File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/79/j79-1040_metho.xml
Size: 107,887 bytes
Last Modified: 2025-10-06 14:11:09
<?xml version="1.0" standalone="yes"?> <Paper uid="J79-1040"> <Title>Contents I ntxoduction Winograd's Understanding System Some discussion of SHRDLU Some more general background issucs Second generation systems Some comparisons and contrasts Conclusion References</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> NATURAL LANGUAGE UNDERSTAND I NG SYSTEMS WITWIN THE An I1 PARADIGM: A SURVEY AND SOME COMPARISONS YORICK WILKS DEPARTMENT OF ARTIFICIAL INTELLIGENCE UNIVERSITY OF EDINBURGH EDINBURGH EH8 9NW </SectionTitle> <Paragraph position="0"> Revised version of Stanford Artificial Intelligence Laboratory Memorandum 237, supported by contract number NIE-P-75-0026 with the National Institute of Education.</Paragraph> <Paragraph position="1"> Zoptp-ight 1976 by the Association for Computational Linguistics ABS'TIWCT: The paper surveys the najcr prcjecte on th.e understanding of natural language that fall within what nay now 5e calle3 t5e artificial intelligence paradigm fcr natural Language syste~s.</Paragraph> <Paragraph position="2"> c4z,'2e null space is devoted ta arguing that the &paradigm is new a red: itv and different in significant respects from the generabive paradiLm ~f present day linguistics. The caalparison's between systems tekt3-c round questions aL.-ut UIC lc~~l, cc~ltralitb* and ~hancnt-n~1c7giial plnusik~ility&quot; of the kn~~wlt~~l~.~~ and inft-rt.nct.s t1.1.1t- must .rv.~~l,~t-'Ltto a systtlm that is to uurldt*rst,~nd ~-\*t-r>*~l.~y ls~lyu~~gr? .</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> References </SectionTitle> <Paragraph position="0"> in his report to the Science Res~arch Council, on the stato aE Artificial Intelligence, Sir Suer, Lighthi11 (1973) qaw &oat cf tho field a eather bad propnosis. One of the few hopful igns ho am war Winwredls (1972) natural language wderstanding system. Yet, only e ye-car later, Winograd had st~ppd wrk on ae system he mnstruuted, and had kwgurr A nnlar one on entirslx different principles.** He went SJ far, in a survey lecture (Winograd '737 of extrnordinnry modesty Ln a field not krtawtl for its mall cwputer systerns designed to understand natural languaga, and \rent on tc describe others second generation' systems.</Paragraph> <Paragraph position="1"> I shall xeturn later to this metaphor of generations, but what is cne to say in general terns of a field where yesterday's brightest spots are today's first generation systems, even though they have not been criticised in print, nor shown in any generally acceptable yay to be fundamentally wrong? Part of the answer lies in the profound role of fashion in Artificial Intelligence in its present pre-scientific phase. A cynital American professor remarked recently that Artificial ~ntelligencd (AX) had an affair with someonels work every year or two, and that, just as there were no reasons for galling in love, so, later, there were no reasons for falling out again. In tho csse of Winograd's work it is imprtant now to resist this fashiony and re-emphasize what a good piece of research it was, as 3 shall inl a m~ment.</Paragraph> <Paragraph position="2"> Another part of the answer lies ,in the still fundamental role of - metaphysical criticism in AI. In the field of computer vision things are bad enough, in that anybody who can - see feels entitled to criticise a system, on the ground that he is sure - he does not see using such and suck principles, In the field of natural language understanding things are worse: not only does anyone whmo can speak and write feel free to criticise on the correspandina grounds, but in addition theze are those trained in disciplines parasitic upon natural language, linguists and logicians, who often know in addition how things bIUST BE DONE on a priori groundsa. It is this metdphysical aspect of the suhject that gives its disputes their characteristically acrbnious</Paragraph> <Paragraph position="4"> what are swre of the outstanding disputes and how testable are the claims being made.</Paragraph> <Paragraph position="5"> If what follows seems unduly philosophical, it should be remetsabered that Uttle - is agreed, and almost no achievements are beyond question. To pretend otherwise, by concentrating only tm the detaibof established programs, ~uld be meretricious and misleading.</Paragraph> <Paragraph position="6"> To euntey an enewetic field like thie one is inevitably to laavo a great deal of excellent work unextiminad, at least if one ia going to do more than give a paragraph to each research project. I have left out of cotasideration at least seven groups of projects: (1) Early work in Artificial Intelligence and Natural Language that has been sumeyed by Wfnograd (1973) and Simmons (1970a) among others.</Paragraph> <Paragraph position="7"> (2) Work by graduate students of, or intellectually dependent upsn that of, people discussed in same detail here.</Paragraph> <Paragraph position="8"> (3) Wxk that derives essentially from projects described in detail here. This embraces several groups interested in testing psychological hypotheses, as re11 as others constructing large-scale systems for speech recognition. I have devoted no space to speech recognition as such here, for it seems to me to depend upon the quality of semantic and inferential understanding as much as anything, and so I have concentrated upon this more fundamental task.</Paragraph> <Paragraph position="9"> (4) Work on language generators, as opposed to analysers and understanders. null They are essential for obtaining any testable output, but are thearetically secondary.</Paragraph> <Paragraph position="10"> (5) All the many and varied reasoning schemzs now availdle in AI, hcluding PLANNER (Hewitt 19691, QA4 fRulifson et a1 19721, MERLIb (*ore and Newell 197a as well as automatic programming (Balzer et a1 1974) (tleidor- 174) and debugging (Sussman 1974) projects, many of which are producing formalisas that appear increasingly like natural lwuaae .</Paragraph> <Paragraph position="11"> (6) Conservative reasoning schames, such as first order predcicatc! calculus, thab have boen appliud ts, or ild~~~uatsrf for, the aqalysis of natural language: tblma~thy and Hayes L?>t.?) tCalrr?n 1972) (Sandewall 1972) .</Paragraph> <Paragraph position="12"> (71 X have also ignored, as one musk in oldor to write at all fro a rapidly changing field, Uaa interpr,etntion given t~ d4inskjw' a (1975) notian bf l&quot;'frmelq auring 1975 by Chsmiak and ScEaadc. During this year Lmth have product4 skctch~a Eer a rrprasantaticn UP knuwladqrs sn a larger scallr than any 3ls;usss.l ln this survey: roughly speaking, they have each produced a schmd Pok a whole story. The value and function of sucll a larger-scale representation is much under discussion at the manent, bdt acne of it invalidates what! is contained here, except for sane qualification to the position of Schank which is noted under the heading &quot;Centrality&quot;' in section 6.</Paragraph> <Paragraph position="13"> The exclusions under (2) above are particularly unfair to the wark in the unpublished theses of Rieger 11974) and McDesmott (1974), and can only be justified, like those of sections (5) and (6) above, by space, bias and the considerations advanced in a later section of this survey concerning what it is for A1 research to be nlmut natural lmquaqe, ra'tht3r than slwut. sanething else, like human psych~l~~~y, or logic, as the bargaining hhaviour of chil-dsen under stress.</Paragraph> <Paragraph position="14"> One final remark &wut the style of this survey. I have not indulged the n~rmal critical luxury of careful and extensive quotation from authoys followed by detailed and meticulous criticism. That would be wholly inappropriate in the present state of things. A great deal of work is being done at the moment, and many of the principal researchers change their views on very fundamental questions between one paper and the next without drawiw any attention to the fact. Cheap self-contradictions and changes of mind are all too easy to find, so criticism and smparisons are best drawn with a very broad brush and a light stroke.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2. Winograd's Understanding System </SectionTitle> <Paragraph position="0"> Winograd's systea runs as a dialogue, in real time, between a human operator who types in meSqes, and the system proper, called SHRDLU, which displays on a screen pictures of a closed world of coloured blocks and pyraida, a box in- which they can be put, and an arm or crane khat can move che objects about.</Paragraph> <Paragraph position="1"> SHRDLU displays its replies to the operator by writing on the samq screen.</Paragraph> <Paragraph position="2"> The following is a discontinuous sample of the dialogue Winograd giws in his book. SHIIDLU1s replies are in the upper case r eick up the big red block,</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> OK </SectionTitle> <Paragraph position="0"> find a block which is taller than the one you are holding and put it into the box.</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> BY IIT', I ASSUME: YOU MEAN TKE BLOCK WHICH IS TALfwER THRN Tm ONE I AM HOLDING </SectionTitle> <Paragraph position="0"> what does the box contain THE BLUE PYMID RND THE BLUE BLOCK had you touched any pyramid before you put the green one on the Lkttle cube YES, TEE GREEN ONE when did you pick it up WEIILE.1 WAS STACKING tlP THE RED CUBE, A LARGE RED BLOCK AND A</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> EARGE WEN CUBE. why TO GET RID OF IT. </SectionTitle> <Paragraph position="0"> Thia, ingenious program was also ahle to accept new definitions.& words, such as steeple'; to make use of what it knew about the p~operties of the objects described, such as that a block could have its top cleared off, but a pyramid could not; and also ta remember what it had done before, as in the sample above.</Paragraph> <Paragraph position="1"> The syn-tactic analySis program was written in PR%RAMMAR, a procedural language related to PLANNER (see Hewitt '69). This means that a familia phase structure rule such as s+NP+w (to be interpreted: a sentence consists of a noun phrase followed by a verb phrase) would be expressed as:</Paragraph> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> (PDEFIME SENTENCE ((( PARSE NP) NIL FAIL) (( PAFSE VP) FAIL FAIL RETURN))) </SectionTitle> <Paragraph position="0"> The details of the notation need not detairius at this point; what Ls important is that Winograd's gramuax 4s not tile cmnventirrnaL list of culas, but small sub-programs Like tha lines above, that actually xaysooant x-o~iiures for iulposing the desirad grstmalieA1 structure, The first leva1 of linguistic pm&ur~s in the system applies a systemic grammar' , dur to M.A. K. Mallidal* (1970) , which inapses a hierarchical structure of clauses on tihe input senterac&s\, which secb tc' b~ dram from a vocabulary of about 175 ims\ls.</Paragraph> <Paragraph position="1"> Winogradas parsing is top down, and depth first, with na automatic back up. The parsing progrim fur each griuxnatical catcq~r~~ is n functional definition in PRfXZWWR, which can be stated either as &VP, fcr SEWPEHCE, or as a flow-chart as below for VP: Here is Winograd's own account of the start of this top-down parsing procedure for the sentence &quot;Pick up a red block&quot; (where the material in C 1 is added explanation and not Winogradas om) : &quot;The CLRUSE program looks at the first wrd, to decide what writ the CLRUSE begins with.</Paragraph> <Paragraph position="2"> If it sees an adverb, it assumes the sentence begins with a single-word modifier tslowly, Jack lifted the book] ; iP it sres a preposition, it looks .for an initial PREP6 Con top of the hill stood a tree] If it sees a BINOBR, it calls the CLAUSE program to look for a BOUND CLAUSE C~efore you get there, we left].</Paragraph> <Paragraph position="3"> In English (and possibly all languages) the firsb word of a construction often gives a very goad clue as to what that construction will be. fn this case, &quot;pick&quot; is a verb, and indicates that we may have an IHPERATIW CLAUSE. The program. starts the VO program witla the initial VG feature. list (VG IWER), looking for a VG of this type. This must either begin with some form of the verb &quot;do&quot; [Do not call me!] or with the main verb itself [Call me!]