e s e n t a t i o n ) . But, t h i s l i s t corresponds almost e x a c t l y t o the p o s s i b l e a c t i o n s a v a i l a b l e from p r e d i c t i o n s (IV.2.2, p. 70)! The problem i s almost s o l v e d . What i s missing, and what was skipped i n the e a r l i e r d i s c u s s i o n of p r e d i c t i o n s , i s some i n d i c a t i o n about the d e t a i l s of these a c t i o n s — i . e . how would they be represented i n LISP code. Number (i) above — what t o look f o r next — i s o b v i o u s l y w e l l s u i t e d to a p r e d i c t i o n based system; One LISP f u n c t i o n , PREDICT, i s r e q u i r e d ; PREDICT takes two arguments, a t e s t and an a c t i o n (plus a number of o p t i o n a l arguments), and posts the a p p r o p r i a t e p r e d i c t i o n . -Number ( i i ) — o r g a n i z i n g what's a l r e a d y been found — i s Dse of I n f o r m a t i o n Sources a Model f o r Discourse a n a l y s i s 105 more complex." I n t u i t i v e l y , what we want here i s a s e t of structured-manipulating o p e r a t i o n s t o permit us to o r g a n i z e the i n f o r m a t i o n i n the d e s i r e d way. These ope r a t i o n s w i l l depend on the l e v e l of the r e p r e s e n t a t i o n , the s t r u c t u r e o f the r e p r e s e n t a t i o n , and v a r i o u s other t h i n g s . C u r r e n t l y , a simple set e x i s t s : (RUMINaTE s l o t ) -takes the component j u s t found, and t r i e s to put i t i n t o the ' s l o t ' i n the c u r r e n t sentence. I t does t h i s i n an ' i n t e l l i g e n t 1 way, i n that i t checks t o see i ) i f the s l o t i s a l r e a d y f i l l e d ; i l ) i f f i l l i n g t h i s s l o t completes the c o n c e p t u a l i z a t i o n ; i i i ) i f any s i d e e f f e c t s should take p l a c e (as, f o r example, when the ACTION s l o t i s f i l l e d ) . (REPLaCE c o n c e p t u a l i z a t i o n ( s l o t f i l l e r ) ) - t a k e s the f i l l e r , and t r i e s t o put i t i n t o the a p p r o p r i a t e s l o t i n ' t h e a p p r o p r i a t e c o n c e p t u a l i z a t i o n . E s s e n t i a l l y , t h i s f u n c t i o n i s used t o rearrange p r e v i o u s l y - b u i l t s t r u c t u r e . I t i s d i f f e r e n t from"RUMINaTE, i n t h a t i ) the f i l l e r can be a r b i t r a r y , and not j u s t the component most r e c e n t l y found; i i ) the c o n c e p t u a l i z a t i o n i n t o which i t i s placed i s not r e s t r i c t e d t o the c u r r e n t one, but can be any p r e v i o u s c o n c e p t u a l i z a t i o n i n the meaning s t r u c t u r e ; i i i ) no c h e c k i n g i s done; REPLaCE assumes t h a t whoever c a l l e d knows what i s going on. Use of Information Sources A Model f o r Discourse a n a l y s i s 106 (EXTRACT s l o t <conceptualization>) - r e t u r n s the contents o f the given s l o t i n the given c o n c e p t u a l i z a t i o n (which d e f a u l t s t o the c u r r e n t c o n c e p t u a l i z a t i o n , i f omi t t e d ) . Thus, the example given e a r l i e r (REPLACE :SENTMEAN: (OBJ (EXTRACT ACTORJ ) ) says t o take the concept which i s i n the ACTOR s l o t of the c u r r e n t c o n c e p t u a l i z a t i o n , and put i t i n t o the o b j e c t s l o t . Besides these t h r e e f u n c t i o n s , t h e r e are a number of g l o b a l v a r i a b l e s (:SENTMEAN :, :CONCEPT:, :CLAUSE:, :COMPONENT:, etc.) where the meaning i s apparent. There are a l s o a d d i t i o n a l f u n c t i o n s whose e f f e c t s should be e q u a l l y apparent. With these f u n c t i o n s , - we are a b l e t o perform the manipulations r e q u i r e d t o use the I.S.'s e f f e c t i v e l y . V.2.2 Information Sources In d e a l i n g with the I.S.'s d e s c r i b e d here, i t must be understood that' t h e y a r e a l l , i n some sense, o p t i o n a l ; without them, a n a l y s i s c o u l d s t i l l proceed, although not as e f f e c t i v e l y . Thus, i n dev e l o p i n g t h i s c h apter, I s h a l l emphasize the added b e n e f i t s bought'by these sources, by g i v i n g an i n d i c a t i o n of what the e f f e c t s ' w o u l d be without them. In the examples given here, I w i l l assume t h a t we are d e a l i n g with a system i n which a l l the modules (I.S.'s) are f u n c t i o n i n g and generating p r e d i c t i o n s , although I s h a l l emphasize on l y one module at a time. Use o f Information Sources A Model f o r Discourse A n a l y s i s 107 staging: •- * I t w i l l be remembered that the major point here i s theme. An example of the use o f t h i s i n a n a l y s i s would be: Mr. Smith's window was broken by the f l y i n g b a l l . The presence'of the theme should be f l a g g e d , - i n t h i s c a s e , as a s i d e e f f e c t of the p r e d i c t i o n which r e c o g n i z e s passive sentences: ( (VERB PASTPART) (PROG (REPLACE : SENTMEAN: (OBJ (EXTRACT ACTOR) ) (SETTHEME OBJ) )) That i s , i n the process of r e a r r a n g i n g the s u b j e c t and o b j e c t components, the theme i s a l s o f l a g g e d . A p r e d i c t i o n which makes use o f t h i s i n f o r m a t i o n might look l i k e : ( (THEME) (EXPECT (MORE-ABOUT THEME))) which, i n the case of the example sentence g i v e n , would produce e x p e c t a t i o n s l i k e : -expect to f i n d out more about Mr. Smith's window -expect Mr. Smith t o be angry Without'some n o t i o n of theme, t h i s sentence would be t r e a t e d i n the same manner a s * i t s a c t i v e e q u i v a l e n t ; i f t h i s were the case, the set of e x p e c t a t i o n s f o l l o w i n g the sentence might be: -expect more about the b a l l -expect the b a s e b a l l game t o continue Thus, use of thematic i n f o r m a t i o n p r o v i d e s the a b i l i t y to 'focus' on a p a r t of the d i s c o u r s e . Other uses might i n c l u d e Use of Information Sources A Model for Discourse Analysis 108 the a b i l i t y to detect a change of theme; t h i s usually indicates a t r a n s i t ion in the focus, which might be s i g n i f i c a n t . cohesion: The important aspects here are information blocks and r e f e r e n t i a l s p e c i f i c i t y . Information blocks, as mentioned, are usually s i g n a l l e d by punctuation or grammatical arrangement. Compare: (a) Peter was running home, and he f e l l down. (b) Peter was running home. He f e l l down. (c) When Peter was running home, he f e l l down. In t h i s example, the act of f a l l i n g i s emphasized, to a successively greater degree, i n each of the three cases. I n t u i t i v e l y , we, would l i k e the representation to be d i f f e r e n t i n each of these three cases, (a) o - |>horae o J>down Peter<=>PTRASS <—Peter <-| >v Peter<=>PROPEL <--Peter <-| ?< / X l< (b) o |> home Peter <=> PTRANS < — Peter <-} A l< 1 then I o |> down Peter<=>PROPEL <—Peter <-| !