Speech Understanding and AI/AI and Speech Understanding

SPEECH UNDERSTANDING AND AI

AI AND SPEECH UNDERSTANDING

Dono.id E. Walker (Session Chairman), Lee D. Erman, Allen Newell, Nils ). Nilsson, William H. Paxton, T er'ry Wino14ro.d, William A. Woods

This panel will review reseilrch in speech undNstanding (SU} and in <Jrtificial intellir,ence (AI) from two perspectives:

The contributions that AI has made to SU the resources in 1\l that have been used in the development of SU c.ys tems. ·

The contribution~; thai SLJ has made to AI ··- the results of the SU prog1·am that have affected or are likely to Affect future AI rP. ~.ea rrh.

Four topics are identified lor major c.on~ ideration:

·Multiple l;ourcc's of l\nowlf~ dge ·- -- which are required; how should \hey be organizccL

System control how to manag-e the complex interactions involved.

Lanr,uage under sl anding- - comparisons of text and speech input.

Organization of research --creating complex, multisource, knowledge> -based sy~t el)"'s.

I_N~gQQ!)~:TJQ~ Donald E. Walke>r, .SRI International

The 1\dvanted flesearch Projects Ag'ency ·of the Dep;~rtment of Defense spo.nscn·ed a fi ve year research program with tile goal of developing a speech understanding sys tem capab le of -eng<~p,i ng a human ap0.rator in a n<~lur<'ll conversation concerning a ~. pe>cific \ask domain. Among the juslific<~lions for pnderlakinr, the entcrprise in 1971 were the developments i-11 arli fi t ial intelligence taking place at that period. Thf! culmination of thi s program in 1976 with the demonstration of a system thai met the target spPci fi.cations set out <'~I the bcr, inning I provides an opportunity for iln overall review of the contributions of artificial intelligence to speech understanding research and of the implications for artifi cial inlelliE',ence of the research on speech understanding.

REFLECTIONS ON OUTAINING SCIENCE· ---- ---Ii:i!~b!:!~~ltili!Ltlli:' ci- s~~-r[~i~ . ~-- Allen Newell, CilrncE',ie· Mellon University

Each sc ience builds its own methodology, primarily by pr art ice rat her than preilc hinl~- However, ~ometimes reflections help a lillie. These · are purely my personal opinion.

M<~rk F. Medress, et a\., "Speech Undersl<~nding Systems, Report of a Steering Committee". SJCART Newsletter, April 1977, 62, 4-8.

One element in AJ's me thodology is lhnl progress is made by building systems that perform: synthes is before analysi s. Why we have . lhi <. rnelhodo logica·l tenet is not far to seek, or why other sciences do not, giving primacy to analysis in various forms.

1 he ·rec:en_tly conclude>d _ARPA SUS program is an import ant dal.a point in a';~ess inE'. the synthesis methodology. My purpose here is to state the data point .(but not argue it) and describe some of it ~; condi t.ions.

• Q I. The question: Cnn sc ienlific progre>ss in AI be made by building perform~nce systems? The increment of evidence from the SUS prog1:am is definitely yes . ·

Purs~ling ·thi s (over brief) answe1' leads-·to other _question-;.