. Since the next word is not: &quot;do&quot; it checks the next word in the input (in U~is case still the first word) :o see whether it is the infinitive form of a verb. If SO, it is to be attached to the parsing tree, and given the additionql feature MVB (main verb). The current structure can be diagramad as:</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> (CLAUSE MAJOR) (VG IWER) (VB MVB INF TRANS ,WRT -------------- pick TWS AND VPRT cmfw fran the definition of the word &quot;pick&quot; when we zalled </SectionTitle> <Paragraph position="0"> the function PMSq for a word.4' Mter this syntactic parsing, a number of &quot;semantic specidlists'' attach cwbantic structures to specific syntactic dnes. A semantic definition of an In the case of &quot;a red cube&quot;, the follo~ing structure is built up by an</Paragraph> </Section> <Section position="9" start_page="0" end_page="0" type="metho"> <SectionTitle> NP &quot;semantic specialist&quot; (GOAL (IS ?X BLOCK)) (GOAL (COLOR ?X RED))_ (EQDIM ?x)-------------------~--- PLANNER description (BLOCK MANIP E%YSOB THING)---------- markers </SectionTitle> <Paragraph position="0"> The first three lines willeventual~y form me bulk of a Micro-Planner progr- which, when evaluated will seek an object X that is a block, is equidimenoPona1 (EQDIM) and is red (where &quot;red&quot; itself has a definition, the system that restricts its application to objects with the feature PHYSOB) The last line of the ftgure is a set of 'semantic features&quot; read off right to left from the following feature tree&quot; The semantic structure of &quot;the red kuheJ&quot;ccan be used by C:?e deductive</Paragraph> </Section> <Section position="10" start_page="0" end_page="0" type="metho"> <SectionTitle> &quot;&WING --- </SectionTitle> <Paragraph position="0"> coapnent of the system, hefore eb*aluatfon resultAng in the actual picking up, to see if such an object 1s JE it wers IIL.~, (an '&quot;~y~idinestsiannl pyramid1&quot; would not be) the system souLd yo dncl try ta re-pitrse Nrn the sentcnco.</Paragraph> <Paragraph position="1"> The meaning of verbs in SHRDLU is mre mmplex. The seaantic c~wpnent has access to n definition fer &quot;pick-up&quot; just as it d~es for &quot;tt5d1' a:~d &quot;klcck&quot; and this definition will enable SHRDLU to translate &quot;pick-up tateaents&quot; into Micro-planner in a mdnner analogous to that for noun phrases.</Paragraph> <Paragraph position="2"> These are two complications here. Firstly &quot;pick-up&quot;, unlike &quot;red&quot;, is</Paragraph> </Section> <Section position="11" start_page="0" end_page="0" type="metho"> <SectionTitle> PLACE PROPERTY -- - </SectionTitle> <Paragraph position="0"> : defined in terms of other concepts in the system: in particular, in terms of GR?SP and RAISEHAND, which axe two of the three basic actiQns in the system. Secondly, there are two rrpes of verb definition, semantic and inferential Winograd does not give the semantic definition for &quot;pick-up&quot;, but here is the one for &quot;grasp&quot; which is a closelb- related verb.</Paragraph> <Paragraph position="1"> which says e~sentiallp that grasping is scmathinq dorle by an animate entity to a oanlpulabla one (flret line).</Paragraph> <Paragraph position="2"> More of the real content of such actions is found in their inferential definition. Here is the one fox '&quot;pick-up&quot;:</Paragraph> <Paragraph position="4"> TSlis definition allows the program to actually carry out the &quot;pick-up&quot; caamand if it is possible to do so in the simulated world, as it would not be, for example, if -re were already a block dn top of the red one..</Paragraph> <Paragraph position="5"> PICKUP is being defiled in edzrms of a number of more primitive s&-actions, such as GRASP and RAISEHAND each of which must be carried out in order that sameming may indeed be picked up. There sub-actio~ themselves have infesential definitions: the one given for GRASP, for example, is somewhat differant from its &quot;C1(&EANS1' definition given above, although the inferential iPefini,tiona are aim, in sane se-e, definitions of meaning as wall as progra~as for actually carrying out the associated conmands.</Paragraph> <Paragraph position="6"> One reason for the enormous impdct of this work was that, prior to its appearance, A1 work was not very linguistically interesting, while the eystems of tho linguists had no place far the use of inference and real world knowledge. Thus a very limited union between the two techniques was able to breed considerable results. Before Winograd &ere were few pmgrams in A1 that could take a reasonable complex English sentace and ascribe any structure whatever to it. In early classics of 'ngtural language understanding' in AX, such as Bobrow's STUDENT (1968) problem solver for simple algebra, input sentences had to be short and of stereotyped form, such as &quot;what is the sum of .... ?&quot; Conversely, in linguistics, there was, until very recently, little speculation on how we understand the reference of. pronouns in such eleqentary sentences as &quot;The soldiers fired at the woslen.and I saw several fall&quot;, where it is clear that the answer is both definite, and that fhding it requires some inferential manipulation of genaralirrtiens dwut the world.</Paragraph> <Paragraph position="7"> 3ha reader should ask himself at this jmint how he kwwe the nsrxact refemnt of the pmnaun in that sentanca.</Paragraph> <Paragraph position="8"> 3. $me Discussion of SWLU S~J -fax, the reaction ts Winugqad \s work has Lmtk wbSrSrla ~ineri t heal. What would crAitics find to attack if tthtzp M~SF 90 a9nde.I: Firstly, that ~~inegrad's linguistic system is highly wnsr~~?atiw, and that UIQ distinction between 'syntax' and 'semantics' my not Lw necessary at all. Sewniily #at his semantics is tied tu the shple referential wrld cf the blccks it? a way *.at muld make it fnextensihle, to any genexal, real tmrld, situatioh. Suppose 'block' were allowed mean 'an obstrxxctidnl and 'a nental ir,k,i*ition', as well as 'a cubic object'. It is dauktful whether Winograd's features and rules could express the ambiguity, and, nore im~rtantly, whether the simple structues he manipulated could decide correctly between the alternative meanings in any given c~ntext of user Again, far more sophisticated and systematic case structures than those hg used might be needed to resolve the ambiguity of 'in1 in &quot;He ran the mile in five minutes and Ire ran the milt. in a pawr kwg , as we11 as 'tllu ccmbination ~,lf case with wrd sense amhipity in 'Ho put tho key in the l~rk' (door lock1 and 'Be threw the key in the luck1 (river leek).</Paragraph> <Paragraph position="9"> The blocks mrld is also strongly deduct-ive and lwical1y closed, Sf gravity were introduced into it, then anything supkwrted that war pushed in a certain way would have, logically have, to fall. But the cazmon sense wlorld, of ordinary language, is not like that: in th~ 'waaen and soldiers' example given earlier, .the pronoun 'several1 can be said to be resolved using same generalisation such as 'things shot at and hurt tend to fall' There are no logical 'have to's1 there, even though the meaning of the ~ronoun is perfectly definite.</Paragraph> <Paragraph position="10"> Indeed, it might be argued that, in a sense, and as r~afis its semantics, Winograd's system is not about natural language at all, but abut the - technical question of how goals and sub-goals are to be crganised in a problem-solving system capable of mani~ulating simple physical ehjects.</Paragraph> <Paragraph position="11"> If n reeqber, for example, that tfre key problem that brought bwn the emrr#rur work on mchiae translation in the Fiftiea and Sixties, was that of the eensa diguity of nattWi.l. language wprds, then we will look in vain to SHRDLU Lor any help with that problem. There seem? to be only one dear exmaple of aur aLnbiguous mrd in the whole system, namely that of 'curitairat as it appears in 'The box contains a red block' and rhe stack Again, if on@ glahces back at the definition of 'pick-up' quoted &ve, ana can see t)cdt it. ie in fact an e%pression rrE a prcrcedure f~r picking up an ob5ect in the SHR~LU .yet=. Nothing about it, for ekample, would help one, understand thep~rfe~tlv ordinary sant;ence 'I picked up my bags PStom the plstforn, and ran for the train', let alone any sentenco not &out a physical action performable by the hearer. One could put the point so: what wa are gFveR in the PLRNFtE3I code is.not a sense of 'pick up1 but an exardple of.its use, just as 'John picked up the volunteer from the aueience by leaning over the edpeeog the stage and WWW her up by means of a rope clenched in his teeth1 is not so much a sense of ,the verb as a use of it.</Paragraph> <Paragraph position="12"> nose who like very general analogies may have.noticgd that Wittgenstain (1953 para. 2ff) devated considerable space to the construction of an e1sentaz-y language of blocks, heaxus and slabs; one postulated on the as tipn that the words of language were basically, as is supposed in modal theory, the names of items. But, as he showed of the enterprise, .nd to the eatisfaction of aav readers, &quot;That philosaphicat concer~t cf, meaning (i.e. af words as the unernbiyuous names uf physical ~bjects---W) has its place in a primitive idea of the way language funotions.</Paragraph> <Paragraph position="13"> But one ca~ aleo say that it is the idea of a lanyuaye more primitive than ougs&quot;. (my italics).</Paragraph> <Paragraph position="14"> To all this, it might be countered that -it has not been shown that the language facilities I have descrlb&d cmt be incorporated in the structures that SHRDLU manipulates. and-th%tt, even if they cwld not, the work muld still be significant in virtue of its orisinal control stzuctura and its demonstration that rpal world knovledge am bv uerged-with linguistic knowledge in a working Whole. Indead. a13mug.1 Winograd has apt tried, in any straightforward sense, tc extend SUEU~LU system one could say that an extension' of this -st is kmir~g stbaptd*$y (am43 wit5 his 'Mliever Systwm' which is a hybrid system -t .hit beliefs that is, in the aanee nf saction 4 blow ''.+mad pouatlmt,, a base analyser Lrolln Bruce ' a ClUWWS amtem (1971) rhlch is e rl+zo r~rw ---late first genesat~un---systPmp in km soaw wnvr & Wbtq~d*u. Otbch in the last catwry that should be- mtentAwm3 Blaavhw bmd fsmdgs'419'?2) e~p3sration of kb csnwpts of luustl ahd ,,cmi8dq kn r mt-ldl & tic-tac-to!&, and Joshils anfcnsion ob it (1973), kt all tbm ra~tp~lt am3 inPluentia2 wxk uf Oh.xx3rn (lQ72).</Paragraph> <Paragraph position="15"> This wrk, mot recsz~thy apylYd to a r&cm-world d ltme~ SOCBt mm, is not discussed in the detail it deserves in this papex. '%be bwd on an augwnted state transition nebork gs-, is udmabtmdly of 't;ba rxaeast mbust in actual use, in that it is less s,ensiUve to 9.4hRTICmS&t input guestions it encounters thdn its rivals. 'Thq reasam fir tmdtbg it in depth is that bath Was and Wimq~ad 'nave an~u.ed 3n print tht thsLs twlo systas are essentially eqiivaLent [Wisrsgrad 1971) 19731, ad so, if they are right, there is no need Its dise:ss bth, and Wixmgxwil's is, within the WI camiaunity at least, the better known of the two, Their ec&ivalenc,e arguments are proLdly mrxect: bth ur gr-based deductive systems, operating within a guestim-anmriqq b~~i-f in a highly limited daain uf disccurse. Wixqsad'3 syst- QX h$aw aa hou to psoceed, within his P-PM gramas, is, as he hhsehf +pints mt, fomallj. equivalent t~ an augmented state transition network, M in prtit vlar tc~ the ordering of cbictes at ndes in PIJodsl system, There is a significant diffexencfe 2n theis mtaphysicaf -, presuppositions about meaning which, howevex, has m inf flu err ace ca the aamL operation of their respective systea~. This diffeme is disguised by the allegiance both givd ta a 'prolcedwal view of meaaiq' T'k difft!zeace is that Molods takes a much more logico-sanantic intqretatiua of that slogan than does Winoqrad. In partisular, fbr Ws the meaniq of an Lnp~t utterance to his system is the procedures within the system that raJ+ipulata tfr truth conditions of the utterancfe and estgblish its truth value.</Paragraph> <Paragraph position="16"> To put the mttex crudely, for Moods an assertion has xi meanhq if hie. system cannot etabligh its tnth or falsity. Wbuqxad has mrtabdy camitted himself to my such extreme~positiom.</Paragraph> <Paragraph position="17"> It is interesting 'to notice that Woods' is, in virtue of his strong position an truth conditions, probably the only piece of work in the ofi~ld of A1 and natural language to satiefy Hayes' (1974) recent demand* that to be 'Lntellectually respectable' a knowledge syptem must have natural model thwretic semantics, in Tarski's senso. Since no-one has over given prec3se tzuth conditions for any interesting piece of discourse, such as, say, Woods1 arm papers, one might claim that his theoretical presupposittons necessarily limit his work to the analysis of micro-worlds (as distinct from everyday language) . However, if Woods ' ' internal ' interpretation of tho 'meanings are procedurest slogah has certain drawbacks, so too does Winograd's, or what one might call the 'external' interpretation. By Ghat I mean Winograd'g concentration on actions, like picking up, that are in fact real world procedures, and ir) acway that tl~e meanings of 'concentrate', 'call', 'have', 'intwpret1, etc, are - not self-evidently rep1 world procedures that we could Set out in PLANNER for a robt. Of course, Ninograd 4s free to concentrate on any micro-world he wishes, and all I am drawing attention to here is thd danger of assuming that natural language is nomaiky about real world procedures and, worse still, the implicit making bf the assumption that we cannot understand discourse about a procedure unless we can do it ourselves. I aa not saying that Winograd is making this evidently false assumption, only that the rhetoric surrounding the application 05 the !meanings are procedures1 sloglsln tx~ his system my cause the unwary to do so.