< (c) o | Mown time o |>home Peter<=>PROPEL<—Peter<-| < Peter<=>PTRANS<--peter<-| That i s , the representation should somehow capture the r e l a t i v e Use of Information Sources A Model f o r Discourse A n a l y s i s 109 s a l i e n c e of the two s e c t i o n s , i n p a r t i c u l a r the one concerning f a l l i n g . P r e d i c t i o n s t o encode t h i s knowledge might look l i k e : ( (WORD AND) (PREDICT ((SENTENCE) (SETQ :C0NCEPT: (: OLDSENT; : NEWSENT:))))) i . e . , i f the word 'and' i s found, look f o r a second sentence to go with the f i r s t , and produce as the meaning the ANDed c o n j u n c t i o n of the i n d i v i d u a l sentences. ( (SENTENCE) then (SETQ :CONCEPT: (:C0NCEPT: < — :SENTMEAN:))| i . e . , i f two sentences are found without any o t h e r l i n k a g e between them assume a d e f a u l t 'then' l i n k , and j o i n the new sentence i n t o the p r e v i o u s l y b u i l t s t r u c t u r e . T h i s corresponds to case (b) above, where the use of separate sentences f o r the two s e c t i o n s serves t o emphasize both of them; ((CLAUSE TEMPORAL) (REPLACE :SENTMEAN: (TIME :CLAUSE:))) i . e . , ; i f a temporal c l a u s e i s found, put i t i n t o a "time' s l o t (thereby r e d u c i n g i t s importance). The r u l e s here are o b v i o u s l y somewhat s i m p l i s t i c ; they are not intended to be complete, but r a t h e r t o i l l u s t r a t e the p o s s i b l e use of i n f o r m a t i o n b l o c k s . Without the s o r t of i n f o r m a t i o n s p e c i f i e d here, a system would miss the e s s e n t i a l d i f f e r e n c e s i n importance; probably, a l l the cases would produce the same r e p r e s e n t a t i o n : Use of Information Sources A Model for Discourse analysis 110 |> home Peter<=>PROPEL<—Peter<-j |> down Many systems do make use of t h i s sort of information. Onfortunately, t h e i r work i s i m p l i c i t , i n that rules concerninq information i n j e c t i o n are not e x p l i c i t l y stated. I f e e l that to function e f f e c t i v e l y , these r u l e s must be recognized by the system designer, and be encoded as such. Refe r e n t i a l s p e c i f i c i t y i s somewhat harder to deal with. A sample of the problem can be seen i n the sentence p a i r s : What we want here i s the a b i l i t y to detect the f a c t that the r e p e t i t i o n of the d e f i n i t e noun phrase i n (b) i s mildly unusual, and that i t possibly r e f e r s to a d i f f e r e n t b a l l (or that the story i s aimed at a younger audience). predictions. Rather, I would envision a separata reference r e s o l u t i o n program, which i s c a l l e d upon to handle any anaphora. This program would have two sections to turn to f o r guidance: i) a l i s t of the concepts currently in working memory, from which the current referent should be chosen; i i ) a l i s t of pronominalization r u l e s f o r discourse, used to t e l l whether the current reference i s over- or (a) John h i t the b a l l hard. I t flew in a long, high arc. (b) John h i t the b a l l hard. The b a l l flaw i n a long, high arc. This sort of information i s not e a s i l y represented by Ose of Information Sources A Model for Discourse Analysis 111 under-specified (a sample of these rules i s contained in Charniak (1972), as borrowed from Lees and Klima (1963)) . Obviously more work needs to be done here. In p a r t i c u l a r , i t ' s not c l e a r what an overspecified referent t e l l s us, or more importantly, how to use that information. Some forms of t h i s are easier to use than others. One example i s 'alternative': John could have been safe, but he didn't run quickly enough. where the negating conjunction serves to emphasize the (unstated) f a c t that John was out. One s i m p l i s t i c way to handle t h i s i s to have i t flagged by the prediction which recognizes 'but': ((WORD BUT) (ASSERT (NEG :SENTMEAN:))) i . e . , i f the word 'but' i s found, assert the negation of the f i r s t clause (in t h i s case; that John was not safe). Again, t h i s i s obviously a r e s t r i c t e d approach; among other things, i t f a i l s to d i f f e r e n t i a t e between di f f e r e n t meanings of 'but'. However, these are e s s e n t i a l l y bookkeeping d e t a i l s , and could be worked out without too much d i f f i c u l t y . Note, i n c i d e n t a l l y , that the content of the second clause i s t o t a l l y i r r e l e v a n t here. That i s , once we see: John could have been safe, but .... we know that he was out, and can assert such without waiting for Use of Information Sources A Model for Discourse Analysis 112 the rest of the sentence. Without t h i s c a p a b i l i t y to use alternatives, the system might never r e a l i z e the i m p l i c i t meaning of the sentence. In the example here, I f the fact that John was out i s not e x p l i c i t l y mentioned elsewhere, the system would never know t h i s — an obvious omission. Another example of c o l l a t e r a l i s foreshadowing: The boys were playing b a l l dangerously close to Mr. Smith's house. where the word 'dangerously* hints that possible trouble i s ahead (this might also be c a l l e d 'evaluative'). A prediction to handle t h i s might be: ( (AD? EVALUATIVE) (EXPECT VALUE)) which i n t h i s case, would produce an expectation l i k e : -expect something bad which, once the inferencer got through with i t , might be -expect Mr. Smith's window to be broken Without t h i s p a r t i c u l a r feature ( i . e . recognition of foreshadowing), the system would s t i l l perform the analysis, but in a d i f f e r e n t manner. The expectation generated would be: -expect the baseball game to continue Thus, the breaking of the window, i f indeed i t happened, would come as a surprise, and some extra processing would be reguired to handle i t . • Obviously, humans would detect a loaded word such as 'dangerously'; the system should do the same. Use of Information Sources A Model for Discourse Analysis 113 structure: This has been discussed thoroughly i n a number of places; what remains i s to show how i t can be incorporated into a prediction system. Given the f a i r l y simple grammar mentioned previously: story --> s e t t i n g + episode • denouement setting —> temporal location f s p a t i a l location J (character) * f continuing event episode —> event \ action + reaction The analysis of a story might begin with -expect a s e t t i n g which would be converted into the c a l l (EXPECT (OR (LOCATION TEMPORAL) (LOCTATION SPATIAL) (CHARACTER) (EVENT CONTINUOUS))) Thus, the sentence John and Peter were coming home from school, besides s a t i s f y i n g a l l the l o c a l (i.e. sentence-level) predictions, would also s a t i s f y the expectation for a s e t t i n g , and hence would produce -expect an episode as the text grammar continues. This i n turn would generate the c a l l (EXPECT (XOR (EVENT) (AND (ACTION) (REACTION)))) and the cycle repeats. The most s i g n i f i c a n t benefit of t h i s i s that i f a story i s Use of Information Sources A Model for Discourse Analysis 114 f i n a l l y recognized (i.e. the expectation 'expect a story' i s s a t i s f i e d ) , the meaning of the story can be organized and manipulated i n whatever way i s desired. E s s e n t i a l l y , we have captured the a b i l i t y to deal with the story at the macrostructural l e v e l , and thus to impose our own organization on i t . This corresponds to the 'semantics' of the text grammar. Without the use of text grammars, two flaws occur in the system. F i r s t , the expectations w i l l be less directed, since the system has l i m i t e d knowledge of how events should follow each other. This i s a comparatively minor d i f f i c u l t y ; the system can get by without the added expectations the text grammar provides. Second,' and more importantly, the system w i l l have no idea of what constitutes a coherent story. It w i l l be able to deal with the- linkages at a l o c a l l e v e l (i.e. between pairs of sentences: causal, temporal, e t c . ) , but w i l l not be able to treat a text as a u n i f i e d whole. The current work with s c r i p t s , etc., i s an obvious attempt to remedy t h i s . context: -Again, there are two relevant aspects: foregrounding and frames. Foregrounding i s e s s e n t i a l l y a superset of the problem of reference (mentioned under 'cohesion'). As I see i t , foregrounding should not be a part of the prediction system i t s e l f , but would have i t s own module, at the 