• Q2. What is science in AI" · II' is 1\nowledge - -- theories, data, evaluations - - thai descr ibes the means to re<~ch a class of des ired ends given certain strucl\.wes iJnd situations. Sciente r_eaches beyond the si tuntion of its generation and becomes a sourre of r,nowledge for . future scientists nnd technologists inves tment rather than consumption. Knowledge of mean-s:_ends relati·ons is what characteri zes the artificial sciences, of whi ch AI is a p<Jrl.
• Q3. Wbal then is the science that carne from the SUS ef fort? This not e can only surnrnilrizP- , !. in<e it s purpose is to iocus on the condiliom'. Other pane·\ members must provide the details. Spenl\illE', brondly, we now have (1) the knowledge from which lo build a techno\OfSY of limited SUSs; (2) a wide sc nl lering of knowledge nlong the way fo higher Cilpability SUSs; ;md (3) several impodant additions to the general stork of knowleclr;e about perceptual systems.
• Q4. What about the effort ndunlly permi tted the results to be obtained'! (I) The performan~e specifications. (2) That several groups were at tempt ing the same go<~l, inhibiting unilalcr<~l judp,mcnl s of unal\ainability. (3) The scientific cnliber and gonls of the mnjor scienti st s involved. (4) The intensive cornrnu11'ily.
• Q5. Would these re:;u\ts have been obtained anyway if the funding had been (.pent in i1 s t ~ndard dyle? Not at all. other results would have been, no doubt. But the base level of the field would not have been lifted.
• Q6. How · did the stri v ing lor performance yield the science rather than detract from it? In forcing attention to aspects that reall y made a difference to performance, and to performilnce feedback to indicate .what was really true. Compared to other AI efforts, the SUS effort was data and analysis rich.
• Q7. Didn't the performance focus distort by deemphasizing speech re .~; u\t s and inhibiting the . front -ends from

being much bel ter'l Limited resources imply chpices: Better front -end~. would have mcnnt wors·e total systems - -· and a loss of the scientific knowl edge of what total systems are sufficient to yield what performance ahd in what way. In the artificial sc iences, sc. ienlific knowl edr;e does not r.eside in the components nlonc.

• Q8. Didn't the r erformance· focus distort by making everyone I hrow sy~;tems tog<.'ther to meet. the specs? The deadline flnvor of tho last six 1'non\hs of the effort was unmi stakabl e. Yet, the sc ientific yield from that s;~me period was very hi8h. Gi,;en the conlini.1atio11 of the program, the salvage value would .have been even hip;her.
• Q9. Didn't the performance focus inhibit sc ientific cooperation and prodt1ce too much competition? Yes, that seemed. to be a cost, though it was kept under moderately good control.
• QJO. Can thi s technique be used in other !·ituations'l Several unique conditions obt ained for ·thi s effort' that seem lo be · critical. (I) There was a single performnnce system (spec s) which several p,roups could ilccerit a~, their operating goal. These sp0cs hnve ex t r r~ me f(lcr>. V<J iidity and h;we survived well the inevi.l able snipif)g. (2) The. groups that accepted the task Wl're not being redirected, with a strong desire tn continue. to do their old things under new lnbel s. ( ~l) Sub5lanliill ;unounl s of local ilulonomy <Jnd stability were granted to the effort (at the end thi s faltered, lwl too late to kP.ep the· progr<~m from 5Urc eP. ding. ) The SUS organizalion<JI paradigm was pilrfiillly atlcmplcd several times, but these conditions· were not s"l ti '.died ;1nd the organization did not work. These failures refl ect limitations on the scope of the organi;aational tec hnique, not on whether atlempting ·performance goals is a succe~.sful methodological tenet for AI.
• Qll. Would I favor using the technique again to generate more AI science? 1\b:.olutely, if the conditions were right and one could Iinke!' with ·the arnmgemcnts a bit..

'!l/H~ IJ:JL\Y.LWI:;_J~l\Bf211:..Q1 Lee D. Erman; Carnegie-Mellon University

would like to concentrate on those experiences of the recent speech lJnder51.anding rese<JI'.Ch that might be relevant to other areas of AI.

Speech undersl anding is a domain with " highly· variable input signal, and it ·requires multiple sources of knowledge which are not w.ell enough understood to be specifiP.d accurately. Thus the problem solution of. necessity must include ·errorful and approximate proc,essing. ThP. requirement of bnilding systems with specific ·performance level s insured that these problems could nof be ignored. The performance requirements al so provided relatively objective measures of success; these were especially· useful in helping select ilmong illl ernative designs.

These projects extended over fi ve (or more) years. It was realized from the start thai several iterations of system design and implementation would be needed. An import ant aspect was keeping each · i1eration balanced with respect to difficulty of the task (that is, constraints provided by limithing vocabulary, syntax, number and types of speakers; quality of the signill, e·l c.) and i\ <, eff ect on performonce (ac<uracy and speed). When sys tems became unbil lant.ed, usuall y because of openinG up ~;ome constn1int precipi1ously, p(>rformance dropped to such low level:; that it was difficult to anal yze the system and Sll [;(',est directions for incremental imp1·ovement.

