METHODOLOGY IN AI AND NATURAL LANGUAGE UNDERSTANDING

! not know of even an a p p a r e n t l y simple example that it did NOT require r e a s o n i n g in order to be u n d e r s t o o d . "

M E T H O D O L O G Y IN AI AND NATURAL LANGUAGE U N D E R S T A N D I N G Yorick Wilks Istituto per Gli Studi Semantici e Cognitivi Castagnola, S w i t z e r l a n d

The above three positions are not i n t e n d e d to be a parody, and c e r t a i n l y not a parody of anyone in p a r t i c u l a r ' s views. I have not in fact heard all three from the same person, even though, in my view, they constitute a coherent position taken together: one which I believe to be not only wrong, and I will come to that, but also harmful. Let me deal with the sociology first, and in the form of a very crude historical generalization.

Are workers in AI and natural language a happy band of brothers m a r c h i n g with their various systems together towards the Promised Land (systems which in the view of mahy well disposed outsiders are only notational variants at bottom), or on +the contrary are there serious methodological differences inherent in our various p o s i t i o n s ? I think there is in fact one central difference, and that it is a m e t h o d o l o g i c a l r e f l e c t i o n of a m e t a p h y s i c a l difference about whether there is, or is not, a science of language. But it is not easy to tease this serious d i f f e r e n c e out from the skein of n o n - s e r i o u s m e t h o d o l o g i c a l discussions.

It is clear that "natural language understanding" has come to occupy a less p e r i p h e r a l place in AI, and much of the credit for this must go to W i n o g r a d (1972). The position, e x p r e s s e d in (I), (2) and (3) above, is in some ways a reaction to that, and in my view an excessive one. Behind the p o s i t i o n s above lurks the s u s p i c i o n that the success of W i n o g r a d ' s system was in part due to its o v e r s i m p l i f i c a t o n s and that we must now be wary, for a while at least, of applications, s u c c e s s f u l or otherwise: that we must, in short, e m p h a s i z e how difficult it all is.

By " n o n - s e r i o u s m e t h o d o l o g i c a l ete." I mean such agreed points as that (i) it w o u l d be nicer to have an understanding system w o r k i n g with a v o c a b u l a r y of Nk words rather than Mk, where N>M, and moreover, that the vocabularies should contain words of m a x i m a l l y d i f f e r e n t types: so that "house", "fish", " c o m m i t t e e " and " t e s t i m o n i a l " would be a better vocabulary than "house", "cottage", "palace" and "apartment block." And that, (ii) it w o u l d be nicer to have an understanding system that correctly u n d e r s t o o d N% of input sentences than one w h i c h u n d e r s t o o d M%. When I say n o n - s e r l o u s here I do not mean unimportant, but only that n o t h i n g t h e o r e t i c a l is in question; so that, for example, it could be only an arbitrary choice whether or not a s y s t e m that u n d e r s t o o d c o r r e c t l y 95% of sentences from a 3000 word v o c a b u l a r y was or was not better than one w h i c h u n d e r s t o o d 98% from a 1000 word vocabulary.

Now there is u n d o u b t e d l y something in this, but it seems to me that the r e a c t i o n may have the p a r a d o x i c a l effect of causing the study of natural language in AI to be given up a l t o g e t h e r . In the last year or two a number of those who seemed to be concerned with the problems of natural language no longer seem to be so. There has been a subtle change: from the analysis of stories, or whatever, to the s e t t i n g out of s y s t e m s of plans which now seem to c o n s t r u c t stories as they go along. It might then seem natural to move further: from the production of stories about tying one's shoe:laces, s h o p p i n g in supermarkets, etc. to plans, for robots of course, that will a c t u a l l y shop in supermarkets, tie their own shoe-laces, play d i p l o m a c y or whatever. And then of course we are back where we started in AI: back to AI's old central interests, robots, p r o b l e m - s o l v i n g and the o r g a n i z a t i o n of plans.

Indeed, the very sizes of the vocabularies and success rates in the e x a m p l e show that such a choice, however arbitrary, is not one we are likely to be called upon to make in the near f u t u r e , so let us press a little deeper.