'meta^prediction* l e v e l . I t s task would be to foreground the desired concepts; as Use of Information Sources A Hodel for Discourse Analysis 115 mentioned a set of rules can be delineated specifying when and how a concept i s to be foregrounded and backgrounded. Foregrounding of i t s e l f serves no purpose, except to move concepts i n and out of the (bounded) working memory. This working memory i s used by other modules such as reference resolution, An example would be: There's a b a l l in the yard. I t ' s green, where the foregrounding routine w i l l recognize the i n s t a n t i a t i o n of a new concept (the ball) , and 'activate' that concept. Thus, when reference resolution i s required (as i n the second sentence), the resolution routines w i l l have a easier job decidinq what ' i t ' refers to (note also that the 'theme' module would probably also help here, by Indicating that 'the b a l l ' i s the theme of the f i r s t sentence). Without t h i s foregrounding c a p a b i l i t y , the system loses a l o t of the power which humans use so e f f e c t i v e l y i n processing language. That i s , without an e f f e c t i v e representation of what i s 'on stage* at the present moment, we are faced with p o t e n t i a l l y explosive searches i n various phases of the processing. Frames, i n the s i t u a t i o n a l sense, provide s i m i l a r power. Like foregrounding, the frame i n s t a n t i a t i o n mechanism would operate at the meta-prediction l e v e l , i n that frame recognition and i n s t a n t i a t i o n would be performed by a separate module, outside the normal prediction system. Unlike foregrounding, Ose of Information Sources A Model f o r Discourse A n a l y s i s 116 however, the i n f o r m a t i o n t h a t the frame c a r r i e s would be i n j e c t e d i n t o the'normal p r e d i c t i o n l i s t . For example, when the frame f o r • b i c y c l e * i s i n s t a n t i a t e d , a s e r i e s of e x p e c t a t i o n s might be spawned; corresponding t o our knowledge about the s t r u c t u r e and p o s s i b l e uses of a b i k e : -expect to hear more about the bike i t s e l f -expect t o hear about the handlebars -expect t o hear about the f e n d e r s • • • • -expect t o hear about someone r i d i n g the bike -expect to hear about someone l o c k i n g the bike • • • • Thus, a combination such as: Jane was r i d i n g her new b i c y c l e . The f e n d e r s were a p r e t t y shade of red. could be handled, s i n c e there would e x i s t an e x p e c t a t i o n f o r fe n d e r s , which would r e s o l v e them as part of the b i c y c l e . T h i s use of u n c o n t r o l l e d forward i n f e r e n c i n q i s o b v i o u s l y e x p l o s i v e ; perhaps b e t t e r power c o u l d be a t t a i n e d through an inference-on-demand ( i . e . deep binding) approach. For the moment, I ' l l l e t t h i n g s stand. The need f o r a frame system i s obvious. Without i t , the sample sentences g i v e n above c o u l d h a r d l y be handled. True, we could t r i g g e r ' a ' s e a r c h through the semantic net, which would e v e n t u a l l y d i s c o v e r t h a t f e n d e r s are PART-OF a b i c y c l e ; however, t h i s approach would- be even more e x p l o s i v e than the frame method, and, b e s i d e s , would miss the e s s e n t i a l f a c t t h a t humans 0"se of Information Sources A Modal f o r D i s c o u r s e A n a l y s i s 117 have t h e i r knowledge ' c l u s t e r e d * I n t o u s e f u l chunks (see S c r a q g , 1976, f o r some i n t e r e s t i n g comments on t h i s ) . N ote, i n c i d e n t a l l y , t h a t t h e g e n e r a l n o t i o n o f frames e n a b l e s us t o s a t i s f y t h e l a s t o f t h e p s y c h o l o g i c a l c r i t e r i a — t h a t memory be an a c t i v e , s e l f - o r g a n i z i n g p r o c e s s . M i n s k y ' s frames a r e based v e r y h e a v i l y on B a r t l e t t ' s schemata, so the cor r e s p o n d e n c e i s no a c c i d e n t . r e a l - w o r l d knowledge: T h i s module has p l a y e d a l a r g e r o l e i n t h e examples t o t h i s p o i n t ; I t ' s t i m e t o e x p l a i n how t h i n g s work. E s s e n t i a l l y , I see i n f e r e n c e s as b e i n g v e r b - d r i v e n ; t h a t i s , t h e s e t of p o s s i b l e i n f e r e n c e s would be keyed by the s u r f a c e - s t r u c t u r e v e r b s found i n each se n t e n c e . T h i s s e t of i n f e r e n c e s would be l o a d e d I n t o w o r k i n g memory, hence would be s u b j e c t t o t h e s i z e c o n s t r a i n t s mentioned p r e v i o u s l y . Thus, we have an upper bound on the number o f p o s s i b l e i n f e r e n c e s a c t i v e a t any one time ( t h i s s h o u l d reduce t h e problem o f c o m b i n a t o r i a l e x p l o s i o n , a l t h o u g h i t w i l l not e l i m i n a t e i t ) . I n f e r e n c e s t h e m s e l v e s w i l l be o f t h e form: ( TEMPLATE ACTION ) where t e m p l a t e - i s t h e p a t t e r n t o be matched i n the c u r r e n t memory (and t h e match can be q u i t e complex) 0"se o f I n f o r m a t i o n S ources A Model for Discourse Analysis 118 action - i s the inference to be made i f the template can be matched. For example: attached to the verb 'go', we might find th» inference !-> Y { ( X <=> PTRANS <—X <-| ) ( X<=> loc (Y) ) ) !-< i . e . , i f we f i n d that person X i s going to location Y, i n f e r that l a t e r on X might be at location Y. Thus, i f we receive the input John was going to school, we might make the inference that John i s at school. Note that Inferences, as described, are s i m i l a r to predictions. Actually, they work at a somewhat higher lever, and have di f f e r e n t c h a r a c t e r i s t i c s . E s s e n t i a l l y , inferences provide two types of information. F i r s t , there i s fac t u a l information that might be reguired for la t e r deductions. This i s the type shown above, where the fact that John i s at school might never be e x p l i c i t l y stated, but might be needed for l a t e r processing. The other type of information concerns expectations about what might be seen next. This was i l l u s t r a t e d e a r l i e r , where the sentence 'John hated Mary' produced the inference that John might do something to harm Mary. This i s probably the more important of the two types, since i t provides d i r e c t i o n to guide t h Q analysis of l a t e r input. That i s , i t t e l l s us not only what to expect next, but also what to do when we fi n d i t . Again, t h i s r e f l e c t s what humans seem to do. I n t u i t i v e l y , Use of Information Sources A Model f o r D i s c o u r s e a n a l y s i s 119 t h e y p e r f o r m a c e r t a i n amount o f i n f e r e n c i n g , i n p r e p a r a t i o n f o r w h a t e v e r f o l l o w s . W i t h o u t t h i s s o r t o f c a p a b i l i t y , t h e s y s t e m i s o b v i o u s l y r e s t r i c t e d , s i n c e i t has no way o f f u l l y c o m p r e h e n d i n g t h e meaning o f t h e s e n t e n c e , o r o f d i s c o v e r i n g l o g i c a l i m p l i c a t i o n s . One example: c a u s a l l i n k s a r e v e r y s e l d o m s t a t e d e x p l i c i t l y , and must u s u a l l y be i n f e r r e d ; w i t h o u t some i n f e r e n c e c a p a b i l i t y , t h e s y s t e m w i l l be u n a b l e t o c o n n e c t e v e n t s t o g e t h e r . The a p p r o a c h h e r e o b v i o u s l y d o e s n o t s o l v e t h e i n f e r e n c e p r o b l e m ; i n p a r t i c u l a r , t h e d a n g e r o f e x p l o s i o n s t i l l r e m a i n s . Some c o n t r o l l i n g mechanisms t o p r e v e n t t h i s have been s u g g e s t e d ; p e r h a p s o t h e r s would be needed. F o r t h e moment, w e ' l l l e t t h i n g s s t a n d . T h i s c o n c l u d e s t h e s e c t i o n on I n f o r m a t i o n S o u r c e s . Of t h e s i x c o v e r e d , i t w i l l be n o t e d t h a t the f i r s t t h r e e a r e ' s t y l i s t i c * , and t h e l a s t t h r e e ' s e m a n t i c * . E s s e n t i a l l y , t h e y a r e m a n i f e s t d i f f e r e n t l y i n d i s c o u r s e , a l t h o u g h t h e i r f u n c t i o n and meaning a r e e s s e n t i a l l y t h e same. T h i s c h a p t e r h a s shown t h a t a s i n g l e , u n i f o r m method — p r e d i c t i o n s — c a n be u s e d t o e n c o d e d i v e r s e t y p e s o f i n f o r m a t i o n . 