A .r.tronc emrha~;is has been on systems design, resulting in system organi zations less ad hoc than is usual. At lc.asl two of these (f)rnr,on/Harpy ·and HPill'!' iiY) are being.explorr>d in other domil ins (includinp, · image unders tanding, signil l interpret ation, protein crystallographic ilnalysis, complex learning, and modelling of humiln reading) <Jncl thus appear to be somewhat GenP-r al.

The Harpy. syf. tem rai ses an iniHresting issue for AI. It s organization, i11 whi ch all r,110Wiedgr~ i<, flilltened into a single, simpl e, common gr ilph repr· ps·e nt e~ tion, and it s ~<e ar. r. h method (the "locus" lnethod), in which the gFaph is m<~tched ngainst the .input d>11'a u·, ing il heuri stic form of dynamic progr<~mming, are rildicall y different from those u!;cd in most othrr /\1 syslroms. The succr,·ss of Harpy force,; the field to understand thi s ilppromh. A few other rc.rent f, ucc.essful AI syr.tem~, (i.r., the Northwe!; l er n Clwss ll.5 ·program and the Univers ity of Pitt sburgh Interni st · diilg no~ ti r prop,r11rn) sc.em. to hilve simililr characteri sti cs, and thus ilcc.ent u'l l e t·he poss ible in)por!ance of doing il large ilmounl of ~.eilrc h over a · l <~ rge but sirnply­ represenled knowledge structure.

The need for effi c. ienl systems while un.der deve lopment (for expNimentalion) 11nd the need for iterating (both on the systern design and the irnplernenlalion of the knowledge sources) led to sr'n';ilivr fll·oblcm'. in r;ystem construction ilnd · maintenance. Tho experiences with lleal:sily-IJ, in particular, havo shown I hill modern techniques· of structured progr ilmming c11n bf;' used to implement (and reimpl emenl) complex AI sy~ tPms in a r e ~.earch e.nvironment, ·striking reasonable bnlanc£Yo, ilrnonp, c.ffic iency, fle:>< ibilily, and user (i.e., rPc.earcher) conveni P.nce. !1) pi!rlicular, good c.ystern implementation dt'~ip,n led lo lhe 8bility · to reimplemenl selec ti vely tho<,e ac;pecls of the sys tem · found to be too inefficient without adve1·sely effecting other part s of lhe system. Thi ~ WiiS compatible with the conceptual dr<~tegy of ~tarting with il very general de~ign for the system <~nd selecltvely specializing .il <IS the need WilS discove-red,

One development c.omlt101' to most of the systems is the separat·ion of the problem-·solving slrillegy from the clomoin­ specific (i.e., speec h) r,rrOwedge. (For example, see the three papers b'y Hayes-Roth and Lesser., Paxton, and Woods in these Proceedings. ) These represent some of the first steps of being able to specify complex control striltegies in a clean way. ·

Another common development, which contr<J~ts with many other current AI efforts, i5 the compil8tion of porticular kinds of knowledge into forms that are spec ialized for applicillion but' which may be very different from their "natural'" external forms (i.e., different from the way in which we specify the knowledge, to the system in the fir~, t place). In some cases the same knowledge is transformed into several different forms, each appropri'ate for a different kind of application.

The use of "semantic" or "pragmatic" grammars, which contain semantically oriented categories instead of the (fewer number of) syntaCtically motivated categories of conventional grarl'lmars was highl y eff ective for the most successful of these -;y~. tems . This development parallels similar development s in ~.eve ral other current niltUrCll langu::1ge efforts (e.g., see the paper by Hendri x in these Proceedings).