</Paragraph> <Paragraph position="18"> There is quite a different and low-level problem about the equivalence of Woods' and Winograd's systems, if we consider what we might call the received co-n-sense view of their work, Consider the following three assertions: (1) s system is an implementation of a transformational grammar (2) Winograd's work has shown the irrelevance of transformational grammar for language analysis - a'view widely held by reviewers of his work.</Paragraph> <Paragraph position="19"> * a view modified in Hayes (1975) where it now seems that programs/procdures would serve as a 'semantics' instead (a quite different, and more reasonable, position, of course).</Paragraph> <Paragraph position="20"> (3) Woods' and Winograd's systems are formally equivalent - a view held by both of them.</Paragraph> <Paragraph position="21"> There is clearly swathing of an inconsistent triad anronyst thosr three widely held breliaFs. T31a txouble probably centxus Gn the r~xact sense which Woods1 nark is formally equivalent to a transf~mational~ graaraslar - not a question that ned detain us here, but one worth plntfng out in passing</Paragraph> </Section> <Section position="12" start_page="0" end_page="0" type="metho"> <SectionTitle> 4. Scwe More General Uckgraund IasurfL </SectionTitle> <Paragraph position="0"> Winograd's hvrk is a central rrxcmplr? of the 'Artificial fnl~lligr?nc*c paradigm of h.ngu&ge', using '~aradicp' in Kuhn's (10701 sense of a 1arqe scale revision in systematic thinking, where the,pa.radigru revise is ,the 'generati- paradigm1 of the Chmskyan linguists fChomsky 1957). Fra the A1 pintof view, the generative linguistic tmrk of the last fifteen years has three principal defects. Firstly, the generation of sentences, with whatever attached structures, is not in anv interesting sense a demonstration of human understanding, nor is the separation of khe well-formed from the ill-fomed, by such methods- for understanding requires, at the very least, bth the generation & sentences as parts of coherent discourse and some attampt to interpret, rather than qer5ly zeject, what seen to be ill-farmed utterances. Neither the transformational grammarians following Chomsky, nor their successors the generative semanticists (Lak-off 1971), have ever-eSplicitAy rknounced thd generative paradigm.</Paragraph> <Paragraph position="1"> Secondly, Chmsky's distinction between pexfomance and competence models, and his advocacy of the latter, have isolated modern generative linguistics from any effecti- - test of the systems of ruhes it proposes.</Paragraph> <Paragraph position="2"> Whether or not me distinction was intended to hdve this effect, it has meant that ay test sxtuation necessarily involves performance, which is wnaLdered vutsfde the province of serious linguistic stugy.</Paragraph> <Paragraph position="3"> And any embdiment of a svstein of rules in a computer, and assessment of its out put, would be perf~rmance. AI, too, is much concerned with the structure of linguistic processes, independent of any particular implementation,** ** Vide: &quot;Artificial Intellige~ce is the s.tudy of intellectual mechanisms apart from applications and apart fra how such mechanisms are realised in the human or in animals. '' (McCarthy 1974) but implementation is never excluded, as it is from competence models, but rather encouraged.</Paragraph> <Paragraph position="4"> Thirdly, as f mentioned before, there was 'mtil recently rio place in the generative paradigm for inferences from facts and inductive generalisatlona, even though vex;y simple examples demonstrate the need for it. This last point, ab~ut the shortcomings of conventional linguistics is not at all new, and in A1 is at least as old as Minsky's (1968,p.22) obawvation that in 'He put the box on the table. Because it wasn't level, it slid off', the last 'itt can only be referred correctly to the box, rqther than the t&le, on the basis of some knowledge quite oth#r than that in a conventional, and implausible, linguistic solution smh as the creation of a class of 'level nouns' sb that a box would not be considered as being or not being level.</Paragraph> <Paragraph position="5"> These points would be generally conceded by those who believe there isam AX paradigm of language understanding, but there wdlrld be Ear less agreement over the psfkive content of the paradigm, The txouble begins with the definition of 'understanding' as applied to a computer. At one extreme are those who say the word can only rqfer to the performance of a machine: to its ability to, say, sustaih dome farm of dialogue long enough arad sensbly enough for a hwn interrogator to be unsure whether what he is conversing with is a machine or not. On the other hand, there are =my, ahst certainly a majority, who argue that more is required, in that the msthde and representations of knowledge by which the pexformance is achieved must be of the right formal sort, and that mere performance based on ad hoc methods does not demonstrate understanding.</Paragraph> <Paragraph position="6"> This issue is closely related to that of the role of dqduction in natural language understanding, simply because deduction is often the structure mant when 'right methods' are mentioned. The dispute between those who argue for, or, like Winograd, use deductive methods, and those who dvocate othex inferential systms closer to cammon sense reasoning, is in my ways a pseudo-issue because it is so difficult t~ define cltarly what a mn-deductive system is, (if by that is meant a system that cannot in principle be lnodellea oy a deductive system) since almost any set of forrpal'procedures, including 'invalid inferences1, can be so displayed.</Paragraph> <Paragraph position="7"> The heart of the matter concdrns the most appropriate kom on an infsrsnce system, rather #an how thss~ infercncas may be &xim~tisFd, and it may well turn out that the most appmpriatc fom for plausLbla rrao~ning in order to understand is indocd non-deductive.</Paragraph> <Paragraph position="8"> Tl~is sum insight has largely defused anot2at.r tlu&med issue: wha ther tl~n app~v~rlate re&xressntstions sl~ould be proer;.dwes or d@claxntions. Winoymdts w~rk was wE the former type, as was shown by his definitions of *urJs like 'pickup' as procedures fur actually pickirry thinys up in U\F Blecks ~~sld. HQWQV~?~, simple pr~xcddral rcprcmntati~nu usually hnvo the disadvantage that, if YOU are QO~IKJ to indicate, Ess every ' it=' of knswldge, how it is tc k5;?.c used, then, if you may use it on a number of kinds uE ocsssi&n, 'y~u will have to store it that numb.es of times. So, if y~u want ta change it later, you will also have to remeinher to ohange it in all the different places you have put it. There is the additional disadvantage of lack of perspicuity: anyone reading the pr~cedural version of (he Winograd grammar rule I gave earlier, will almost certainly find the c.onventilc?nal9, declarative, version easier to understand, So then, the fa&hion far all things procedural has to some extent abated (see Winograd 1974). There is general agreement that any system should show, as it were how it is actually to be applied to langua~e, but that is not the s-e as demanding that it should be written in a ~r~c~dura3 language,line PLANNER. I shall return to this last pifit later.</Paragraph> </Section> <Section position="13" start_page="0" end_page="0" type="metho"> <SectionTitle> 5. Second Generation Svstems </SectionTitle> <Paragraph position="0"> To und'ewstand what was meant when Winograd contrasted his own with what he called second generation systems, we have to remember, as always in this suject, that the generations are of fashion, not chronology or inheritance uf i3eas. He dedcribed the work of Simmons, Schank and myself among others in his s-ey of new approaches, even though the foundations and terminology of those approaches were set out in print in 1966, 1968 and 1967 respectiyely, What those approaches, and others have in mmon is the belie. thak understanding systems must be able to manipulate very complex linguistic ohjects, or semantic structures, and that no simplistic approaches to understanding language with computers will wrk.</Paragraph> <Paragraph position="1"> In a very influential recent paper, Minsky (1974) has drawn together strands in the work of Charniak (1972) and the authors above using a teminolagy of 'frames': tvA frame is a data-structure for representing a stereotype situation, like a certain kind of living room, or going to a children's birthday party, Attached tea each frasne are several kinds of information. Sbme of thia is information about how to use the frame, Somc is about what cne can expct to happen next, Some is about what to do if those axpectatione are not confimed.</Paragraph> <Paragraph position="2"> We can think of a frame as a network of nodes and relations. The top levels of a frame are fixed and represent things that are always true about the supposed situation. The lower levels have many terminals --'slots1 that must be filled by specific instarices or data. Each terminal can specify conditions its assignments must meet .... Simple conditions are specifia by markers that might require a terminal assignment to be a person, an object of sufficient value, etc..,, II The key point about, such stxuotures is that they attempt to specify in advance what is going to be said, and how the world encountered is goihg to be structured. The structures, and the inference rules that apply to them, ate also expressions of 'partial information' (in MKarthyts phrase) that are not present in first generation systems. As I showed aarliar, with the 'women and soldiers' example, such loose inductive information, seeking confirmation Erom the surrounding context, is required for very shple sentences. In psychological and visual terms, frame approaches ~nvisage cnn undezstander as at ieast as much a looker a? a seer.</Paragraph> <Paragraph position="3"> Thus, we might, very tentatively, begin by identifyiq what Winograd called 'second generation' approaches with those making use 05 very general notions akin to what Minsky called 'frames'. But this is no more than a temporary device, for convenient initial classification of the field, because later we shall have reason to question the first-second generation distinction, and, as noted earlier, Minsky's notion of 'frame' is itself a highly fhid one in the process of definition and refinemgnt.</Paragraph> <Paragraph position="4"> Let us now turn briefly to five approaches that might be called sewnd generation.</Paragraph> <Paragraph position="5"> Charniak The new work which owes most to Mirrsky Is advocac,y is Ckarnink s .</Paragraph> <Paragraph position="6"> Hea studied what: sorts uf inforezitiizl infomation Charnlak 7 '73, ''741 would be needed to rrasulm pronoun Wiyui ties in ciai3dren l a s torjCr::, and AJ% that sehso tu understirrail t.htxu. Orla of his ex,mgic, ' otouicls ' is: 'Jane was invited tu Jack's birtl~day party. She wnderd if he wuld like a kite. A friend tcld June that Jack already had a kite, an3 that ha would mku her take it back it refers to the first kits uasrti~ned c?r tha se~wnd. Charniakts analysis begins by pirating out that a great deal cf what is required to understand that story is implicit: Iun~ilrledge aLwut the giving of p~esents, knoililedge that if one possesses one of a certain sort of thing then one may well not want another, and *so on, Charniak's system does not actually run as a preyram, but is a theosetical structure of rules called 'demcnsl khat correspond roughly to what Minsky later called frames. A demon for this exanple would be, If we sop that a person might not like a present S, then lock Pclr S being returned to the store where it was bought. Zf we st?e that hngfening, or even being suggested, zssert that thp rcasor why is tht P docs not like S1.</Paragraph> <Paragraph position="7"> The imprtant wards these are 'lt-k fey', which suggest that t31r>r~ may well bo confirming hints to be found in the :=tory and, if t4lerc are, than this tentative, partial, inference is cursect, and we have a definite and correct answer. &quot;I'his approach, of using partial {not necessarily true) inferences, in order to assext a definite answer, is highly characteristic of 'second generation' systems.</Paragraph> <Paragraph position="8"> The demons are, as with Winoyraa's work, expressed in a procedural language which, on running, will seek for a succession of inter-related 'goals'.</Paragraph> <Paragraph position="9"> Here, for example, is a demon concerned with another story, about a child's piggy bank (PB) and a child shaking it, looking fcr money but hearing no sound. The demon, PB-OUT-OF, is fomalised as:</Paragraph> <Paragraph position="11"> Again, it is not necessary to explain the notation in detail. to see that conditions are being stated for the contents of a piggy bank having ken emptied.</Paragraph> <Paragraph position="12"> The pattern being sought by the demon in operation is tire third line. TE a chin of demons can 'reacht one of the passiblo xaferente in a story then there is a suct+ass registered and the ambiguity of the corresponding pronoun is resolved.