All this would be a pity, not only because someone has, as always, to be left holding the baby of natural language analysis, but b e c a u s e it is too soon, and AI has not yet had the b e n e f i c i a l effect it is capable of having, and ought to have, on the s t u d y of natural language. There are at least four of these benefits; let me Just remind you of them:

C o n s i d e r the following three points, which I will name for ease of s u b s e q u e n t reference: (I) T h e o r y and practice: "Trying hard to make a system work is all very well, but it's too s u c c e s s - o r i e n t e d , what we need at the moment is more t h e o r e t l c a l work".

(i) e m p h a s i s on c o m p l e x stored s t r u c t u r e s in a natural language u n d e r s t a n d i n g system: frames, if you llke (Minsky 1974)

(2) AI a~d ~ i e n c e : "What we are after is the right set of rules, and e x p r e s s i o n s of real world knowledge, for understanding natural language: no approximate, 95%, s o l u t i o n s will do, just as they won't do in physics".

(ii) e m p h a s i s on the importance world, i n d u c t i v e knowledge, in the s t r u c t u r e s of (i)

(3) Where to st~: "Since difficult examples clearly require reasoning to be understood, we cannot even begin without such a theory because, without it, we could

(iii) e m p h a s i s function 130

of real expressed

on the communicative of sentences in context,

l I I I I

i I ! ! J

I

t ! I I

scientific, one and that attempts to draw a n a l o g i e s between science and AI work on language are not only overdignifying, as above, but are intellectually misleading. Conduct with me, if you will, the following Gedankenexperiment: suppose that tomorrow someone produces what appears to be the complete AI u n d e r s t a n d i n g systems, including of course all the right inference rules to resolve all the pronoun references in English. We know in advance that many ingenious and industrious people would immediately sit down and think up e x a m p l e s of p e r f e c t l y a c c e p t a b l e texts that were not covered by those rules. We know they would be able to do this just as surely as we know that if someone were to show us a boundary llne to the u n i v e r s e and say "you cannot step over this", we would p r o m p t l y do so.

i.e. the finding of the correct-in-context reading for a sentence, as opposed to the standard linguistic view, which is that the task is the finding of a range of possible readings, independent of context (iv) emphasis on the e x p r e s s i o n of rules, structures, and i n f o r m a t i o n w i t h i n an operational/procedural/computational environment. C o n v e n t i o n a l linguistics has still not a p p r e c i a t e d the force of these points, which are of course c o m m o n p l a c e in A.I. Let me now turn to the position sketched out earlier under three headings, and set out some countervailing considerations. It should be made clear that in whaa follows I am making only m e t h o d o l o g i c a l points aout the a s s e s s m e n t of systems in general. No attack on the content of anyone s system is intended.

! ! ! ! !

1 1 I !

Do not m i s u n d e r s t a n d my point here: it is not that I would c o n s i d e r the one who offered the rule system as refuted by such a counter-example, particulary if the latter took time and i n g e n u i t y to construct. On the contrary, it is the counter-example m e t h o d o l o g y that is refuted, given that the proffered rules expressed large and interesting generalizations and covered a wide range of examples. For the simple m e t h o d o l o g y of r e f u t a t i o n is the method of idealised science, where one awkward p a r t i c l e can o v e r t h r o w a theory*. In the study of language such a m e t h o d o l o g y is no more a p p r o p r i a t e than it is to c o n s i d e r the definition of fish as s o m e t h i n g that swims and has fins as b e i n g "overthrown" by the discovery of a whale. Of course it is not, nor does the d e f i n i t i o n lose its power; we simply have special rules for whales.

First, to the theory a~d practi~ point. It seems to me worth e m p h a s i z i n g again that there can be no other ultimate test of a system for u n d e r s t a n d i n g n a t u r a l language than i t s success in doing some specific task, and that to pretend o t h e r w i s e is to introduce enormous confusion. Considerations of logic or psychological p l a u s i b i l i t y may indeed be suggestive in the construction of AI language systems, but that is quite another matter from their ultimate accountability, which can only be w h e t h e r or not they work. Suppose some system had all d e s i r a b l e logical properties, and had moreover been declared by every r e s p e c t e d p s y c h o l o g i s t to be consistent with all known experiments on human reactions times and so on. Even so, none of this would matter a jot in its j u s t i f i c a t i o n as a c o m p u t a t i o n a l system for natural language.