SPINOI:.f_~OM _5_!:'~[Q·U}12JfJERST ANDING RESEARCH William/\. Woods, £3oll Beranek and Newman

The focu~. of my prcsenlation wi ll be on what one can ilbstr<Jcl from thP f!Xper·ior\t f.> of lhf! Speech Undf!rstanding Program that can have generill applicability or at least sur,ges ti ve direrlion·. of appro ::1c h for other Artifi cia l Int elligence problerM . spec ific ally other high- level pPrceptuill ta ~;l-.s ~. ur h <~5 vhral sr:em! i~1t erpre tati on and the anillysis of -dialog ;ond di sc our ~.e structures in naturetl langrr ilp,e, lo~f'the r with thP early robot v isi on pi'Ojeds, the SUf~ prOject rAnr,s as une Of \he few instanc es where a tot al sy~. ! Ptn has been conc.tructP. d <Jnd faced with rP.al wor ld dill~ that has not been ab<dri1dr~ d ;,nd simplified to eliminate noi se and make tlte pro l :i l <~tn ea·.det. Although a great many things were learned during this pr·ojcd about the low leve l signal char~cterislics of e. p<>Pch ·:;ounch, I do not expect these to have p, rca t carryove r inlo other .1reas. However, in the techniques for inter! ac ing high !(-·vel hypollw~; is formation and evalu ;~ tion lo · such low·-levc l s.OLwc cs, I believe the speNh tmders_t ;mcfing w·orr, has m~d.e some generalizabl e contribul ions.

The m;,jor two ~- trch contributions, in my opinio"n, are \he c.)(plicit e~p lor<'ltior1 of control str-uctures And s\ra tegies, Clnd the. demonstr nt ioi1 of the I~OWf'l' of what l will call fo.ct orerl knowle>dge r. t r uclure ~ .. In tlw fonr1cr arecl, I thinr, that the discovery of the dcn,_,i ty c.coring strategic~ h;~s interesting consequences in the are<~ of ~;eM-ch and optimization, since it represent s a new technique no·t subsumed by the A* algori thm.2 In the l <~t l e r Mea, the pervas ive use of vari ous f~ctOI' ed knowledge re pr· c ~. e nt a tion structures, in which 'the common part s of many different _pal terns or schema! a are me rged, has :;ignii iumt impor t for the pr·oblems of generalized percc·ption and knowl e·ci ((e repre$entation.

Ely fil ctorPd ~; trutlurc·s , I ref ~·r to suc.h structures ils the phonet ic segment 1.1\ t icc ilnd the tree st rue! ured diet ionary re>pre·;ent alion·; thai Me used in lhe BBN HWIM sys tem, the trer. c. tructur<:>d g r~rnrnilr u ~; ed in CMU's. Harpy sys tem, ;md the /\TN r,ran111lill. f ornr<11i ~. m used in rnnny ~ystems by now. In general, a fildored rcprp·.entation is any l~nowledp,e ~;\ructure in which cornmon part s of different r,nowledge elements 11re merged in such '' WilY that retriev ;;~l processes can 11cc.ess them inc.rernent illly to create more and more specific hyp~thPse s as >1ddi lional d il ta or measurements on those data are obtained. The ~;irn.p l es t examples of such structures are

deci ~.i on trees or di r.c rirnination nets.

Such representat ions nre used many places in the BBN speech sys tem to organize its internal information about alternative theories, and io. of cour se the major organizing principle of the Harpy system at CMU. If one shifts to the contex t of fr ilme .. ba~.ed language understanding systems and considers the problem of determining the f.rame that one should be in nt any given. poi tit (in a sydc rn that · contilins thousands of such tramPs, ;my numlwr of which might match . son1e initial pori ion of a dialog), then it see.ms cl ear th <~t sirnililr such fact ored ~;\rutlures can be use·ful. Thi1 t is, one would like to have internal sl nics corresponding to the results of sequC' ncc~; of meiJ~; ur· c .. ments on the input stimu li th;-11 constn1in the po~;s ibl c interpr-et ati ons of the input -without enurne ra tin L~ nil of the po~;si bilitics explicitly. These stal es can then indk ,1\e fur ther measurement s to be made on the · input , and transitions to new inter nal stat e<; corresponding to more specific hypotheses that can be made as <1 result of such measurements. · I expect thi s kind of f actored structure to have more ilnd more application in artificial intelligence especially in v ision. and . niltural l<1nguage underst ;111ding.