</Paragraph> <Paragraph position="13"> It can be seen that the information encoded in the system is of a highly specific sort - in the present case it is not about containers as such, and how to get their contents out, but about Piggy Banks in particular, and everything relies on that partfcular knowledge having been put in. Not all the knowledge is of this general sort: in a recent paper (Charniak '74) whws tha 'kite' stary is reconsidered there are rules ~f considerable ganrrrsuty snd interest. One such is that Charniak calis a R+SSA rule: 'XE the .tory give8 information which would make it plausible to infer mt PERSON is favourably inclined towards action A, and PERSON does SI a signif icmt subaction (SSA) of A, then f nfer that PERSON is doing A An important azsumption of Charniakls is what I shall call his 'decoupling assunaption', namely that this work on the knowledge-based detemahation of reference can be effectively decoupled from any partichlar syntactic or semantic representation, or of its apgdication, arid can be explored Fn isolation. Thus, the apparently English words in the PB-OUT-OF d-n above do not have any particular status ower than that of English words that happen ta occur in the stories being examined. I shall return to this assumption later.</Paragraph> <Paragraph position="14"> Corny has (Colby et a1 '71), (Enea and Colby '73), (Colby and Parkison, '74) developed over a pezfod of several years an on-line dialogue system thatsimilatss a paranoid patient called PARRY. A masuxa of its syrucea~s is that f* is perhapa the arost used A.I. praqsam in cr~istencsr usbit, mat La, by those. outside ths rle~earch psojact that developed it, Herd is n discontlnuouar sati~pla of its output, wain with ma. rryskeia2a mspnoao in the upper case: Ban you been hospitalised before3 SWIS IS W FIRST TXblS.</Paragraph> <Paragraph position="15"> How long ham you bssn there? ABanlWQpltE;~.</Paragraph> <Paragraph position="16"> &my headachara? NEI llEALT&quot;H IS FINE* Are you having; any difficulty with your ~tressllosy? m.</Paragraph> <Paragraph position="17"> Then perhaps you can tell me who is the President of the United States? NIXON IS PRESIDENT.</Paragraph> <Paragraph position="18"> What day is it today? mMY IS THURSDAY.</Paragraph> <Paragraph position="19"> What season is it? LE;TS GET ON WITH THE ZNTSEIVZGW.</Paragraph> <Paragraph position="20"> Colby's stm was to produce a systtm that was not fragile' in the way that most A.I. systems are: one that did not break dawn, at give up, if We input was not exactly what it expected; but me that carried en scamhaw in a11 circumstances, in the way that conversing humans do. Fragility is avoided by having m conventional syntax analyser, and by xiot even attempting to take account of all tihe mrds in the input. This is a considerable aid, since any parsex that begins to parse a more #an usually polite request such as 'Would you be so kind as to ...*' is going to be in trouble. British English speakers arriving in the UiS. quickly learn to delete such phrases, since they cause great confusion to human listeners in stores.</Paragraph> <Paragraph position="21"> The input text is segmented by a heuristic that breaks it at any occurrence of a range of key words. Patterns are then matched with each segment. There are at present about 1700 pattesns on a list (Colby and. Parkison, in pxess) that is stored and matched, not against any syntactic or semantic repfesentations of words but against the 'input ~mrd *string directr and by a process of sequential deletion.</Paragraph> <Paragraph position="22"> So, for example, &quot;What is your main problemr' has a root verb &quot;BE' substituted to became WHAT BE YOU WIN PTtOBLBH, ft is then matched awcessively in the following farms after successive</Paragraph> <Paragraph position="24"/> </Section> <Section position="14" start_page="0" end_page="7" type="metho"> <SectionTitle> BE YOU MAIN PROBW WHAT YOU WIN PROl3LW WHhT BE MAIN PROBLEM WHAT EIE: YOU PROBLEM WHAT BE XKl WIN </SectionTitle> <Paragraph position="0"> ond only We ppulthate line exists as one of the stored patterns, and so i8 wrtchd, Stored in the same famat as the patterns are rules expressing the conrepuenc3a for the 'patient1 of detecting aggression and over-friend-Liness in the intanrlewer's questions and remarks.</Paragraph> <Paragraph position="1"> The matched patterns Pound be then tied directly, or via these inference rules, to response patterns which are generated.</Paragraph> <Paragraph position="2"> Enormous inyenyity has gone into the heuristics of this system, as its popularity testifies. The system has also changed considerably: it is now calltad PARRY2 and contains the above pattern-aratchinq, rather uan earlier key Wrkt heuriatica. It has the partialt or what scme would call 'pragmtf~~, rule& about wpctation and intention, 9ndvthese alone might qualify it as 'swmn8 genarqtion' on some interpretations of the phrase. A genexator i+ alro being instaXLd to avoid the production of only *lcannedl remponsas .</Paragraph> <Paragraph position="3"> Colby and his associates have put considerable energy into actually wing to find out whether or not psychiatrists can distinguish PARRY'S re8ponaes fran those of a patient (Colby and Hilf '73). This is probably the first attempt actually to apply Turing's test og machine-person distingulshbility. There are statistical difficulties about interpreting the results but, by and laxye, the result is that the sample questioned cannot distinguish the two, Whether or not this will influence those who still, on principle, bel~eve that PARRY is not a simulation because it 'does not understand', rains to be seen. It might be argued that they are in danger of falling into a form of Fapext's 'human-superhuman fallacy1 of attacking machine shulations Mause *Aoy d not perfom s~pmk-n talkse like trcnslate mtxy, taakr that arcma p~pla c@rbhniy can - d.cr but #1i7 anajority cannot. Whdn such aceptlcr say Ult PARRY dws mt und~tstad they hava in mind e lave1 of uderstsndLng that is mrtalnky high - cna could extd their case iyonLcaJCly by pointing out that my EQN px?pl Warstad the content af oantences in the daptn am detafl that an and,~cic philosophar does, and a vszy gd thing toe. But tPI~ce can be +&t #at Paany p~gla on many omasian% !XI oaem to Wrrxstand in ti:@ way that PMY does.</Paragraph> <Paragraph position="4"> S mns The remaining three systms differ fs\# the t:w in Uaa&r attempt to provide soma repressntationrtl structure quite different frc~a L!it of the English input, This mans the use of cases, and of cxazpl~x structures that allow inferences to be &am from the attsibution of case in ways Z shall axplain, There as also, in the remaining syst=sr same attwzpt ta construct a primitive, or reduced, .-c&ulaxy irito which the lampage represented is squeezed.</Paragraph> <Paragraph position="5"> Simmons1 wxk is often thought of as a 'r~eaory Wed', though he dees in fact wy tsbore attention to wrd sense &iguitla, and ts actual recognition in text than do many other authors. For !:in the Eunhcntal ncscian is that of a 'soarantic networkq, defined aussntialfy bp the sratxmnt of relational triples qf f~mr aRb, where R is the name ut a relatlou and s and b are the names of nodes in the network. Shmorts' wrk with this general formalism goes back to at least (Shams et al, '66) but, in its never fern with case foxmalisn, it has been reported since 1973 (Shamans '?Obi, (Shaons and Bruce 1 , (Simmns and S&ocum '72) , (Sicmons '73) , ar,d (Hexdxix et a1 '73) may reasonably be considered a further implementation of Simmns' methods.</Paragraph> <Paragraph position="6"> Simmons considexs the example sentence 'John broke +the window with a hamer'. This is analysed into a network of nodes C1, C2, C3, Cd corresponding to the appropriate senses of 'John1, 'Bread', 'Windcv' and 'Hammer' respectively. The linkages between the nodes are labelled by one of the following 'deep case relations ' : CAUSAL-ACWUtT (a. , CA2) , 'EENS, =US, S03JRC6 and GOAL. Case relations-are specifications oT thd way dependent parts of a sentencer or concepts corrtesponding to parts of a sentence, depend on the main action. SO, ih this example, John is the first causal actant (CAI) of the breaking, the hammer is consideted the second causal actant (mi?) of that breaking, and the window is the theme of the breaking. Thuss the heaxt of the analysis could ba repzesented by a diagram as followd: Huwwer, thisis not the full representation, and my addition of the word 1-1s to the diagram is misleading, since the nodes me intended to be nar~ea of senses of wozds, related ta the actual occurrence of the corxesporufing wad in a text by tho relation TOK (for token), In an inrplemntatfon, a node would have an arbitrary name, such as L97, which would then rumit a stared sense definition. %, for a sense of 'apple Shmanrr suggests an ao~otiated set of featwes: NBR-singular (S), SHAPE-spherical, COLOR-red, PRINTTHAGS-apple, THEME-eat, etc. If the name of the node tied to this s~t of features was iweed L97, then that We might becane, say, 65 on being brought into saa sentence representation during parsing. Thus the diagram 1 gave must be thought of supplemented by othez relational ties fram the nodes; so that the.ful1 sentence about John would be represented by the larger set of triples: Wxd eense ambiguity i~i taken account of in that the node for one sense of 'hamaex' would be different fram that corresponding to some other sense of the same wzd, such as that meqaing Mwam3, slightly strained alternatiw for this ssntenca, The network above is slro o represantrrtion aC t)U Collouing rsnt,nger which can bb .Ihouoht of as surface vaqlmta of p shqle 'uI*ll~tlyLq~~ S~NG twe: John broke the wind~w with a hakmmr John broke the whnaow The hammer br~ke the widow Thw window broke, kt all parts of t;harknetwrk will be sat up each QP these sentences, ~i caurse, but tha need fex same item tO fill an appmpriate s! t can 3Ye infer~'dj i,e. of the first c~us~~cW~ [John) in the last tw sentences, The sentences ah~ve are recognised by mans of the 'ergative paradigm2 of orderec matching patterns, of which the Eollowing list is a part: (CAI. 'lxaxls CA21</Paragraph> <Paragraph position="8"> These sequences will each match, as left-rigw ordered itarns, one sf the dwve sentenceq. It will b2.s clear that Simmons' method of ascribing a node to each word-sense is mt in any way n prhltive sy~t~~ by which tmean a system of classifiess into which all word sanses,are mapped.</Paragraph> <Paragraph position="9"> Shmns is, however, cansidexing a system of paraphrase rules that wlould map from one network to another j31 a way that he claims is equiwlent to a system of primitives. Thus in (Simmons '733 he considers the sentence: John bought the boat from Mary Mazy sold the boat to. Job which would noracally be considered approximate paraphrases of each other. He then gives 'natural' representations, in his system, as fallews in the same order as the sentences: SL~wna opts for the first .EUom of representation, given Ule poalibility of a transfer rule gohg Ptopl either clT the shalkower representations tb Ihc other, while in (Hendrix et a1 l73), the ather apprdach is adopted, wing a primitive action B&CI.fANGG inetead of 'transfer'.</Paragraph> <Paragraph position="10"> ThIe implementation under c~nstruction is a front-end ,oarsex of the Woodat augmented transition network type (see Woods '701, and a ganefation system going gram the sea~tic networks to surface strings described8in detail in (SWns and S1ocu.m '72). SimPons has also given considerable tb (SFrPPJons and Bruce '7 1) to the auloPPatic ttanslation of the networks into a cmxrespontling Efrst-order predicate calnilus format of the sort developed by Sandenall. (1971). This last is particularLy valuable because, LZ penatalirablr, lt rhwr that any linguistic di,ng in network Ecm -- can b translated in- some form of the predicate calcults, if that Formalism and it# crssvclated proof kachnigurs oan be shorn to be app?!!priate for carbin p#~~blaa$ La the area of natural language analysis.</Paragraph> <Paragraph position="11"> Sehank SchanScqa is a rich system of semantic representation, developed aver a psiad of six yeaxsf with the collaboration af a number of talented students. Its graph system of notation has influenced psychologists like &belson (1973) , amony others. Schankls contribution has been the notational syste'~ representing the structure of natural lancruage sentences, and this has &ern progr-ed by various collaborators over #e years. In its present version, caJled WGfE (schank et a1 '73) it has an ana'lyser of I&quot;;nal&h t%~to Zdesbeck (19741 s seiixmtic memory Ewpnent due to RiegW (1974) , and t genetatox of English due t;o GoLban (1974). MaRGIE produces cxutput in two modes, demonstrating the sort of wnceptual FnShrencfng that goes on at -the Bevel s'f the stmmUc repkesenbUwB: Tb Pm md th I- amtifa. Samples of 'input wd outwt tm rad ho t& ba rPdw em h shm thus: hsy did because she was unable t~ i-le air azla she was u&Ls to inhale sme air because John grahkd her DIF& The ah of Schadc's aystem has ahways hen txa pmvida a --Urn of meaning in terns of which these aLsdl ether tasks, such sa &i,m indeprsdent of any pasticullax language, and of syntax, /ityq fS'sbrca, Oe all surface structure whatever.