7.3 A D e t a i l e d Example To c o n c l u d e t h e c h a p t e r , we w i l l f o l l o w t h e a n a l y s i s o f one c o m p l e t e p a r a g r a p h - l e n g t h s t o r y . The s t o r y we s h a l l be c o n c e r n e d w i t h i s : A D e t a i l e d Example A Model f o r Discourse A n a l y s i s 120 John and Peter were coming home from school. As they were walking through the playground, they saw an animal i n the gra s s . I t was a r a b b i t . The boys chased a f t e r i t , but i t hopped away. They went home. He w i l l present the a n a l y s i s here i n a manner s i m i l a r t o t h a t used f o r the s i n g l e - s e n t e n c e example shown e a r l i e r , except t h a t fewer d e t a i l s w i l l be given. Assume t h a t a l l of the I.S.*s mentioned e a r l i e r are a v a i l a b l e , and that the t e x t grammar we a r e using i s the one s p e c i f i e d before. At the s t a r t , we have the standard d e f a u l t p r e d i c t i o n (a) ( (NG) (RUMINATE SOBJ) ) plus the f o u r p r e d i c t i o n s spawned by the e x p e c t a t i o n f o r s e t t i n g . (b) ( (LOCATION TEMPORAL) (RECORD SETTING) ) (c) ((LOCATION SPATIAL) (RECORD SETTING) ) (d) ( (CHARACTER) (RECORD SETTING) ) -(e) ( (ACTION CONTINUING) (RECORD SETTING) ) The p r e d i c t i o n manager looks these over, and d i s c o v e r s t h a t two of them (the d e f a u l t one f o r NG, and the one f o r ch a r a c t e r s ) are e s e n t i a l l y l o o k i n g f o r the same t h i n g . I t merges these, producing a new p r e d i c t i o n , with a higher s t r e n g t h (not shown): (f) ((NG) (PROG (RECORD SETTING) (RUMINATE SOBJ))) The p r e d i c t i o n s a c t i v e are now (b) , ( c ) , (d) , (e) , and (f) . Note t h a t (b); because of i t s p a r t i c u l a r s t y l e of c a n c e l l a t i o n , has not been k i l l e d o f f by the merger. The phrase John and Peter i s now found, s a t i s f y i n g ( f ) , which removes i t s e l f , but spawns a new p r e d i c t i o n : A D e t a i l e d Example A Model for Discourse Analysis 121 (g) ((VG) (ROMINATE ACTION)) Again the '-predict ion manager recognizes t h i s , and merges i t with (c) , to form (h) ((VG)< (PROG (RUMINATE ACTION) (RECORD SETTING)) ) Thus, analysis proceeds through the sentence, i n a more-or-less straightforward manner. At the end of the sentence, the only predictions active are (b) , (c) , (d) , and (e). However, the fact that the expectation f o r •setting' has been s a t i s f i e d causes a l l of these to be removed. At t h i s point; some other modules begin to contribute. Since the setting has been s a t i s f i e d , the text grammar recognizes t h i s , and an expectation for an 'episode' i s generated. This translates at the lower l e v e l to a prediction for an event, or an action/reaction pair: (h) ( (EVENT) (RECORD EPISODE) ) (i) ( (ACTION) (PREDICT (REACTION) (RECORD EPISODE) ) ) The foregrounding module has also been at work, d u t i f u l l y noting that i n the f i r s t sentence, Peter, John, home, and school were a l l mentioned. These four concepts are thus foregrounded, i . e . , loaded into working memory. The frame module"has also'been active , recognizing that the 'school* frame should probably be instantiated. This creates expectations for school-house, playground, parking l o t , and various other facets. The real-world component i s dormant at t h i s point. I t has noted that John and Peter are moving from school to home, but has also noted the progressive tense, which indicates that the A Detailed Example a Model for Discoucse Analysis 122 action has not yet been completed. The representation to t h i s point i s : Peter p Peter J-> home & <=> PTRANS < S <--! John John |-< school (note: the "p1 above the * <=>' denotes the progressive tense; i n general tense markers w i l l be omitted from the diagrams, for the sake of s i m p l i c i t y ) . and the set of predictions active i s something l i k e •. .. 5 •. . . ... (h) ( (EVENT) (RECORD EPISODE)) (i) ((ACTION) (PREDICT (REACTION) (RECORD EPISODE))) (j) ((PLAYGROUND) (FRAMEREF SCHOOL)) plus predictions for other aspects of the 'school* frame, not shown here. Since we have no strong predictions as to the form or content of the second sentence, we s t a r t o f f with the "default* set (in order): (k) ((NG) (RUMINATE ACTOfi)) (1) ((PG) (RUMINATE (TIME SPACE))) (m) ((CLaUSE SUBORDINaTE) (RUMINa*P2 (IN3T MANNER))) (note that t h i s i s remarkably s i m i l a r to the standard 'start* node of an ATN: with one s i g n i f i c a n t difference: prediction (1) w i l l pick up A Detailed Example A Model for Discourse Analysis 123 l o c a t i v e or temporal phrases anywhere i n the sentence; with an ATN, the arc would have to be duplicated, at every node at which i t could conceivably occur.) Thus, we begin parsing the second sentence. The subordinate clause i s picked up f i r s t . As i t i s being processed, the c o l l a t e r a l module has two b r i e f moments of a c t i v i t y . F i r s t , the pronoun 'they* i s interpreted; t h i s causes no d i f f i c u l t y , since the theme of the previous sentence was 'Peter and John 1, and the concept 'Peter and John' i s also foregrounded. In addition, the use of a pronoun at t h i s point in the discourse i s consistent with the somewhat s i m p l i s t i c rules f o r s p e c i f i c i t y . The other s i g n i f i c a n t point occurs when 'the playground' i s encountered. The use of the d e f i n i t e NG i s explained by the expectations spawned by the 'school' frame, so everything i s a l l right. However, the frame module jumps i n again at t h i s point, to i n s t a n t i a t e the 'playground' frame, with appropriate expectations f o r grass, swings, etc. To return to the top l e v e l : the system finishes picking up the subordinate clause. At t h i s point, the c o l l a t e r a l module again becomes active, r e a l i z i n g that the use of information blocks here i s s i g n i f i c a n t . In accordance with i t s p r i n c i p l e s , i t overrides the default prediction, which i s t r y i n g to treat the clause as either instrument of manner, and instead places the clause i n the 'time* s l o t of the newly b u i l t structure. Parsing continues, with the rest of the sentence being handled i n the obvious way, to produce: A Detailed Example A Model for Discourse analysis 124 John o & <=>MTRaNS <—animal < Peter a time John o John l-> & <=>PTRaNS <-- & <-| plygd Peter Pater f-< I loc grass again'nothing unusual has happened; the theme has remained consistent, but a new concept, an animal, has been introduced. Mote that the d e f i n i t e reference to grass i s v a l i d , since grass i s mentioned i n the playground frame. Interestingly, d i r e c t mention of grass In'the school frame would be i n v a l i d ; the playground frame must be invoked f i r s t . Scragg (1975, pp. 9-11) discusses'this problem of 'continuity of contexts *. Since' the conceptualization