2 See my paper, "Shortfall and Density Scoring Str<1legies for Speech Un.derst anding Control" in this conf.erence proceedings.

Other a5pect s of the spccc.h unders tanding systems that I think will have wider applirilbi lil y include the use of Bayes ian probfl bilit y eslirn<J tns to cornhine informa tion from different knowledge source~; , the use of analysis --by-synthes is ver ification a~ a ·r.ource of informiltion in percept-ion systems, and the clevelopn1ent of middle-out parsing <!lgorithms for ATN and oth~ r phrase -structure grammars.

EXPmiMENTAT ION IN ARTIFICII\L INTELLIGENCE ---.. - \viili~r-11-H~· Pa~i·~~. SRI- l;~t~~~~~ i~~-;j3 _____ --

My pre5ent ation will locus on · the c.ontributi.ons of the SUR program to AI methodology, in part icul ar on the value of e)(periments ·to aiel in underst;,nding the effects and inter actions of sy~. l em design fe il t Llres. The various speech sys tems are l i1r ec ilncl c.on1plux. It is often difficult for the designers thcrnsc lves to. ltndcr ~; t and the operation of the sys tems, 11nd lr <~di tion <JI te<hniques such ilS traces of sample runs are of little usc IJ:cr.;n.r~e of the compl ex it y. of the control slr ;, tegieo. and tl 1e varia ti on ill\1ong utterilnCP.S. A pMticular system feature m<~y improve pr.rforma1;ce in one case but make>· it wor !;e in a110llwr, so judt;ments abo.ut the value of a feature mu~:t not be ba5ed on intuition or casua1 tests of a. few sPntence!;. · Mon.·over·, the comple~ ity and size of the sys tems make analy tic methods of little uso.

These consideri'llions lead us to ildopt an experiment al ~pproach, but what Y.ind o! e·xper irn,.. nts should we perform? ~lo.w should we Cil iTY orrl 01 n experimental study so th<11 it will help us to under stand how the sys tem works - - help us to see which design fea tur·c:' arc import ant and why they have the effects they do? . The .i;ydem de!; igners will of cour se ·have ideas about which il i'C the import ant features, and, if they hilve done their job well, they will olso know what the main alternative~ ill:e. For exarnple, in most of the speech sy$\erns the cle ~ i gner s felt that it was import<1nt lor the sys tems to be able t.o islr:wd- drive, that is, to construd interpretations star ting with words found anywhere in the input. The alternati ve to island-dri vi ng is a more constrained cont r- ol striltegy ~· urh as strid left - to - right proces·sing of the input. In this instilnce, a simple method to determine the effect of isl and-dri vi ng is lo look at the difference in performance with island -dri vi ng versvs with left-to-right process ing.

Such comparati ve . tes ts are often possible for the major rles ip.,n features, p~rtirutarly if the sys tem is constructed with testing in mind. As <1 11 experimental ;~pproach, compilrati ve tes ts have scvet·ill a ltr;~ctive <~!tributes, indurling independence of absolltle perlormilnce leve ls and compa tibilit y with powerful stati stical methods. Co.nsequently, we su~;ges t tl~at !'uch l.e«ts should be a stand·ard technique. to aid in understanding complex systems.

3 Now at Xerox Palo Alto Research Center.

Since the <> im Of the tests i<; to understand how the sy~; t ern work,; , ;~nd not just to optimize performance, auxiliary me<>surements. of ~.ys l en1 operat ion rnus1 be made in addition to measurement s of primary cri teriil .such as speed and accurilcy. Ag<~ in, we rely on the dc~;igne r's ~.nowledge of the system to decide what· to mea~ ure so that we witt have the necessary inform.ation ava il able to exp l;~in the ob.served results. In a speech unders tanding system, auxilia.ry mc<~sures would include " var iet.y of things such <IS number of cor rect and incorrect wordr. accepted nnd dorage usage. These auxiliary mc;~sLII"c-s provide the intermediate steps in explanations of the effeds of the sys tem features. Thus, lor exo:~mp le, the ef fect of i~,l~nd -driving on system accuracy might be explilined by reference to its · effect on storage us <~gc in ~onjum. lion with information . about the · relation between storage usage and accurilcy .· ·