</Paragraph> <Paragraph position="12"> The fcml structure of Schank's grans is that of (Bays '64), ad the items in gsaw axg pf f~ur typesq ar cmtegories. null They are symblisd as PP, Am, PA asld M, vhfdr am mzzuqms, ht which correspond closely (for the purpose of um3e.r~- th.ix feam=tkd tb those of a noun, verb, adjective and adverb, respctidy.e+ T'Ik basic</Paragraph> <Paragraph position="14"> * Schank distinguishes 'ccnceptuall and 'searnantic' zepesematia&~ im rw that is important fox hh within his own syttep. Bowever, Z obrll tmm the terms indiffexently since, in this brief ad stprfichl W-, nothing hangs upon the distinction.</Paragraph> <Paragraph position="15"> **This is a considerable overs~plifica~~n~ in cx&w to gim r 4 self com.taine3 ,descrSigrion. ;Butl in fact, aany Ex@ km represented as Am's: chair, pen+ honesty, am3 'trwItfcm, structtare is called a conceptualiaation, and is normally intrbduced with a etraightfomard dependency structure such as, for the sentence 'The man teok a book': an & take + book Here 'pf itldicatea past, and is the aepndendy symbol liking a PP to We ACT ('take') which is the hub of the conceptualization, as with Simon& ?'he '0' indicates the objective case, marking the dependence of the object PP on the central ACT. There is a carefully constructed syntax of linkages between the conceptual categories* that will be describpd only in part in what follows.</Paragraph> <Paragraph position="16"> The next stage of the notation involves an extended case notation and a set of primitive ACTS, as well as a nqer of it:ems suoh as PHYSWNT which indicate ather stqtes, and items of a fairly simplified psychological theory (the dictionary entry for 'advise', for example, contains a subgraph telling us that Y 'will benefit' as part of the meaning of 'X advises Y' [Schank '73). There Axe four cases in the system, and their subgraphs are as follows: 'Phere are at present fourteen* basic actions forming the nubs of the graphs, as well as a default action DO. They are: PROPEL, .MOVE, INGEST, EXPEL, GRMP, PTRANS, MTRANS. ATRANS, SMELL, SPEAK, LOOK-AT, LISTEN-TOI COW and MBUZLD. The notions of case and primitive act are related by rules in the develOpment of conceptualizations. So, for example, the primitive act INGEST has as its instrument the act PTRANS.</Paragraph> <Paragraph position="17"> mere are also other infer--- -* Since the publicabion of (Schank 73a) their number has been reduced to ezeven (plus DO) by the elimination of SHELL, LISTENID, IXK)KAT and COQC, and the adation of ATTEND, ances fram any ACT classified as an INGEST action, such as that the thing ingested changes its fomt that M tAe UA~G Awp&ac3 is&LUe kb In gestet becomes lmrs nourishodl stem (oao Scbnk '73, pp. 38tf .I. This will all kcam clsaxar if we consider tho trmsitl~n Zrcm a dietiwary entry Pox asr action to a filled-in mnceptuahirati~n. Hem Is tthe d~ccionary entry for the action 'shwt' : can consider this entzy als an active 'frAxte-like~ object seeking filler itms in any context In which it is activated. Thus, in the sentence 'John sho the girl with h riflet, the variables will be filled in frcm context and the case inference will be made fm the main act PROPEL, which is that its hstruuent is lkSOVEI GRASP or PFtOPRL, and so we will arrive at the whole con- null This case inference muSF b~ made, according to Schank, in order to achieve an admate zepresentation. There is, in the last diag~am, a cextain redundancy of expression, but as we shall see tn the next section this often happens with deeper semantic notations.</Paragraph> <Paragraph position="18"> More recently, Schank, together with Rieger, has developed a new class of causal inferences which deepen the diagrams still further. So, in the analysis of 'John's cold improvPQ1 because I gave him an vplel (Lrt Scfrank '74a) the extended diagram contains at Ih~t four yet lower levels of causal =rowing, including one corresponding L.ht the notion of Juh cans+Ncting the idea (WBUTLD) that he wants to ea% an apple.</Paragraph> <Paragraph position="19"> So we can see that the undexlying explication of mean* here is not only in the serlso of Iinpulistlc prLmltLws, but in tern of a theory of mental acts as well. Now Ulsra ate a number of genuine ~positi6na]. difficulties here for the euml&tator faced with a epstm of this complexity, One aspect of thi~ is the atages of developnt pf the ~ystam itself, which can bc seen ae a consimtcntpmcesa of produrlng what was argued for in advance. For ekampla, Schank claimed early on to be a constructing system of semantic mtructures undatlyirrp the 'surface of natural language', alehaugh initially them were no primitives at all, and qa late as (Schank et a1 '70) there was only a single primitive TRANS, and most of the entries in the dictionary conmisted of the Bnylish wards coded, together with subscripts. Since than the primitive system has b&ossomed and &ere are now twelve primitives tor hCTS including three Ebr the original TRANS itself. Each axposition of the system recounts its preceding phrases, from the original primitive-free one, throuqh to the present causal inference form; rather as each human foetus is said to relive in the womb all the evolutionqy stages of the human race. The only trouble with this, fxm an outsider's point of view, is khat at each stage the representation has been claimed, to be the correct one, while at the same ti- Schd admits, in moments of candor (Schank '731, that there is no ad to the conceptual diagrrVbraing oE a sentence. This difficulty my well reflect genuine problems in language itself, and, in its acuteat form concerns a three-way confusion between an attractive notation for displaying the 'meanings of wordsv, the course of events in the real world, ad, finally, iibtual procedures for analysis to be based on the diagrams, This raises the, to me, inrportant question of the application of a semantkc system, that I shall touch on again later. Schank, for example, dues mention in passing the questions of wd-sense ambiguity, and the awful ambiguity of English prepositions, but there are in no way central for him, and he assumes that with the availability of 'the correct repxesentationWt his sysh when UnpleIm?nted must inevitably soive th- traditional and vexing questions. M procedures are hinted at along with the graphs as to how tnrs is to be done. A distinction cof importance may be becoming apparent here batween Schank s work and Riegerls: in Riagar's thesis (RLqer '74) the rules of inference appear to craatg clclparata and new rubprrphr wnicn may swd in an lnfersntial celatlon to each other so as tsl produce cloncluaiona &but: ~~~~~a of, gayt pronoun rabr~ence, etc. But in Schmk's cormsponUng papers the s- ihfctrsncss urn not applid ke actual problems (Schank '74a) but only be- to amplaxity th. conceptual graphs yet further.</Paragraph> <Paragraph position="20"> Closely connected witla this raattex is the quaation af the survival of the mlvface sr'tructure in tho diagrams. Until very recently p~~h,itiuisation applied only to verbs, tht of nouns being Left to Mehr [Wekwr 272) Most recently, though, noun wds have been disappearf ng from dfagraps and been replaced by.categorfes such as +PZIYS0=* But At is cleax that the swface is only slowly disappearing, rathex than having been abhorred all along.</Paragraph> <Paragraph position="21"> In a mra recent publication CSchank '74b) there are signs that this. trend of infinitely proliferating diagrams (for indivihal sentences) is feversing. In it Schank considsrs the application of his approach to the repraqentation of text, and concl :des, correctly in my tP4etq, that the representations of parts of the text must be interconnected. by causal arrows, and that, inlordm to present@ 1rv.cidity, the conceptual diagrams for individoal sentences and their partvi must be abbreviatedr as by triples such as POEPLE PplWS PEOPLE. her^ indeed, the surface simply has to surviw in -the representation un1esF one is prepared to camit oneself to the axeme view that the ordering OF sentences in a text is a purely superficial and arbitrary matter. The Fensein wnich this is a welcome reversal of a trend should be clear, because in the 'causation inference' development, mentioned earlier, all the consequences and effects oE a conceptualization had to be drawn within itself. Thus, in the extreme case, each sentence of a text should have been represented by a diagram containing most or all of We text of which it was a part. Thus- the representation of a text would haye been impossible on such prihciples.</Paragraph> <Paragraph position="22"> Pay own system constructs a semantic representation for small natural language texts: the basic representation is applied directly to the text and can then be 'massaged' by various forms of inference to became as deep as is necessary for well defined tasks demonstrating understanding. It is a uniform representation, in that information that might convenionall~ be considerea as syntactic, s-ntic, factual ox rnrerencial LS weu axpressed within a single type of struwra. The fundamental unit in the mnatructioe of th2s beaning representation is the template, which is intmded to correspond to an intuitive notion of a basic message of agent-action-object fom.</Paragraph> <Paragraph position="23"> Templates are rigid format networks of more basic butldlnp blocks called fomulas, which correspond to senscln of individual worde. In order to cohskruet a cctnplete text representation templates ate bound byethat by two kinds of higher level structures called paplates and inference rules. The templates themselves are built up as the construction of the representation proceeds, but the. formulas, paraplates and jlnference rules are all present in the system,at the outset and each of these three types of pre-stored structure is ultirpately constructed frm an inventory of eighty semantic primitive elements, and from functions and predicates ranging over those elements.</Paragraph> <Paragraph position="24"> The system runs on-line as a package of LXSP, k&ISP and MLISP2 program, Wing as input small paragraphs of English, that can -be made up by the uber from a vocabulary of about 6QO word senses, and prduzfng a good French t:ransL&tion as output. This environment provides a pretty clear teat of lmguage undarstding, bcauaa E'rench translations for everyday. pmse are either right or wmng, and can be seen' to be 910, while at the same titPe, the mfot difficulttee of understanding ptogtams - word sense ambigratty, case anbigufty, difficult pronoun reference, etc. - can all be represented within a machine translation environment by, for example, choosing the words of Lhe input sentence containing a pronoun reference difficulty so th&E the potsible alternative references have different genders in French. In that way the French output mdkes quite clear whether or not the program has made the correct inferences in order to understand what it is transla_ting. null The program is reasonably robust in agtual prformance, and will even tolerate a certain amount ob bad grammar in the input, since it does not pzfbm a syrkax a~LysI;s &=%he sense, hut snnkn message forms representable &XI the semantic smctures employed.</Paragraph> <Paragraph position="25"> vpical input would ble a sentence such as 'John lives out: QE tQtm hi &inks his wine out of a bottle, Mo than throws the httlas out uf Uae for each of tke thraa occurrences of 'out oft, since it raslisarrs that they diffexenco must be reflected in the Fwonch, A sentonce such as, *Give Use monkeys b;mands although they are not ripe because they ay~ very I~wQry' produces a translation with different equivalontd fur t)ra tm eccurrmwr of lUaey1, bocause the syslm oorr@ctlp realiasst iraQlSI wlrat 4 shall describe below at preference considerations, that tlae most sensible intct~retaticn is one in which the first they' refers to the bananas and the second tc the monkeys, +and bananas and monkeys~have different genders in French. These two exmph are dealt with in: the 'basic de' of the system.</Paragraph> <Paragraph position="26"> (Wilks 73a) Inmany cases itcmotxesolve pronoun ambiguities by the sort of straightforward 'preference considerations1 used in the last exaaple, where, roughly speaking, 'ripeness' prefers to &-predicated of plant-like things, and hunger of animate things, Even in a sentence as sbple as 'John drank the wine on the table and it was gdtt such considerations aye inadequate to resolve the anbigtlity of 'it' between wine and table, since both my be good things. In such cases, 02 inability to ras~lve within its basic moder the program deepens the xepresentatio~ of the text so as to tq and set up chains of inference that will reach, and su prefer, only one of the possible referents. I will return to these pzocesses in a nment, but first I shall give sane brief description of the basic representation set up for English.</Paragraph> <Paragraph position="27"> For each sense of a word in its dictionqxy the program sees a fohula. This is a tree structure of semantic primitives, and is to be interpreted formally using dependency relations. The main element in any fonnula is the rightmost, called its head, and that is the fundamental category to which the fonnula belongs, In the formulas for actions, for example, the head will always be one of the primitives PICK, CAUSE, CHANGE, FEEL, HAVE, PIXME, PAIRl SENSE, USE, ,WANT, TELL, BE, 5X), FQRCE, W, -THINk, FLOW, W, DROP, STRZK, FUNC or HAPN.