i s recognized as an action, a prediction for 'reaction' i s spawned, together with the ones s t i l l active from the school and playground frames. as we move into the next sentence, the word ' i t ' i s picked up f i r s t ; This causes a couple of i n t e r e s t i n g e f f e c t s : f i r s t of a l l , the pronoun is'resolved; t h i s causes no problem, as there i s o n l y one concept currently foregrounded which i t could match, and t h i s binding i s performed. Second, the theme module, on looking t h i s over, discovers that the theme has changed. This t e l l s i t that•we are not t a l k i n g exclusively about Peter and John, but " rather * about an event in which they happen to be involved. This^causes the system to 'back o f f " a l i t t l e , and reorganize the global structure i t i s building. Processing of"the sentence continues; withont any unusual happenings. However, when* the sentence i s interpreted, the system r e a l i z e s that i t i s but a further s p e c i f i c a t i o n of a a Detailed Example ft Model for Discourse analysis 125 previously mentioned concept. Thus (in accordance with and the p r i n c i p l e of 'integrated memory') i t does not build a separate conceptual structure for t h i s , but instead adds the new information to the old s l o t . The l a s t conceptualization now looks l i k e : John o & <=> MTBftNS <-- rabbit . . . . Peter a ! loc grass (note that I have been using proper names and generic names to label concepts; of course, these are only 'token' instances of •type' nodes i n a semantic net — the l a b e l l i n g used here i s for convenience) . Moving on, we begin the next sentence, and pick up the f i r s t clause: 'the boys ran af t e r i t ' . One point i s worth mentioning here — the use of 'the boys' rather than 'they' to refer to Peter and John i s consistent, because t h i s concept was not mentioned i n the previous sentence, hence faded 'off-stage'. Thus, the use of 'they' here would have caused problems in resolution. Once the clause i s picked up, a structure i s b u i l t f or i t : John o John |-> rabbit & <=> PTSaNS < S < j Peter Peter \-< at t h i s point, the inferencing module, at l a s t , i s allowed to say something. Triggered by the verb 'chased', i t loads a series of inferences: a Detailed Example A Model for Discourse Analysis 126 o |-> y |-> (X <=> PTBANS <—X <-| ) (Y <==f frightened ) ) ,-< |-< i . e . , i f something chases an animal Y, that animal might be frightened. |-> o |-> away (Y <==! frightened ) ( Y <=> PTBANS <--Y <--| ) ) !-< 1-< i . e . , i f an animal Y i s frightened, that animal w i l l move away from whatever i s frightening i t . o |-> Y X |-> (X <=> PTBANS <—X <-| ) ( <==| physcont ) ) !-< Y 1-< i . e . , i f X i s chasing Y, X might catch Y. Figure 10 - Set of Inferences for the Verb 'Chased' The f i r s t and t h i r d of these can go immediately, and the second i s f i r e d by the hypothetical s i t u a t i o n which i s instantiated as a r e s u l t of the f i r s t (this chaining effect i s very s i m i l a r to Bieger's work). Thus we have three expectations: -that the rabbit w i l l be frightened -that i t w i l l run away -that the boys w i l l catch i t Armed with these, the system analyzes the second part of the sentence. Conveniently, the structure found (that of the rabbit hopping away) corresponds to one of our expectations (the second). Thus, the chain of causal inferences joining the two A Detailed Example (a) ( (b) ( (c) ( A Model for Discourse Analysis 127 events i s instantiated, with the f i n a l form represented as: John John I-> rabbit S <=> PTRANS < & < 1 Peter Peter |-< A 111 c 111 !-> rabbit <==| frightened ...,-<. A 111 C . y III 1-> away rabbit <=> PTRANS <-- rabbit | |-< Figure 11 - A Causal Chain One more point remains, however; the use of the word 'but' s a t i s f i e s the c o l l a t e r a l check for alternatives, hence the only remaining r e s u l t of the f i r s t clause Is negated: John & Peter -i |-> /\ <====| physcont rabbit |-< At t h i s point, we have found an action/reaction sequence, which s a t i s f i e s the c r i t e r i a f o r an episode. Thus, the text grammar proceeds, and predicts for the only remaining feature, denouement. • • ....... The l a s t sentence s a t i s f i e s t h i s expectation, and the story i s f i n i s h e d , with the r e s u l t being: A Detailed Example A Model for Discourse Analysis 128 John p o John |> hone 6 <=> PTRANS <— & <-] Peter Peter |< school / / / se t t i n g / !_ John o time John o John |-> & <=>MTRANS < — r a b b i t < & <=>PTRANS <— 6 <-| plygd Peter A Peter Peter |-< I l o c A grass I | then l John o John | ->rabbit John & Pater -» | -> 6 <=>PTBANS < & <-) / \ / \ <====| physcont Peter Peter |-< rabbit |-< A 111 c I I I 1-> rabbit <==| frightened |-< A I I I c I I I l-> away rabbit <=> PTRANS < — rabbit | • !-< / / / denouement / !_ John o John |-> home & <=> PTRANS < — 6 < -| Peter Peter |-< Figure 12 - F i n a l Representation of »Rabbit» Paragraph A Detailed Example A Model for Discourse Analysis 129 V.U The Implementation The system described here has been p a r t i a l l y implemented, to the point that the remaining work should be (hopefully) straightforward. The system i s written i n LISP/MTS (Wilcox and Hafner, 1974), and runs in 290K bytes (including the LISP interpreter) on an IBM 370/168. Of' the sections diagrammed i n the thesis (p. 92), the analysis ( i . e . , s i n g l e sentence), prediction control ( i . e . , execution), and prediction manager have been f u l l y implemented. The modules corresponding to the various I.S»s remain i n d i f f e r e n t stages of completion. Theme, real-world knowledge, and foregrounding are r e l a t i v e l y complete; text structure and c o l l a t e r a l , somewhat l e s s so; frames and cohesion are s t i l l very f l u i d . • I f e e l that the system could be completed i n a reasonable amount of time; the remaining work consists mainly of empirical observations, to provide breadth for the system. However, there s t i l l remains the p o s s i b i l i t y of some unseen conceptual block destroying an unsubstantiated premise; Of the work remaining, probably the most i n t e r e s t i n g would l i e in modification to the prediction manager. That i s , the prediction manager as currently implemented i s somewhat s i m p l i s t i c , because of an absence of thorough testing. The next step should be to run the system on a large corpus of examples, observing the decisions demanded of the prediction manager, and The Implementation A Model for Discourse Analysis 130 making extensions accordingly. With respect to the sections Implemented thus f a r , probably the most i n s t r u c t i v e has been the prediction control system I t s e l f . The basic nature of predictions has changed several times, as increasingly complex actions and interactions have been demanded. The current form of predictions -- with TEST, ACTION, STRENGTH, SOURCE, CANCEL, and OBL? — was developed as successive needs arose; there i s no guarantee that i t w i l l not be changed again. This work has also provided considerable insight into the questions of control of large systems; in pa r t i c u l a r a deeper understanding of psychologs was gained. Probably the largest single benefit gained i n the implementation e f f o r t was an appreciation of the complexity of natural language, and the structures required to handle i t . In conclusion, the implementation has been educational, i f not completely successful. The Implementation 131 VI. Conclusions VI.1 Review This paper has been an attempt to delineate the problems involved In discourse analysis, together with some possible solutions. Discourse