Typically, there will be several design features that ·are believed to be importilnt, and it will be desirable to . test the features simull aneOLI '·IY to see how they interact: Moreover , the f e11tures wi ll u~uill l y hilVf' good elfec: t·!. in .some casP.s and l;ad ef fect s in ol het·s. The stati stical method for dealing with suc h ~ sihtillion i~ caHPCl <lnillysi~, of var iance. This ·technique makes it possible~ to compute the probability that obser ved effects ;,nd intNadions ;1re rcillly C(IU~;ed by the experiment al variables rather than by chance. We will briefl y illustrate the 11se of thi '; technique by ·sketchin~ some of our experiments on control strat egy destgn cho1ces.

In conclusion, if it is worth building il system as part of an AI reseilrch project, it is cedainly worth ma~,i ng an effort to undersl and how I he sys tem actually works, and experiment<1tion i ~, iln import<>nt techniqLte ·for doing thi ~; . Simply d~rnons lr ;~ l ing a work-inp., sys tem should no longer be . enough; let us begin to demilnd !hill the AI systen1 designer specify the supposedly ·import ant sys tem features and their alterniltives, do the · experirnents to show .the features' effects, <1nd provide explanations of why the features have those effects.

LANGUAGE UNDERSTANDING AN[) REP8~ENTAT!ON -·· ·-·-:[~-;.-;.·,; ·wir~~g~~~l.·-st·;;;~orc.JLJt,iver sity

Although I h;we not been directly involved in research on speech unders landinR, I have followed the work ci05ely, since I bel ieve that · it is a harbi nger of things to co·me in AI research in general. The speech projects are the first major AI efforts whi ch have ploced primary emphasis .on the system organization required for making . ur.e of diverse sources of knowledge in a t ask whose structure defies simply structured programs. Over the nex t few years, other AI researchers will be~;in to attack problems which demilnd this kind of robu~tness - - the ability to come up with an ilnswer when the input daln are messy, the combinatori11l possibilities are explosive, and. high leve l knowledge Ciln b() of great influence in determining the answer. In programs for vi·s·ion, scientific anillysis (a Ia Dendral), and langu11ge, I foresee a shift in thi s direction.

4 Further details are given in Walker and Paxton, et al., "Procedures lor Integrilling Knowledge in a Speech Under standing System" in the proceedings of this conference; a full description of the experiments appears in Paxton, A Frnmework for Speech Understanding, Ph.D. Dissertation, Stanford University, 1977.

In programs ·for c0111prehcndi ng li!nguage ·-- even those !hilt deal with text -- we need to 1~1ovc tow;Jrd handling more natural inputs, wilh ;~II of the inaccuracy, incompleteness, and ill-formedncs~. we have long ignored. This will demand program and knowledge organiz~tions I hat are based on ideas !hill I he speech work has begun to explore : · multi-process communication; hypothesi s formation and verification; the intermixture of goal-dri ven and event -dri ven ·processing; careful att ention to the interlace lo.nguo.ges lhnt make it possible to give meaningful slructur·e to the communication between components; and the importance of performance evaluiltion tool s thnl help us m~~.e sense of whitt is hilppening in a complex multi - process environment.

In the panel, I would like to . see a di scuss ion along two · lines: Whnt ·are the m<~jor insights to be gained from the speech pr.ojecl experi ence !hill can be of use in orgai1izing other AI programs' 1 and what «re the obvious gaps to be filled in the next round of experiment alien? Since my major current conrern is with representation languages, 1 would like to I ake part in a discuss ion concerning the ·problems I hat were encountered with the representational systems available for speech work, and the features that wifl be important for systems of the same degree of complexity in the future.

UNDEii~J~~i':!�_R_f~,_EARC:.!:i Nil s J. N.il~;son, SRI Internationi!l

Artificial Intelligence has recently completed · an extensive and coordinilted cxplornlion ihto the terra incognita of l;~rge scale, knowledge -based sys tems. To the brave and resourceful t~ xplorers, we stay - ill - homes must say, "Congratulations and well -done! We enjoyed· your slide shows and marveled nt your specimens". BLil when the celebrations are over it will be important for ·US all to digest the new knowledge uncovered by these. explor·ations.