The fact of the m a t t e r is surely that we cannot have a serious theory of natural language which r e q u i r e s that there be some boundary to the language, outside which u t t e r a n c e s are too odd for consideration. Given sufficient context and explanation a n v t h l n ~ can be a c c o m m o d a t e d and understood: it is this basic human language c o m p e t e n c e that generative linguistics has systematically ignored and w h i c h an AI view of language should be able to deal with. We know in p r i n c i p l e (see Wilks 1971 and 1975a) what it would be like to do so, even if no one has any c o n c r e t e ideas about it at the moment*: it would be a system that could discover that some earlier i n f e r e n c e it had made was i n c o n s i s t e n t with what it found later in a text, and could return to try again to understand. And here, to be interesting, the b a c k t r a c k i n g would have to be more than simply the following of some

In a similar vein, it seems to me highly misleading, to say the least, to describe the recent flowering of AI work on natural language inference, or whatever, as t h e o r e t i c a l work. I w o u l d argue that it is on the contrary, as p s y c h o l o g i s t s insist on r e m i n d i n g us, the e x p r e s s i o n in some more or less a g r e e a b l e s e m l - f o r m a l i s m of intuitive, common-sense knowledge, revealed by introspection. I have set out in c o n s i d e r a b l e detail (Wilks 1974) why such an a c t i v i t y can h a r d l y be called "theoretical", in any strong sense, however worthwhile it may be. That it i_~sw o r t h w h i l e is not being q u e s t i o n e d here. Nor could it be, since I am engaged in the same activity myself (Wilks 1975b). I am making a meta-, m e t h o d o l o g i c a l , point that the a c t i v i t y does not become more v a l u a b l e by being described in value-added terms. The w o r t h w h i l e n e s s , of course, is shown later by testing, not by the intuitive or aesthetic appeal of the knowledge represented or the formalism adopted.

*The bad i n f l u e n c e may not come directly from science, but via "competence theory" in linguistics. * W i n o g r a d ' s thesis, of course, had a system for c h e c k i n g i n f e r e n c e s and new i n f o r m a t i o n against all that it knew already, t h o u g h it is not clear that such a direct method would extend to a wider world of texts. In (Wilks 1968) there was a very crude p r o g r a m for finding out that an assignment of sense, earlier in a text, had gone wrong, but it was almost c e r t a i n l y an i n e x t e n s i b l e method.

Let me turn to position (2): A_~I Science. It seems clear to me that our activity is an engineering, not a 131

to try and counter, in a brief i n a d e q u a t e manner, what I see as the effects of the where t__qostart view.

branch of a parsing that h a d been ignored earlier: it would have to be something e q u i v a l e n t to p o s t u l a t n g a new sense of a word, a new r e f e r e n c e of a pronoun, or even a new rule of inference itself. It is surely these situations that the "AI paradigm of language understanding", and perhaps it alone, will be capable, in principle, of tackling, in the future, and it is these features of language, that require such maneuvres, that show most clearly why the "100%-Scientific Rqle" picture does not fit language at all, and why time spent trying to make it fit may be a d i v e r s i o n of attention from really key areas like the heuristics of m i s u n d e r s t a n d i n g and contradiction.

and bad

The view is an a l t e r n a t i v e to a more simple-minded view which goes as follows: "we should now concentrate on difficult examples, r e q u i r i n g reasoning, when s t u d y i n g natural language u n d e r s t a n d i n g , because the basic semantics and syntax have been done, and we are t h e r e f o r e right to focus on the remainder". This view is simply h i s t o r i c a l l y false about what has been done, so let us leave that and turn to the much s u b t l e r where t__oo start view which holds that, on the contrary, the basic s e m a n t i c s of natural language u n d e r s t a n d i n g have not been done and cannot even b__eestar~ed without a full theory of reasoning capable of tackling the most difficult examples, because, without such a theory, we can't know that it isn't needed, even in the a p p a r e n t l y s i m p l e s t cases. The a r g u m e n t is llke that against the employment of paramedical staff as a front line in community medicine: we cannot have a h a l f - t r a l n e d doctor t r e a t i n g even influenza, because unless he's fully t r a i n e d he can'~ be s u r e i t isn[t pneumonia.