i t s e l f was described, and six features staging, cohesion, c o l l a t e r a l , text structure, context, and real-world knowledge -- were Ide n t i f i e d as being p o t e n t i a l l y useful i n the analysis process. After appropriate psychological and computational preparation, a model for discourse analysis was presented. This model r e f l e c t e d two design features: i) use of a modified prediction system, to handle the complexities of discourse; i i ) a 'sufficient-information* assumption, which s p e c i f i e s that natural language analysis must proceed by extracting the maximum amount of information available at each point i n the processing. The model was specified In d e t a i l , and examples were presented. I f e e l that the system described here i s reasonably sound. The l i s t of information sources provides an extensive, although Review Conclusions 132 not complete, characterization of the useful features of discourse. The development of the prediction-based system of analysis i l l u s t r a t e s one possible method; there are undoubtedly others. Host importantly, t h i s thesis has discussed one approach to the problem of discourse analysis, and shown that i t i s viable. VI.2 Future Work The next step- in t h i s framework would be to f i n i s h the implementation. Rs discussed i n section V.U, t h i s i s semi-complete, but much remains to be done. The e f f o r t of specifying the system at the l e v e l of code would probably reveal some hidden flaws, as well as some intere s t i n g e f f e c t s . Also, a complete implementation would make the system avai l a b l e for testing of various sorts, concerning e f f i c i e n c y , completeness, etc. >..... After t h i s , more empirical work could be performed. That i s , the system could be run on a large corpus of examples, to determine: i) the c h a r a c t e r i s t i c s of each information source. The system developed in Chapter I I was somewhat s u p e r f i c i a l ; more d e t a i l i s needed. i i ) a d d i t i o n a l information sources. There are obviously features of discourse which have not been mentioned here (e.g., presupposition, d i c t i o n ) . These and others must be empirically developed. Future Work Conclusions 133 i i i ) general control reguirements of the system. The current control structure (diagrammed on p. 92) represents the present system. With more thorough te s t i n g , t h i s would evolve, as increasingly complex demands were put on the system. This work would augment the breadth of the system something that i s e n t i r e l y too r e s t r i c t e d at present. The f i n a l step would be to i t e r a t e , and return to the design phase, t h i s time with a more s p e c i f i c set of constraints. Probably a number of things wuld change here. The use of Conceptual Dependency as a representation, while adequate for the demands made of i t so f a r , seems to be approachinq the l i m i t s of i t s effectiveness. A more e f f e c t i v e representation would be needed; t h i s would involve a number of questions about the use of primitives, prepositional vs. analogical forms, etc. The prediction system would also undergo modification. Based on the (somewhat belated) discovery that i t Is a form of psycholog, I would try to use the results of work on PSs (in p a r t i c u l a r , knowledge-source ones l i k e DEMDRAL and MYCIN) to provide a computationally (and theoretically) more elegant system. Other, minor, modifications would also occur. The prediction manager, i f i t s t i l l existed, would be redesigned, on a more t h e o r e t i c a l basis. The "analysis 1 module, and some of i t s methods, would probably also need revision. Thus, the next step i s — keep going. Future Work Conclusions 134 VI,3 The Future of Discourse Analysis I f e e l " that- discourse analysis has a d e f i n i t e place in computational l i n g u i s t i c s . Discourse i s one of the r i c h e r areas of l i n g u i s t i c s ; at the same time, i t i s one of the most structured. This inherent structure f a c i l i t a t e s the process of analysis, and provides a l o t of gratuitous information for our use. Why not use i t ? In conclusion, the thesis has indicated that we need a f l e x i b l e , dynamic system to handle discourse, and has discussed one such system — predictions. The Future of Discourse Analysis 135 VII. Bibliography. Ammon, P.R., The P e r c e p t i o n of grammatical r e l a t i o n s i n sentences, J . V e r b a l L e a r n i n g and V e r b a l Behavior, 7, 1968, 869-875 "~ ™ ~ " ~~ Anderson, J . , Language, Memory, and Thought, Potomac, Erlbaum, . - • • 1976 ~ — Anderson, J.R. and G.H. Bower, Human A s s o c i a t i v e Memory, Toronto, Winston, 1973 ~~ " ~ """ B a r t l e t t , F., Remembering: a Study i n Ixperiienta1 and S o c i a l hology, Cambridge, Cambridge U n i v e r s i t y ~ P r e s s , ~ T 9 3 2 ~ Becker, A.L., Symposium of the paragraph, College. Composition and Communication, 17, 1966, 67-72 Bever-, T. , The C o g n i t i v e B a s i s f o r Language S t r u c t u r e s , i n C o g n i t i o n and the Development of Language, J.R. Hayes "(Id.T,"New~York7~Wiley7~1970"" ' " Bobrow, D. and D.'Norman, Some P r i n c i p l e s of Memory Schemata, i n R e p r e s e n t a t i o n and Understanding, D. Bobrow and A. C o l l i n s (Id.), New York, Academic Press, 1975 Bock, J.K. and W.F. Brewer, R e c a l l of sentences with a l t e r n a t e s u r f a c e s t r u c t u r e s , J . Exp, Psych., 103, 1974, 837-843 Borgida, A., Topics i n the Understanding of E n g l i s h Sentences by Comjjuter, T * U n i v e r s i t y of Toronto, 1975 B r a n s f o r d , J. and J . Franks, The a b s t r a c t i o n of l i n g u i s t i c i d e a s , C o g n i t i v e Psychology, 1, 1971, 331-350 B r a n s f o r d , J . , J . B a r c l a y , and J . Franks, Sentence memory: a c o n s t r u c t i v e vs. i n t e r p r e t i v e approach, C o g n i t i v e £S2£k2l2ai# 3,1972, 193-209 Carney, T., Content A n a l y s i s : A Technigue f o r Systematic I n f e r e n c e from Communications, Winnipeg, U n i v e r s i t y o f Hanitoba~Presi,~1972 ~ Cercone, N., Representing E n g l i s h i n Extended Semantic Nets, TR 7 5 - 1 9 , " u n i v e r s i t y of Alberta,~1975~~~ ~ Chafe, W., Meaning and the S t r u c t u r e of Language, Chicago, U n i v e r s i t y o f Chicago P r e s s , 1970 B i b l i o g r a p h y 136 Language and memory. Language, 49, 1973, 261-284 A , Language and consciousness, Language, 50, 1974, 113-134 Charniak, E., Toward a Model o f C h i l d r e n ' s Story Comprehension, AI-TR-266,~M7I.T. ,~1972 ~ ' Chomsky, N,, Aspects o f the Theory, of Syntax, Cambridge, M.I.T. Pre s s , 1965 •, Deep S t r u c t u r e , Surface S t r u c t u r e , and Semantic I n t e r p r e t a t i o n , i n Semantics, D.D. Steinberg and L. Jak o b o v i t s (Ed.), Cambridge, Cambridge U n i v e r s i t y Press, 1971 C h r i s t e n s e n , F., Symposium on the paragraph. College Composition §Hd Communication, 17, 1966, 60-66 C l a r k , H.H. and E.V. C l a r k , Semantic d i s t i n c t i o n s and memory f o r complex sentences, Quart. J . Exp. Psych., 20, 1968, 129-138 C o f e r , C.N., The S t r u c t u r e of Human Memory, i n San F r a n c i s c o , Freeman (Ed.), 1976, Cohen, P.F., A Prototype N a t u r a l Language Understanding System, TR-64, U n i v e r s i t y of Toronto, 1974 ~ Cr o t h e r s , E., Memory S t r u c t u r e and the R e c a l l of D i s c o u r s e , i n Language Comprehension o f Knowledge, J . C a r r o l l and R. Freedle -(Ed7)7~Washington, Hinston7 1972 Davidson, J.E., Sources o f Information i n Discourse A n a l y s i s , Proc. l§t CSCSI^SCEIO NatJ.1 Conference, Vancouver, 1976a , An E v a