We need to ask more than "What have ·we le·arned?"· It is too ternpling lo answer that question using our current vocabulary. We rni ~; ht for example filii into adopting .the rough anrl ready fronlier pilrlance and metaphors of the explorers themselves a·nd start speaking about "multiple cooperating sources of ~.nowtedge ", "blilckboards", and "island growing". Or we might illtempt to describe the: new· vistas with older find perhaps inadequate phra~,e.s such as "rule­ based systems", "left-to-right· parsing!', an.d "heuristic search".

The ques tion before us, I think, is hard~r than "What have we learned?" "it is "How are we going to express what we have lparncd?" A major eX"pedition just completed is too precious .an occasion to let pass heralded only by a·ccounts from the explorers. It is an opportunity for . attempts at synthesis and for i.nvQnt ing new concept s and new par adigm~ .. We ~hould not he di !;couragt·d merely · because there i ~; no gUMilnlee that these attempts will be SUCCessful Or because the odds ;,gai nr. t useful new pilradigms are ~lways high. We have just spent about 100 rnan -years on exploring. · We can afford to follow this up with a few man-years of thinking about how to say what we have learned. ·

As good a!. they are ill speech understanding, it is unfortunat ely true that 11/\RPY and HE/I.RS/\Y and friends cannot speak f·or them,,elves. The major product o·f the 100 man-years is not the rot<JI body of code that was produced nor is it what ·that code accompli shed in the demonstrations. The memorabl.e output, what can be taught to future generations of students, will be a description of that code. It is not even necessary that the descriptions be completel y accurate. Simplifications iltid even fabri cations ilre justified if they have pedagogk va iLte and do not over ly · mi slead poster ity. My major poi nt is that it is important . !hilt these descr iptions be eler.anl and thai they have a certain, hard- to­ define, estheti c ;,ppeill so that they will be memorable, easy to use for leaching purposes, and provocative for 1 he design of new systems.

In creating the kinds of descriptions !hill I thin~, will be important, inventive talent will be more important than reportorial !>~.ill. Suppose, for example, lhill one could invent some irnaginnry sys tem that was something like one of the · adu;,l . speech unders tanding sy<> l erns but different in many detail s. Since o"r imaginary sys tem doesn' t really have to run on a computer we can strip it of the various ad hoc features of real systems so necessary for efficiency. Now maybe .we· can reorganizP it a bit to give it a more coherent intern.al organization and to relate ·it more closely to existing well .. undPrs tood AI meclvmisms. Thet'e may be some tension in try ing to do thi s. Maybe the ex ist ing AI mechanisms aren't so well -ttnder stood or a!> general iiS we thought. Perhaps the effort of try ing to b.uild our imaginary system out of these mech<misms stretches ·them a bit. Maybe we'll be fortunat.e enough to think of a niajor generalization of some of these mechanisms to rnake the.m more useful for our ., fictional system. ·Now, maybe we'll reorgatii ze the fictional system some more and go through the loop again. Once in a decade or so, and if our interes ts are broad, we might notice that the new AI concepts just invented could also be profit ably used to describe the results of other explorations. At the very least our new synthes is will geatly simplify the process of designing new systems of ·a simililr kind.

These steps are important if a field is to grew into a mature !>c.ientific or engineeri.ng di sc ipline. Artifi cial Intelligence has to take several such steps before it can be as productive as we all would like it to be. AI has not ye>t really developed what could be called a set ?f universally adopted methodologies that can be (ollowed in the design of new systems. If six different AI laboratories were given the task. of building a nt.le-bo.sed .syste.m for some well -understood application, I would not be surprised to see several quite different designs. Much of the terminology used by AI people is still pre-technical at best and meaningless jargon at worst. Let"s try to use the plentiful and excellent experiences of the speech understanding projects to climb a rung or two in the conceptual understanding of our field.