</Paragraph> <Paragraph position="28"> Here 18 tnts eLee stxuctufe for the. action of drinking: (mu PART) Qace again, it is n~t necessary to explain the formalism in any detail, to see that this sense of Idrink* is being expressed as a causing to mve a liqyid object (F'WH m) by an animate agent, into that saine agent (containment case indicated by IN, and formula syntax identifies SELF with thB +gent) and via (direction case) an aperture (TLIRU PART) of the agent. Template structures, which actually represent sentences and their parts are built up as netwcrks of formulas like the one above. Templates always consiat of an agent nude, and action node and an object node, and other: nodes ttat laay depend on these. Sot in building a template for 'John drinks,wine', the whole of the above tree-formula for 'drinks' would be piaced at the central action node, another tree structure for 'John' at the agent node and so on. The complexity of the system comes from the way in which the formulas, considered as active entities, dictate how other places hn the same template should be filled.</Paragraph> <Paragraph position="29"> Thus, the 'drink1 formula above can be thought of as an entity that fits at a template action node, and seeks a liquid object, that is ~ say a f~rmula with (FLOW STUFF) as its right-most bzanch, to put at the object noda of the same template. This seeking is preferential, in that formulas not satisfying that requirement will be accepted. but only if nothing misf actan ca lXZ'-fotUEa. TIie -EUElliplate Uif ly esWIisned Tm 3 Tragment of text is the one in which the most formulas hive their preferences satisfied. There is a general principle at work here, that the right interpretation 'says the least1 in inforreation-carrying terns, T)rh wry simple device is able to do much of the work of a syntax and wxdysnse wzibigukty resa1vi;tap pxagraa Pnar cusq~e, LZ the a,mteme kd been 'John drank s whole pitcher1, the fomulr tor th. 'pitcher of klquidb wuld hawe hen pr.rsEerreCI to that for thar human, sfm the subf~mS;a (FLOW STUFF) could be apprtopriateAy located uithirr it.</Paragraph> <Paragraph position="30"> A tonsidarable tamwnt af squeezing af this sbapl~ eansnkcal Lorn of template is necessary to We it fit tha mmplexfty of language: texts have to bt Eraymented initfalLy? then. in fragments which am. say, grapositional phrases there is a daaay agent Lapasad, and the prepsitions1 phrases thexe is a dummy agent imposed, and the gremsiticmai LomuLa functions as a pseuda-action. There are special 'less preferred1 oaliers to deal with fragments not in agent-acti~n-object order, and so on.</Paragraph> <Paragraph position="31"> men the local inferences have been done that set up the agest-action object templates for fragments of input text, Shd system attempts tm tie these templates together so as to provide an overall initial structure fox the input. One form of this is the anaph,oxa tie, o-f the sort discussed fog the monkeys and bananas example above, but the =re general PSom is the case tie. Assignment of these would result in the template far the last clause of 'He ran the mile in a paper bag' being tid to the action &e of the template for the first clause ('He ran the mile'), and the tie king l~~~ed CONTaiment. These case ties are made with the aid of-another class of ordered stxuctues, essentially equivalen* to FMlrPore s case f ruses, called praplates and which are attahea t~ the formulas for English prepositions8. SO, for 'outof', for =aaple, there h-ould be at Least six ordered paraplates, each of which Is a string of functions that seek inside templates for information. In general, paraplates range across two, nat necessarily contiguous, templates. So, in analysing 'He put the nuhr he thought of in the table', the successfully matching paxaplatz would pin down the dependence OP the template for the last of the three clauses as DIREctior., by Wing as ampment only that particular template for the last clause that contained the formula for 'a numerical table', (and - not a template repxesenting a kitchen table) and at would do that because of a function in that paraplate seeking a similarity, of head (SIGN in this case) between the tt~m appropriate objecr, ConqularJ for ~numbarl and 'tabla'. The other template cumtaining the tfurnftuzsq formula for 'table* would naturally not satisfy the function brcaure SIGN would niok ba the kesa of this amme foxmula for %able9, The structure of mtuailly cc~nnected templatars that has hen put togsthax thua fu conetitutau a 'atmiantic blockg, and, if it can h con&txucted, then ar far as the mystrtm is concerned all osmsntic and referential ambiguity has hen reaolvd an8 it will begin to generate French by unwrapping the bldck again. ?&quot;ha generation arrpects of this work have hen dracriW in (Hor~akovitr 73 . One aspect of the general notion of preference is that the aystan should never construct a deeper or mre elaborate oqmattc rraprssantation than is necessary. fox the task in hand and, FE the initial block can be constructed and a generation of F- ;rich &one, rto 'deepening* of the representation will, be attempted, HOW~V~X, wmy exmples cannot be resolved by the methods of this 'baeic mode' and, in particular, if a ward sense arPbiguity, or pronoun reference, i~ still unresolved, then a unique semantic block of templates canrnot be constructed and the 'extended mode' will be entered.&quot; In this &a, new template-like forms are extracted fran existing ones, and then added to Me template pool ftom which further inferences can be made.</Paragraph> <Paragraph position="32"> So, in Ula tm~lata derived earlier for 'John drinks wine', the system enters the Loarula for 'drinks', and draws inferences corresponding to coach case sub-Eorrmula. In this t~xtmple it will derive template-like forms equivalent to, in omJf~ry English, 'The wine is in Jobt, 'The wine entered John via an aperture' and so on.</Paragraph> <Paragraph position="33"> The extracted templates express information already implicitly preser.t in the text, wen though many of them are partial inferences: anes that may not necessaxily, be true.</Paragraph> <Paragraph position="34"> -n-sense inference rules are then brought down, which attempt, by a s-fe strertegy, ta construct the shortest possible chain of rule-linked tmmplate fo-8 from one containing an ambiguous pronoun, say, 50 one c2ntainhg one of its ~ssible referents. Such a chain then constitutes a solution ta the ambiguity problem, and the preference approach assumes that the shartest chain is always [the right one.</Paragraph> <Paragraph position="35"> So, Yn the case of 'Jahn drank tha wine /on the table/ and it was good', (in three temprate-matching fragmenb as shewn) the camact chain t~ 'wine' uses the two rules to wht wa may xeao~rmbPy Iwk out tor in a gkvrn rttutkon, net? to ubt WSP happn. Tha hypotl,.sir irere btrt wderrtahl&np can only trkr plracs oh the baefs of .akpptble rufoo that: are mafixred by the eunlext af ap~lication. In this axample the chain constructed uy ba expressed as (w8srinq the &ova sguam bracket rmtaatio~~ to cont&bn nut a representation, but sisxxply an indiedtion, in BngldsA, of the template contents): The assmption here: is mat tw ehain ushq ather inference niles wttl,d have reached the ' t&1q1 solution by using less ~.aa two sules, The chief drawback sf this sp!irm is that dings consisting entirely of primitives have a considerable amount of bo'eh vagueness and redundancy For example, ns reasonable coding in terms of structured primitives could be expected to distinguish, say, 'hmerl and 'mallet'. That my net matter provided the cdings can distinquish iapostantly differe~': of words. Again, a template for the sentente s he sheperd tended his flock' would contain considerable repetition, each node sf the template trying, as it were, to tell, the whola story by itself, again, the ~refasence cziteria are not in any weighted, which might seegn a dxahcack, and the prefexential chad LET@I criteria for hference chains miqht v~ff seem too crude. Whether or not such a, system can remain s-le with a WrutderabLe vecabulary. of say several thousand words, has yet to be trrtd.</Paragraph> <Paragraph position="36"> ft will ba ovidrnt tso any reader that Zha laat twa systems described, Bch.nktm ud my moun, share a great deal in cccpmon. IWsn tha apparent Qtff*rence In notation is reduced if one see$ the topological similarity mat rorults from mnrlderfng the head of a formula as functioning rather lLko a Schwk bait action. If one thinks of khe dependencies of the case eubprte of a fornula, rot &ranged 1 lneargy along the. .bottom of a tree, but radiating out Ex- the head in the centre, then the two diagtsms actually have identical topologies under interpretation. A difference vises in that the 'filled-in entity' for Schgnk is the conceptualization centred on the basic action, though for me it is the network oE formulas placed in relation La a t~lopLate, whexe there is indeed a basic action, the he& of the action formula, but there is also a basic entity in the agent formula and SO on. OX, to put FL another way, both what-is and what-is-expected are represented in the templates: the agent formula represents the agent, Pox exampla, but the left-hand pact of the action formula alsp represent3 what atgrant was expected or sought, as in the (*ANT SUM) sub-formula of the *&inkt formula, A~thou~h developed in isolation initially, these twlo systems have also influenced each other in more recent years, probably unconsciously. For eatample, conceptual dependency now emphasises the agent-action-object far~rcrt -re than befoxe, and is less iverb-cent.red' and t~heless while, ronvez-sely, rrty own system now Wes much more overt uuc= of ~les of wtfa1 LnfcanaaUun than in its earlier versions. Again, b~th systems have intellectual conneotions that go back before either generation of A1 systeam. In my view, both these systems have roots in the better parts of the Computational Linguistics movement of the Fifties: in the case of Scfrank's systm, cane may think of the earlier systems of (Hays '64) and (Lasrb '661, and the arkow-structured primitive system of (Farradene '66) ~~EB&e~~OLa3s-~~~prcceaents.Fnthp .Parkex&odes '61) system of classiffc&ion awl the early seamtic structures of (Richens '61) and (Casrsterman '61). In 1961 the last author was arguing that 'what is needed is a disoiplina that will study suantlc nsolrge camaction in a, way malogous to that in which r&nrba#omatfes now rtudiro ~~thmtLcal connaetion, and to that in which ~Ulm~tLcal Ilngulstic~ AQW studla8 SPUC~~C CWU3acti08l1 * (LUd. , p. 31 This historical pint x&fsrso s final bns that is, I feel, of prgsing interest. Then seem t~ bs two rsscarch styles in this field: one is what sight b9 callled the 'fully finished style1, is whish ma wxA exists only in one ccmphte fow, and is not issued in iaerly ar dove1 vets iodns, The best example of this io Winoptad's wrk. The other type, examplified by all the other authots di8scussad hate, to same extent, is the det-eloying style: work which appears in a n-r of vlersiens over tke years, one bps with gxadual hprovt?ments, perkraps in attmpts ta tackle a wider range of lfncpistic or other inferential phenomena. There are &vatages to both styles, but even in the latter one hws that any proposed stxuctura ox system will, in the end, be found wanting! in Lhe balances of language, so it can only be a question of when one will have to abandon ~t. The interesting question, and one to which no answer could possibly h given here, is just how far is it worth pushing any given structural approach before starting again fram scratch? 6, Sane Cormpisans~, CVkd,. -,l'qn-$r-ass In this sactivn X shall -para and co~~trast, under some nine interconnected headings, the projects ae.;cribed in be My of the papex, This is not easy to do, particularly when, the pxesent author is among the writers discussed, though that is easily mrdied by be reader's Wing an appropriate discount. A more serious problem is thb , at this stage of research in artificial intelligence and r:ztural language, the most at-Ctractive #istinctions dissolve on more ds~ailed scrutiny, laxgely because of the lack of any p~ecise theoretical statement in =st, if not a11,'tha major prn jects. There are those who think that it therefore follows that this is not me lzrwrent for any form of critical camprison in this field, and that no more is needed than a 'psitive attitudey towards all possible pxwjeeb. OrtZythsewho feel tht,.