l u a t i o n of Conceptual Dependency as a Representation f o r Language and Knowledge, unpublished term paper, U.B.C., 1976b Davis, R., B. Buchanan, and E. S h o r t l i f f e , P roduction-Rules as a Representation f o r a Knowledge-based C o n s u l t a t i o n Program, A7l7 Memo 26?7 Stanford~University7~1975 ~~ Davis, R. and J . King, An Overview of Production Systems, A.I. Memo 271, S t a n f o r d U n i v e r s i t y , 1975~ van D i j k , T.A., Some Aspects of Text Grammars, The Hague, Mouton, 1972 ~ ~ •"' E a r l , L., Use o f word government i n r e s o l v i n g s y n t a c t i c andsemantic ambiguity. Information Storage a.nd R e t r i e v a l , 9, 1973, 639-664 B i b l i o g r a p h y 137 E a r l e y , J . , An e f f i c i e n t c o n t e x t - f r e e p a r s i n g a l g o r i t h m . Comm. A.C.H., 13, 1970, 94-102 Erman, L. and 7. L e s s e r , A M u l t i - L e v e l O r g a n i z a t i o n f o r Problem-^solving Using Many, D i v e r s e , Cooperating Sources of Knowledge, Proc. 4th I n t ^ l J o i n t Conference A r t i f i c i a l I h t e l l i g e n c e 7 ~ T b i l i s i , " l 9 7 5 , 126-133 Fabens, W., PEDAGLOT and understanding n a t u r a l language p r o c e s s i n g , Amer. j . Computational L i n g u i s t i c s , 1975, mf 32 Fiegenbaum, E. et a l . , On G e n e r a l i t y and Problem S o l v i n g -- A Case Study I n v o l v i n g the DENDRAL Program, i n Machine Intelligence 6, B. Meltzer and D. Michie (Ed.f, Edinburgh, U n i v e r s i t y P r e s s , 1971 Fillenbaum, S., On coping with ordered and unordered c o n j u n c t i v e sentences, J . Exp. Psychology, 87, 1971, 93-98 F i l l m o r e , C.J., The Case f o r Case, i n UniversaIs i n L i n g u i s t i c Theory, E. Bach and R. Harms (Ed.), New York Winston, 1968 Fodor, J . and M.F. G a r r e t t , Some R e f l e c t i o n s on Competence and Performance, i n P s y c h o l i n g u i s t i c Papers, J. Lyons and R. Wales (Ed.), Edinburgh, U n i v e r s i t y Press, 1975 Fowler, J . , LOTUS - A P r o c e d u r a l Environment, M.Sc. T h e s i s , Computer Science Department, U.B.C, 1976 F r a s e , L., S t r u c t u r a l a n a l y s i s of the knowledge t h a t r e s u l t s from t h i n k i n g about t e x t , J . E d u c a t i o n a l Psych., 60, 1969, Monograph F r e d e r i k s e n * C., E f f e c t s of context-induced p r o c e s s i n g o p e r a t i o n s on semantic i n f o r m a t i o n a c q u i r e d from d i s c o u r s e . C o g n i t i v e Psychology, 7, 1975, 139-166 , Semantic P r o c e s s i n q U n i t s i n Understandinq Text, i n Discourse P r o d u c t i o n and Comprehension, R. F r e e d l e (Ed.), H i l l s d a l e , ~Erlbaum7 1976 "~ ~ - 7 ~ Goffman, E. , Frame An§iZ§iss, Cambridge, Harvard U. P r e s s , 1974 G r i g g s , R.A., The R e c a l l of L i n g u i s t i c Ideas, J . Exp. Psychology, 103, 1974, 807-809 Grimes, J.E., Kinds of Information i n d i s c o u r s e , Kivu,ng, 4, 1971, 64-74 , O u t l i n e s and o v e r l a y s . Language, 48, 1972, 513-524 B i b l i o g r a p h y 138 , Thread o f D i s c o u r s e , The Hague, Mouton, 1975 H a l l i d a y , M.A.K., Notes on T r a n s i t i v i t y and Theme i n E n g l i s h I I , J o u r n a l of l i n g u i s t i c s , 3, 1969, 199-244 -, Language S t r u c t u r e and Language Fu n c t i o n , i n Sew Horizons i n L i n g u i s t i c s , J . Lyons (Ed.), London, P e l i c a n , 1970 H a l l i d a y , M.A.R. and R. Hasan, Cohesion i n E n g l i s h , London, Longman's, 1976 Hausenblas, K., On the c h a r a c t e r i z a t i o n and c l a s s i f i c a t i o n of d i s c o u r s e s , Trayaux L i n g u i s t i g u e s - de Prague, 1, 1 964, 67-83 Havens, W.S., Can Frames Solve the Chicken and Egg Problem?, Pr_oc. Fir . s t CSCSI/SCEIO latf.1 Conference, Vancouver, 1976 Hays, D., Dependency the o r y : A formalism and some o b s e r v a t i o n s , Language, 40, 196 4 H o l s t i , 0., Content A n a l y s i s of S o c i a l S c i e n c e s and H u m a n i t i e s , Reading, Addison-Wesley, 1969 ~ ~ ~ "~ Hunt, E. and S. P o l t r o c k , The Mechanics of Thought, i n Human Isf2ES§£i2S El22§ssing: T u t o r i a l s i n P e r f o r i a n c e and C o g n i t i o n , B. Kantowitz l i d . ) , Hillsdale7~Erlbaum7~1974 James, C., Theme and imagery i n the r e c a l l of a c t i v e and passive sentences, J . V e r b a l Learning and Verbal Behavior, 1, 1972, 205^211 J e r v i s , J . , An Augmented T r a n s i t i o n Network Grammar f o r E n g l i s h , Master's T h e s i s , Computer Science Department, U.B.C, 1974 Johnson-Laird, P.N., The P e r c e p t i o n and Memory of Sentences, i n fi§! Horizons i n L i n g u i s t i c s , J . Lyons (Ed.), London, P e l i c a n , 1970 Kaplan, R., Augmented t r a n s i t i o n networks as p s y c h o l o g i c a l models. A r t i f i c i a l I n t e l l i g e n c e , 3, 1972, 77-100 , On Process Models f o r Sentence A n a l y s i s , i n E x p l o r a t i o n s i& C o g n i t i o n , D. Norman and D. Rumelhart (Ed.), San F r a n c i s c o , Freeman, 1975 Kinneavy, J.L., A Theory of D i s c o u r s e , Englewood C l i f f s , Prentice-Hall7~1971 " K i n t s c h , W., Representation o f Meaning i n Memory, H i l l s d a l e , Erlbaum, ~ 9 7 4 B i b l i o g r a p h y 139 t Memory Representation of Text, i n Information P r o c e s s i n g and C o g n i t i o n , R. Solso (Ed.), H i l l s d a l e , Irlbauo7 1975 , Memory f o r Prose, i n The S t r u c t u r e of Human Memory, C.N. Cofer (Ed.), San F r a n c i s c o , Freeman, 1976 K i n t s c h , W. et a l , Composition and r e c a l l of t e x t as a f u n c t i o n of c o n t e x t u a l v a r i a b l e s , J . V e r b a l L e a r n i n g and V e r b a l Behavior, 14, 1975, 196-2*14 ~ ~~~ Kummer, W., O u t l i n e s of a model f o r a grammar o f d i s c o u r s e . P o e t i c s , 3, 1972 L a k o f f , G., S t r u c t u r a l complexity o f f a i r y t a l e s . The S t o r y of Man* 1» 1972, 128-150 Lees, R.B. and E.S. Klima, Rules f o r E n g l i s h p r o n o m i n a l i z a t i o n . Language, 39, 1963, 17-29 Lenat, D., BEINGS: Knowledge as I n t e r a c t i n g Experts, Proc. 4th I n t ' l J o i n t Conference a r t i f i c i a l I n t e l l i g e n c e , T b i l i s i , 19757 483-490 ~~ ~~ ~ Lesgold, A.M., P r o n o m i n a l i z a t i o n : a d e v i c e f o r u n i f y i n g sentences i n memory, J. V e r b a l L e a r n i n g and V e r b a l Behavior, 11, 1972, 316-326 :~ ~ ~ Lindsay, P. and D. Norman, Human Information. P r o c e s s i n g : An IS.iE2^S£ti2a t.1 Psychology, New York, Academic P r e s s , 1972 Longacre, R.E., Sentence s t r u c t u r e as a statement c a l c u l u s , Language, 46, 1970, 783-815 McCalla, G. and M. Kuttner, An e x t e n s i b l e feature-based p r o c e d u r a l guestion-answering system to handle the w r i t t e n s e c t i o n of the B.C. D r i v e r ' s Exam, CSCSI^SCEIO Newsletter, 2, 1975 McCalla, G., An Approach to the O r g a n i z a t i o n of Knowledge f o r a Model of C o n v e r s a t i o n , T h e s i s Progress Report, O.B.C., 1976 McDermott, D., A r t i f i c i a l i n t e l l i g e n c e meets n a t u r a l s t u p i d i t y , SIGART Newsletter, 57, 1976, 4-10 Marcus, M., Wait and See S t r a t e g i e s f o r P a r s i n g N a t u r a l ItaiiaJii.at» working paper 75, M.I.T. A.I. Lab, 1974 , Diagnosis as a Notion of Grammar, i n T h e o r e t i c a l Issues i n N a t u r a l Language P r o c e s s i n g . R. Schank and B. Nash-Webber l i d . ) , Proceedings o f the conference, Cambridge, 1975 B i b l i o g r a p h y 140 Barks, L.E. and M i l l e r , G., The r o l e of semantic and s y n t a c t i c c o n s t r a i n t s i n the memorization of E n g l i s h s e n t s n c e s , J . V e r b a l L e a r n i n g and V e r b a l Behavior, 3, 1964, 1-5 Meyer, B.J.F., The O r g a n i z a t i o n o f Prose and. i t s E f f e c t s on Memory, &msterdam7~North-Holland7 1975 M i l l e r , G., The magical number 7, p i n s or minus 2, P s y c h o l o g i c a l Review, 63, 1956a, 81-97 , Human memory and the s t o r a g e of i n f o r m a t i o n , I.R.E. Trans, on I n f o r m a t i o n Theory, IT-2, 1956b, 129-137*" , E n g l i s h Verbs of Motion: A Case Study i n Semantics and L e x i c a l Memory, i n Coding Processes i n Human Memory, a. Melton and E. Martin ~(Ed.