-wt theemkrary, any kbe is asgod @Len arms at2 u~ @puo *Xqoa ayr7 aaoqa - puey am a- m @a-1 TB~~~PU xo,; uo~~~uasexdax 30 TaheT aawj~dcmdd~ ay? surojum iimpmxdd~ uo~qarauab puoaaii ay7 B?mm zcpdqg 7na=rxnr, p vt up~3o3-a-w 60 p-&quot;z -sn.lrwlm a3=Tpla;rd u? passardxa msXs s , auo axaw qua;EPdstrea aq pvcm q~a3 aqq p anbv -yaaq bu~~ozd-asxoaqq pzapu~qs Xxa~ awzs 73123 q axaw sanljm- wtfazapq s,euo qew zea3 am aq pTn= 'aldmoxa ro3 &uossaz peq B :asam azo~dbm a eaqd am qcm sdeqrad araq put? uo~qqolu prams e bqp~oha xoz stmmaz PW ~ltl~ P6 tn~ am *SPP~F~ a3e3~~33d ~aha lo (EL, fl-~ LD ax-) sap= tmoTasnpcud tq passaxdxa eren smq9Xs sqq yo ;I~T xaysaa qanm aq X~xwl3'.p~na~ &pd s~w 30 3s- aqq 316511 &XPB 03 SF 1.pssnasTp siayrm yamem ftm 49 pasn smzbalp pua rpuopqou aulsla33fp so qa~ew s~ 3apzw-q s2swxauo~ gua suomrrodans wyqe0~130.)38,03 raqqoyy 'pax.dB03 aq 03 suarcud atp Xluo pup naantm tp~yrr Am! 30 qasqno am qle #nap S'J qj 0s tsaq3euzdd~ ~to~~?riruo6 5~2~3 saw0 UP TTarJI sta &quot;yrm UMO s,Pz)Z~~~U~ ran- X~upqxa;r pTnm qupkl pumas oy7 *raho -arw -7nq qndu~ am u~ X~q~q'ldxa quasard qw us~~wo3u~ qqm uraauoa TTe pawaqasa re3 os seq aq log Jpuo3as ar(;) ~TM swtriqs xou 4auydI asx~y en WTn Aqlq) zo ye~lrr~w xanru qw mnon, lpooxdd~ resentatfm, er rather, that they whd a#r xrtrlilard at a 8c at &ma& of magrssentaths~, d~pendimg on th mbjer=t area. khm &a bubt a86 the r.egg~scntathw in terns M -isriatam tBm% bm im 'B$s ~lrk mar,$ to 31e Ln we-tow cxxrorpwdbam vlsh ropLM m.</Paragraph> <Paragraph position="37"> The slsaqest how-kewl appmach fs -y ht ollrcm;t$J dm that this dispute is ultimateby one oE degree, simx no aae clnilr that every lccution recognized by an $nltelligaat rryl- art br i.Qpd hto a 'deep' representation. To taka an extmme auab amy -tam tikt uag@ 'Gxd W8rninqT into a deep sewm%hc represemgltkm that the carsect sespsns8 was also rGcad &arnhxjt wauld $Ch & seirfms thesretical mistake.</Paragraph> <Paragraph position="38"> Hawever, themst serious arqument fa ira mn-tqpar%fshr$ -tmUm Ss mt in kerns 09 the av~Hdaqam of ammbm~ d%g%%dUw, & closely tied to the defence of se&anthc pahb2hwm %EI m, u&%& Ps a large subject not to be unde~taken here. Cbe a% the tmxUes & tic prh,itivqs is that they are open ka - bad &e%eaces, ihumua z;cW than increase their plausibility. For -e, users a;JF fur linguistic representation have declared them to hve scme d -WU existence and have implied that there is a 'right sett ~Wwes bpq to ernpiriqal discovery. On that view the essentially hqd.sUc of structures sf prbitives is lost, because %t is an usmwkLd fe&are af a language that we can ch-e its vocabuhry as function ui& dLt&matf.a vocabularies. But if there is a zight set of prLmiUwm, utmt~ mk~~% are the awes of brain-items, then that essential SF=- uxald km h&t-What is the is that there is a considerable amount of psychologicill evidence that Geople a2e able to recall. the content of uhat they hear and understand without being able to recall either the actual words or the syntactic structure used, Thare is large literature on this subject, from which two sample references would be [Wettler '73) and (Johnson-Laird '74) . Thesc results are, of course, no proof df the existence of semantic primitives, but they are undoubtedly supportkng evidence of their plausibility, ao is, on a different plane, the remlt from the encoding of the whole Weboter's Third International Dictionary at Systmv Develagmefit Corporation, where it was found that a rank-ordered frequency count of the words usad to define other words in that vast dictionary was a list (omitting 'the1 and a which corresponded almost item-for-item to a plausible list of ssmantic primitives, derived q ptioxi, by those actualhy concerned to.codel the structure of wmd md sentence meanings.</Paragraph> <Paragraph position="39"> Zt is important to distineish the dispute Ibout level from the, closely connected, topic that I shall call the centrality of khe #nowledge required by a language understanding system.</Paragraph> <Paragraph position="40"> Centrality What X aria calling the centrality of certain kinds of information concerns not its level of representation but its non-specifidty: again a contrast can be dram between the sorts of infomiltion required by Charniakls s~st~, 0x1 on@ hand, and that required bySchankls* and my om on the obhar. Charniak's examples suggest that .the fundemental form of information is highly spacific** to particular situations, Like parties and the giving of presents, while the sorts aPS information central to Schank's and my own systems are general partial hssertions abut human wants, expectations, and scr on, my of which' are so general as to be almost vacuous which, one misht argue, is why their zple in understanding has been ignored for so long. * Though as noted earlier, Schank in 1975 has adopted Rbelson9s (1973) notion of 'script', as a largar-scale 'frame1, in such a way as to incorporate much less 'central1 knowledge.</Paragraph> <Paragraph position="41"> **In a recent paper (1974), Charniak gives much more general-rules, such as his 'rule of significant sub-action', mentioned earliw.</Paragraph> <Paragraph position="42"> If I were a reasanably Eluent spaker of, say, G8man, 1 might we13 not understand a Gem conversation about birthday presents unless Z had tietailed 8factuaS. information &but: hw Gens~~~ns organixa the giving of! presents, which Plight be considerably different the way we do it, Converselys aP course, 3 migl.rt umiarstd much og a twkmlcal artLcle abut a subject in which I was an expert, even th~~rgh 1 knew wry Ifttka af the language in which it was written, These az'e certainky wnrL3exatians that tall Lox Charniak's approach, and it La perhaps a paradox that the s~rt of r~aturrl language understalridler that wuld tend to COJIP~~ his apswp,t tons ~muld be one concerned with disooursa &L.r;)ut, say, the details aL reyapking a t$Otor CU, where factual. infsmathon is what is centsalt yetr imnically, Charniak has concehtrated on something as general as childtents stories, with their need of deep assumptions about hurwn desires and khaviour. In the end 'this difference may again turn out to be one of enphasisj and of what is most appropriate to diSSerent subject areas', though there niay be a vexy general issue lurkiw somewhere here. It seems to me not a fuolish question to ask whether much of what appears to be about natural language in A.I. research is in fact about language at all, Even if it is nbt that may in no way detract fran its value. Newell (blolore, Newoll q73) has argued that A.1. work is in fact 'theoretical psychology', in which case it ceul8 hardly be research - on natural language, When describing Winograd's work earlier in the paper, Z raised this question in a weak farm by asking whether his definition of Ipickup1 had anything to da with the natural language use of the word, or whether it was rather a description of how his system picked samething up, a quite different matter.</Paragraph> <Paragraph position="43"> Suppose we generalize this query samewhat, by asking the apparentky absurd question of what would be wrong with calling, say, Charniakls work an essay on the 'Socio-Economic Behaviour of American Children Under Stress? In the case of Charniak's work this is a facetious question, asked only in order to make a point, but with an increasing number of systems in A.I. being designed not essentially to do research on natural language, but in order to have a natural language 'front end' to a systm that is essentially intended to predict chemical spectra, or play snakes and ladders or whatever the question becomes a serious one. It seems to me a good time to ask whether we ahould expect advance in understanding natural language from those tackling the problems head on, or those coroncerned to build a 'fr8nt andv. It i~ cLtt,xly the case that - anpiece coulL bp esrcntial to the understanding of sane story.</Paragraph> <Paragraph position="44"> The question is, does it follow that the epehifict.tion, organieation and formalization of that knowldge la &a studf oP l:.aga, because if it is then all human enquiry ftm physics and history to medicine is a linguistic enterprise.</Paragraph> <Paragraph position="45"> And, of cowrr, that poaskbility has actually been entertained within certain strains of darn philosophy.</Paragraph> <Paragraph position="46"> itowaver, I am not wing hefa, to breathe fresh life into a philosophical distinction, batween being aLuut lmpunge and - not being about language, but tather introducing a practical distinction, (which is also a consideration in favour of optiqg, a3 I have, to work on very general and central areas of howledge) between specific knowledge, and central knowledge without which a syartem could not be said to unilexsttind the language at all. For example, I might know nothing of the arrangement of American birthday parties, but could not be accused of not understanding English even though I failed understand sme pazticular 'children's story. Yet, if I did not have available acme very general partial inference such as the ane people bainq hurt an8 fallingr or one about people e*avouring to possess things that they want, then it quite possible that my lack of understanding of quits airtple aentencee would cause observers to think that I did not underrW Englbsh. An interesting and difficult question that then arises is whether those who concentrate on central and less central areas of discouse could, in principle, weld their bodies of inferences together in such a :gay as to create a wider system: whether, to put the matter another way, natural language is a whole that can be built up fxm parts? Pken-noloqica level Another distinction that can be confused with the central-specific one is that of the lphencmenological levels1 of inferences in an understanding system. I mean nothing daunting by the phrase: consider the action eating which is, as smatter of matmica1 fact, quite often an act of bringing the bones of my ulna and radius (in my arm) close to that of my lower mandible (my jaw). Yet clearly, any syatQIP OP CCEPgeK).n sense inferences that considered such a truth when reasoning about eating would be making a mistake. One might say Ulat the phenoeenolqtcrl lrvrL of the anraly_sis was mng even thourgh all the InF'amznces it: !!ad8 ware Uue ones, The stme wuld be true of any W.I. system that wade everyday inferences about physical objects by mnsiQaring their quantum structure. Schank's analysis of eating rontaias the inf'matian &st it la done by uovirsg the hands to the mukh, and it might be argued that yven ulis is goisrg too far ftom the '@aaningl of eating, whataver that my bar towsrds generally true information about ma act which, if always inferred &ut aU acts of qating, will carry the systesrs nruamageably fax.</Paragraph> <Paragraph position="47"> Therq is no denying that this sort of infomatioar might be useEul to have around somewhere; Wt, in Minsky's terms, the 'default1 value of the instrument for eating is the hand brought to the mouth, so that, if we have no contrary infomation, then that is the way to assue that any gfvm act of eating was performed. Nonetheless, there clearly is a danger, and that is all X am drawing attention to here, of taking inferences to a phenolnencw ldgical level beyond that of uammn sense. A clearer case, in my view, would be Schank's analysis (1974a) of mental. activity in which all actions, such as kicking a ball, say, are preceded by a rsrenta9 action af conc~iving or deciding to kick a ball. This is clearly a level of analysis untrue to caumn sense, and which can have only harmful effects in a systea intended to mimic corxlaPon sense reasoning and understanding.</Paragraph> <Section position="1" start_page="7" end_page="7" type="sub_section"> <SectionTitle> Demupling </SectionTitle> <Paragraph position="0"> Another general issue in dispute concerns what I shall call demupling, which is whethex ox not the actual parsing of text or dialogue into an 'understanding system.' is essential. Charniak and Minsky believe that this initial 'parsing1 can be effectively decoupled from the interesting inferential work and simply Qssumed. But, in my view, that is not so, because many of the later inferences would actually have to be done already, in order to have achieved the initial parsing. For example, in analysing 'He shot her with a colt', we cannot ascribe any structure at all until we'can make the infexences that guns rather than horses are instruments for shooting, and so such a sentence cannot be represented by an 'inference-but-no-parsing' structure, without aremi- that language doas not have one of its esgential charaeteristlca, namely ayptemrrtie ambiguity.</Paragraph> <Paragraph position="1"> The essence of decoupling is allowing roprersntational etructures to have significance quite indtlpendant of theirapplication, and that may lead one to a eituatMh lot essentially ditfstont frm that of the logician who simply asserts that ouch-and-much ie the 'right structuxel of sme sentence.</Paragraph> <Paragraph position="2"> The inferences required to resolve word aense ambiguities, and those ad tb reaolva pronoun reference pxobletast are not of different typos1 oftan the two pmblaas occur in a eingle sentence and must be resolved together. But Chatniak's decoupling has the effect of completely separating these two closely related liniguistic phenomena in what seems to me an unraallstic aanner. His system does inferencing to resolve pronoun ambig2 uttfes, while sense ambiguity is presumably to be done in the future by sapre other, ulti.mately remupled, syste'~.* Wodulari ty Madularity concerns the deccwposability af a firogran or system into (interacting) parts, and fhe nature of the relationship between t+e parts. Winograd's program, as we saw, contains syntactic, semantic and deductive BegmentJ which interact in a way he describes as 'heterarchicl (as oppo$ed to 'hierarchic8) which means that different wents can be in controlaat dif foreht tiswc .</Paragraph> <Paragraph position="3"> Qn the other hand, ScW and Wilks have argued that it is not necesaaty ta absarve efther the syntactic-semantic, or the semantic-deductive, dlatfnctlon in an understanding program. On that view there 0 no part;icular,virtue in integrating syntax and semantic rbutlnes, since +here was no need tm separate them.</Paragraph> <Paragraph position="4"> Charniak, h~~verr wbld argue that, in same sensg, one should makq a syntax-setplantics distinction here if one c+n.</Paragraph> <Paragraph position="5"> This would be consisterit with his view on decoupling, and for him it wuld be convenient to decouple at a module, as it were, such as syntactic analysis.</Paragraph> <Paragraph position="6"> But decoup* Although Chaxniak would aque that sense ambiguity could be introduced into his system in its present fona.</Paragraph> <Paragraph position="7"> and s-ng modularity are not the same thing, Wineqrsldi1s progru, for example, is madulak but not at all ddcouplecl kropp SUX~~C?! trxt, Ava&lability of suxEaee at.ruct:ura An issue close ho that of the spproprlcrtp level of repreaer\letlun in a system i$ that aP the availabflity of! Qa surface sbuctum FP the language mcnlysedt or, to put it more crutlely, Ute availability during subsequent analysis of tho actual words &wing antlkysed, Tt~ee~s axe thtrerty available in ~olby, and ar~ indirectly availabh in S ns7, Nkrxq~ad'~~ and my awn system* but Schank mkos a pulrat uf Uaa iiqwrtance af their nmavailability, on the grounds that an ided r-epresentalhcn skmuld be totally independent of the. input surface structuxta and wrds, There axe Sxrth theoretical and pxactf cal aspects to aims claim sf ScbcEk %: f r. the lhf t , the osder of the sentences of a text is part aPS its surface structure, and pres\rm&ly it is not intended ta &andon this 'superficial inf~rmaticn' In one of his recent papers 91974bI Schank sems to have accepted sme limitation on the abandonment Of surface structuse.</Paragraph> <Paragraph position="8"> The other, practical, pint concerns the form of representation employed: in the (1973) hnplementation of Schank's systffi using an analyser of input text, a matwry and a generator of responses, it was intandd that nothing should ha transferred fxan the input program to the eufyut pr~yriuw -cept a rapresentatian ded in the structures sf primitives discu:::dl earlitlr,* The question that arises is, can #at structure specify and disting~ ish word-senses adequately without tuansf erring inf~m~ticn spcifically associated^ with the input word? Schank clearly believes the answer to this question is yes, but that cannot be considered established by the scale of cmputations yet described in print.</Paragraph> <Paragraph position="9"> A suitable envir-ent in which to consider tke question is that of translation from one language to another: suppose we are analyging a sentence containing the word 'nail1 meaning a physical object. It is clear that the translation of that word into ~rench should not be the same * '~nis point is to some extent hypothetical since, as we saw, Schankls conceptualizations still do ccntain, cr aspear t~ c~ztain, 3aF.y surface items; in particular nouns, adjectives an3 adverbs. Iizwever, tLis is a transitional'natter and Lley are in the course of r'epkace.zezt, as noted, by non-superficial items.</Paragraph> <Paragraph position="10"> as the translation for 'screw or 'peg1. Yet is it plausible that any dascription of the function of these three entities entirely in terms of arawmtf~ prWtima1 ma without any explicit mention of the wrb name and its connection to its French equivalent, will be sufffcienG t43 ensure that only the right match is made? Blication 'Shis pint ia a ganeralioation of the last Lwp, and concerns tho way in which differant ryutaa8 display, in the etzructures they manipulate, the actuel. Ipracdltres of application of those structures ta input text or dialogue, null 'Thiri is a matter dlEferent from computer implementation of the aystm. Xn the case of Colby's patterns, for example, the form of their application to the input English is clear, even theugh the @a&hing involved could be achieved by many different implementation algorithms. Xn the case of my awn system, I hold the same tio Be true of the template stzructurres, even though the time the input has reached the canonical template form it is considerably different from the input surface structure. The system at We extxeme end of any scale of perspicuity of application is Wincgradls wheke the procedural notation, by its nature, tries to make clear the way FR which the structures are applied. At the other end are the sys~QSS of Schank end Charniak, whaxe no application is specified, which means that tha regrslrrentatfono are not only cmpatible with many hapiementation aLporitba, which doers mtmtter, but axe also compatible with many syste~ls of: Ilnguiskic NILS, W~IOO~ specification is an essential piece of inquiry, and whose subsequent production may cause the basic system to be fuhdamentally dFf Perent.</Paragraph> <Paragraph position="11"> Application is thus different fram decoupling, for SChankls system is clearly coupled to language text by Riesbeck'sqarser, though his stzuctures do mt express their own =lication to language text.</Paragraph> <Paragraph position="12"> English pxepusitions will serve as an example: in Schaxk's case notation there is no indication of how the case discriminations are actually to be applied to English prepositions in text.</Paragraph> <Paragraph position="13"> So, for example, the preposition *in1 can correspond to the containment case, time location, and spatial loeatioqw amow others. As wa saw earlier, tiha B1serhinatian rnvolved in actual analysis is a matter of apciEying wry delioat. srmtic rulss ranging ovex the basic atatahtic otxvctursw the structures iLlEd case aytltem thmseL~as B~BLP~ to mh to bg leoaantlally dependent on the nature and apphicabiUlty olP such EU~~S~ and as this application of tlae sy8tem should have an obvious place in the ~Ve~aLli structuxas, It is nat sawthing to be delegated to ta mra ' impl9s~entation' It epugh of the linguistic intractablasfi of English analysis we= to be delegated out of this segmtlentatiun, &.I, muld be uffexlny no more to the analysis of nature1 language than the hgiciarts tllho pmEEer the predicate calculus as a p]rausibLe strircture for English.</Paragraph> <Paragraph position="14"> In sane of his -re recent writing's 'Minograd has begun to develop a view that is considerably stronger khan this 'application1 one: in his view the control structure of an undkrstanding progran is itself of theoretical significance, for only in that way, he believes, can natural la-Forward inference great outstanding dispute perspicuous.</Paragraph> <Paragraph position="15"> whether one should make massive forward inferences as one goes through a text, keeping all one's expectations intact, as Charniak and Schank hold, 0s whether, as I hold,, one should adopt some 'laziness hypothesis1 &ut understanding, and generate deeper inferences only when the system is unable to salve, say a referential problem by mre superEicia1 methods. Of, in other terns, should an understanding system be ~roblern-, or data-, driven.</Paragraph> <Paragraph position="16"> * This is not meant to be just bland assertion. I have written at same length on the relations between application and the theoretical status of linguistic theories in (Wilks '74).</Paragraph> <Paragraph position="17"> **The differences between Minsky's (19741 notion of 'default value' and what I have called 'prefexence' can be pointed up in terms of application. MhsQ suggests 'gunt as the default value of the instrument of %he action of shooting, but I would claim that, in an example like the earlier 'He shot her with a colt', we heed to be able to see in the structure assigned whether or not what is offered as the apparent instrument is in fact an instrument and whether it 'is the default or riot. In other words, we need sufficient structure of application to see not only that 'shcotlng1 prefers an instrument &at is a gun, but also why it will chaose the sense of 'colt1 thatcis a gun rather than the one which is a horse.</Paragraph> <Paragraph position="18"> ATtlx>ugh Schank sametinee writes of a system making 'all possible1 inferences a8 it p10ceBd8 though a textt this ie not in fact the heart ot tho dispute, since no one would want ta defend my atmng definittior oL the tom 'all poesibla infetences'.</Paragraph> <Paragraph position="19"> Chacniakqs argument 4s that, unless certain fornard inferences we made during an analysis ofr say, a e-ry - forward inferencest that is, that are not problem-driven; not made in rerrpcnss ta any particular problem of analy.ysia then known to the ayrtsm - than, ar a matter of empirical fact, the system will not in general be able to solva srPbiguity or rofetence ptoblems that arise later, because it will never in fact be possible to locata (while looking backwards at the text, as it were) the points Ohere those forward inferences ought to have been made. This isr in very crude summary, Charniak's case against a purely ptoblw-driven inferencer in a natural language understander , A ditficulty with this.argument is the location of an axample of text that c~nffrms the pofnt in a ncn-contentious manner. Chatniak has found an excerpt ftcm a book describing the life of apes in which it is indeed hard to locate the reference of a particular pronoun in a given passagQ. Chamiak's case is that it is only possible to do so if one has made eertaln inon-prublm occasioned) inferances earliez in the story. But a nuabet QE readers find it quite hard tb refer that particuXe pronoun anywayl which might sweet that, the text was simply badly written.</Paragraph> <Paragraph position="20"> Another difficulty 16 that it is not always clear whether the argument is about what pple are thought to do when they understand, or about how one should mnstruct an wdexstandLng system.</Paragraph> <Paragraph position="21"> This is a difficult matter about which to be precise: it would be possible, for example, to agree with Charniakts argurnentmd still construct a 3urely problem-driven inferencer on the ground that, at the mment, this is the onlv way one can cope with the vast majority of inferences for understanding, since any system of inferences made Fn response to no particular problem in me text is too hard to control in practice. Indeed, it is noticeable that the mst recent papers of Schank (1974a and 1974b) and Charniak (1974) have been considerably less forwardinference oriented than earlier ones.</Paragraph> <Paragraph position="22"> This Bispute fs prhaps mly one of degree;&nb about tha posalbility of befining a degree af forward inf~rlene~e that alds the adutisn of later semantic problem without going Lntta wwe~srry depth. This right be area where paychelogicab invastrigatloras wukd be eE F~'IO.~US hdp It9 'The,justflication of systems -= Finally, one might useZuLly, WwpA bslaf?Ly, curntxast ths Shffarent des of JustiPicathan iolplieftly appealed ta by the syst:.~~~s deucxikd earlier hn this paper. These seem to FIM? to &u:e to EOW t (i) Tn terns of the pwex of the inferentiak syskea enpisyPd.</Paragraph> <Paragraph position="23"> This fama of j.ustifhcation has underlain the- earLy predicate caXculus-based language programs, and is behind Hayesr (1974) recent deaand that any foxmalism for natural language analysis should admit of a set thecttetic s~tics, in the Tarskfan sense, so as to gain 'Intellectuai respectability1, as he puts it. The same general type of justificatim is appealed to in sane degree. by systems8 with BLMNER-type f~xtmlikas.</Paragraph> <Paragraph position="24"> (ii) In terms of the provision and formalisation, in any terms including Elqglish, of We sorts of knowledge recpirea LO understand areas oe ~~SWUS~.</Paragraph> <Paragraph position="25"> (iii) In terns af #e actual performance of a s):st~m, i~plementd on a ckmguter, at a task agreed tu demonstrate understanding.</Paragraph> <Paragraph position="26"> (iv) In terns of the lfnquistic and or psychclcgfca~ plausibility of the proffered system of representation.</Paragraph> <Paragraph position="27"> Oversimplifying considerablyr one might say that Charniakls system akpeals mostly to (ii) and somewhat to (i) and (iv); Winogr3S1s to {iii) and scmewhat to the other three categories; Colbyls (as regads its natural language, rather than psychiatric, aspects) appeals almost entirely to (iii); Simmons largely to (iv) , and Sthank's and my own to dif f ereng mixtures of (ii), (iii) and (iv) .</Paragraph> <Paragraph position="28"> In the end, of course, only (iii) counts for enpiricistsu but there is considerable difficulty in getting all parties to agree to the terns of a test.* A cynic might say Chat, in the end, all these systems analyse tho setltenres atit they analwe orl to put the same point a little more Wwtetically, these is a sense in which systerms, those described here and tho st^ elsewhereb each define a natllcleal, languaqe, namely the one to which it applies. The difficult question is the extent to'which those mnv and mall natural lacpages resemble Engli~h.</Paragraph> </Section> </Section> class="xml-element"></Paper>