}, Washington,"winston,~ 1972 M i l l e r , G. and L.A. McKean, A chronometric study of some r e l a t i o n s between sentences, Quart. J . Exp. Psygh,, 16, 1964, 297-308 Minsky, M., A Framework f o r Representing Knowledge, AIH-206, M.I.T. A.I. Lab7~1974 Moran, T.P., The Symbolic Imagery Hypothesis, Computer Science Department, Carnegie-Mellon O n i v e r i s t y , 1973 M o r r i s , C.W., S i g n s , Language, and Behavior, Englewood C l i f f s , P r e n t i c e - H a l l , 1946 Newell, A., P r o d u c t i o n Systems: Models of C o n t r o l S t r u c t u r e , i n U s u a l Information P r o c e s s i n g , w. Chase (Ed.), Washington, Winston, 1973 Norman, D., Memory and A t t e n t i o n , New York, Wiley, 1976 Norman, D. and D. Rumelhart (Ed.), E x p l o r a t i o n s i n C o g n i t i o n , San F r a n c i s c o , Freeman, 1975 Olson, D., Language and thought — a s p e c t s of a c o g n i t i v e theory of memory. Psych. Review, 77, 1970, 257-273 Palmer, S., V i s u a l P e r c e p t i o n and World Knowledge, i n IS£l2E§£i2M i l C o g n i t i o n , D. Norman and D. Rumelhart (Ed.), San F r a n c i s c o , Freeman, 1975 Paxton, W. and A, Robinson, system i n t e g r a t i o n and c o n t r o l i n a speech understanding system, Amer. J . Qomputat.io.aal ii£2fii§ti£§* 1975, mf 34 P e t r i c k , S., On N a t u r a l Language Based Query Systems, Res. Rpt. #5577, I.B.M7~T7j7 Watson,~1975 B i b l i o g r a p h y 141 P h i l l i p s , B., Judging the coherency o f d i s c o u r s e , American J . Computational L i n g u i s t i c s , 1975, mf 35 Post, E., Formal r e d u c t i o n s o f the g e n e r a l c o m b i n a t o r i a l problem, Ameri J : Hath., 65, 1943, 197-268 Propp, v . , Morphology o f the F o l k t a l e , A u s t i n , 0. of Texas P r e s s , 1968 Q u i l l i a n , M.S.-, The tea c h a b l e language comprehender: a s i m u l a t i o n program and a theory of language, Comm. A. CM., 12, 1969, 459-476 Seddy,• R., L. Erman, and Neely, A model and system f o r machine r e c o g n i t i o n of speech, I.E.E.E. Trans. Audio §.34 E l e c t r o a c o u s t i c s , QU-217 19737 229-238 Seddy, R. and A. Newell, Knowledge and I t s Representation i n a Speech Understanding System, i n Knowledge and C o g n i t i o n , Gregg (Ed.), Potomac, Erlbaum, 1974 Rieger, C.J., Conceptual Memory and In f e r e n c e , i n Conceptual ISf2£iation P r o c e s s i n g , R. Schank (Ed.), Amsterdam, North-Holland, 1975 Riesbeck, C , Computational Understanding: A n a l y s i s of Sentences i n c o n t e x t , Ph.D. t h e s i s , computer Science Department, St a n f o r d U n i v e r s i t y , 1974 , Conceptual A n a l y s i s , i n Conceptual Information P r o c e s s i n g , Ri Schank (Ed.), Amsterdam, North-Holland, 1975 Robinson, J . , A tun e a b l e performance grammar, Amar. J. Computation L i n g u i s t i c s , 1975, mf 34 Rodgers, P., A d i s c o u r s e - d e n t r e d t h e t o r i c of the paragraph. C o l l e g e Composition and Communication, 17, 1965, 2-11 Rosenschein, S., How does a system know when t o stop i n f e r e n c i n g ? , Amer. J . Computation L i n g u i s t i c s , 1975, mf 36 Rowat, P., Of C r a y f i s h and Production Nets, Proc. F i r s t CSCSI^SCEIO Natf.1 Conference. V a n c o u v e r " 1*976 Rubin, A. , Grammar f o r the People: Flowcharts o f SHRDLIP s Grammar7 ATM-282, M.I.T7 ~A7I. Lab"7~1973~ Rumelhart, D., Notes on a Schema f o r S t o r i e s , i n Rep r e s e n t a t i o n and Understanding, D. Bobrow and A. C o l l i n s ^ I d .), New York, Academic P r e s s , 1975 B i b l i o g r a p h y 142 Rumelhart, D,, P. L i n d s a y , and D. Norman, A Process Model f o r Long-term Memory, i n The O r g a n i z a t i o n of Memory, E. T u l v i n g and W. Donaldson (Ed,), New York, Academic Press, 1972 Sachs, J . , Recognition memory f o r s y n t a c t i c and semantic aspects of connected d i s c o u r s e , P e r c p e t i o n and Psych^oghysics, 2, 1967, 437-442 Sanders, Gerald A., On the n a t u r a l domain of grammar. L i n g u i s t i c s , 63, 1970, 51-123 S c a r l , E., A F l e x i b l e ATN, a F l e x i b l e Morpher, i n f o r m a l t a l k , F i r s t CSCSI/SCEIO N a t ' l Conference, Vancouver, 1976 S c h a c t e r , P., Focus and r e l a t i v i z a t i o n . Language, 49, 1973, 19-45 Schank, R., I d e n t i f i c a t i o n of the C o n c e p t u a l i z a t i o n s U n d e r l y i n g N a t u r a l Language, i n Computer Models of thought §.nd Language, R. Schank and K.I. Colby (Ed.), San F r a n c i s c o , Freeman7 1973 t C a u s a l i t y and Reasoning, TR-1, I.S.S.C., 1974a t Understanding Paragraphs, TR-6, I.S.S.C., 1974b , Representation and Understanding o f Text, NATO ASI, Santa Cruz, 1975a (Ed.), Conceptual Information P r o c e s s i n g , Amsterdam, North-Holland7 1975b Scragg, G., Frames, Planes and Nets: A. S y n t h e s i s , TR 19, I.S.S.C7, ?976 Shannon, C., P r e d i c t i o n and entropy o f p r i n t e d E n g l i s h , B e l l Sys. Tech. J . , 30, 1951, 50-64 Simmons, R., Semantic Nets: T h e i r Computation and Use f o r Understanding E n g l i s h Sentences, i n Computer Models o f Thought and Language, R. Schank and K.M. Colby l i d . ) , San F r a n c i s c o , Freeman, 1973 Simon, H., How b i g i s a chunk?. S c i e n c e , 193, 1974, 482-488 S u l i n , R. and D. D o o l i n g , I n t r u s i o n o f a thematic idea i n r e t e n t i o n of prose, J . Exp. Psych., 103, 1974, 255-263 T a y l o r , B.H. and R.S. Rosenberg, A c a s e - d r i v e n parser f o r n a t u r a l language, Amer. J . Computational L i n g u i s t i c s , 1975, mf 31 B i b l i o g r a p h y 143 T e s l e r , L., H. Enea, and D. Smith, The LISP70 Pattern Hatcher, Proc. T h i r d l n t * l J o i n t Conference A r t i f i c i a l I n t e l l i g e n c e , T b i l i s i , 1975 T u l v i n g , E. and W. Donaldson (Ed.), The O r g a n i z a t i o n o f Memory, New York, Academic P r e s s , 1972 de V i l l i e r s , P.A., Imagery and theme i n r e c a l l of connected d i s c o u r s e , J . Exp. Psych., 103, 1974, 263-268 Waterman, D., Adaptive Production Systems, Complex I n f o r m a t i o n P r o c e s s i n g Working Paper No. 285, Carnegie-Mellon U n i v e r s i t y , 1974 Wilcox, B. and C. Hafer, LISP/HTS User Guide, Mental Health Research I n s t i t u t e , Ann Arbour, ?974 Wilks, Y., Preference Semantics, AIM-206, Stanford U n i v e r s i t y , 1974 ~ " ~ Winograd, T., Procedures as a Representation f o r Data i n a Computer Program f o r Understanding N a t u r a l Language, MAC-TR-84,~M. 17-T. ~, 1970 , Breaking the complexity b a r r i e r : again, SIGPL4N N o t i c e s , 10(1), Jan. 1975, 13-22 , Frame Representations and the D e c l a r a t i v e / P r o c e d u r a l Controversy, i n Repres e n t a t i o n and Understanding, D. Bobrow and A. C o l l i n s (Ed.), New York, Academic Press, 1975 Woods, W.A., P r o c e d u r a l Semantics f o r a Question-Answering Machine, Proc. AFIPS 1968 F.J.C.C., 1968, 457-471 Woods, W.A., T r a n s i t i o n Network Grammars f o r N a t u r a l Language A n a l y s i s , Comm. A.C.M., 13, 1970, 591-606 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0051779/manifest

Comment

Related Items