Perhaps a m o m e n t ' s further dilation on the role of c o u n t e r - e x a m p l e s is w o r t h w h i l e here. Consider two counter-examples: one produced against the " e x p e c t a t i o n as basic m e c h a n i s m of parsing" h y p o t h e s i s of R i e s b e c k (Riesbeck 1974), and one against my own " p r e f e r e n c e as basic m e c h a n i s m etc." (Wilks 1975c) hypothesis. Riesbeck considers s e n t e n c e s such as "John went hunting and shot a buck", where, putting it simply, the concept of hunting causes the system to expect more about h u n t i n g and so it r e s o l v e s "buck" c o r r e c t l y as the animal and not the cash. One then i m m e d i a t e l y thinks of "John went h u n t i n g and lost fifty bucks".

One o b v i o u s trouble with the argument, in both its l i n g u i s t i c and m e d i c a l forms, is its openness to reduct~o ad absurd~m replies. It follows from that position, if taken seriously as a theory of human understanding, that no one understands a n y t h i n g until they are capable at least of understanding everything. So, for example, a child could never properly be said to understand anything at all, nor perhaps could the o v e r w h e l m i n g m a j o r i t y of the h u m a n race. There is clearly s o m e t h i n g untrue to our e x p e r i e n c e and c o m m o n - s e n s e there.

Conversely, in my own system I make much of the p r e f e r e n c e of concepts for other concepts to play certain roles, so that for e x a m p l e in "John tasted the gin", "gin" will be r e s o l v e d as the drink and not the trap, because of the p r e f e r e n c e of tasting for an edible or potable object like the liquid gin. Someone then, p l a u s i b l y enough, comes up w i t h "He licked the gun all over and the stock tasted good", where the p r e f e r e n c e on a small scale would get the wrong "soup" sense of "stock", and not the "gun part".

I am not t r e a t i n g this position with the seriousness it deserves in the space a v a i l a b l e here. In a weaker form it might draw u n i v e r s a l agreement. If, for example, it w e r e put in the w e a k e r form that it was not really worth starting machine t r a n s l a t i o n in the way they did in the 1950"s, because they knew they had no s e m a n t i c mechanisms, and so w i t h o u t some a b i l i t y to go further, it was not even w o r t h s t a r t i n g there. In that weaker form the a r g u m e n t looks far more plausible.

It should be clear that these counter-examples are to what appear t o b e , s u p e r f i c i a l l y , opposed t h e o r i e s of parsing. My point is that in ~ case do the examples succeed in showing a theory useless, i.e. neither "preference is no good" nor " e x p e c t a t i o n is no good" follow from the p r o d u c t i o n of the c o u n t e r - e x a m p l e s . What is needed of course, and what in fact both parties are trying for, is some s u i t a b l e m i x t u r e of the approaches. But, and here is the key point, there will not be any magic right m i x t u r e either. There can only be a c o m b i n a t i o n that will itself go w r o n g with s u f f i c i e n t l y i n g e n i o u s examples. Only a r ~ e o v e r v m e c h a n i s m will save us, Just as it saves people, who misunderstand all the time. There will never be, nor could there be, a RIGHT combination, in the way that F : k,,,__~L gives a right theory of g r a v i t a t i o n ~ h e n , and only when, n : 2

What I am questioning here is its stronger form: and again the reply is the same, namely that the position is another version of the 100%-rule fallacy: that in s c i e n c e you have to have a c o m p l e t e theory to h a v e any o r t ~ theor~ at all. This is untrue to language and diverts our attention from application and from an s y s t e m that could misunderstand and recover. Let me s u m m a r i s e the p o s i t i o n paper: it is an attack on what I have called the 100%-rule fallacy, alias the use of scientific methodology and assessment in work on AI and natural language. In my view this p o s i t i o n ~as four u n f o r t u n a t e aspects:

Finally, let me turn to the third aspects of the initial position, w h i c h I called w h e r e t o start. This brings up the very difficult q u e s t i o n about the r e l a t i o n of r e a s o n i n g to natural language, and I have made some remarks on that in the paper in section 2 on "Primitives". Here I just want 132

I I i I ! I !

i I |

! i ! !

I

! I I !

I. It requires holding, usually implicitly, the false metaphysical position that there is some boundary to natural language over which one cannot step. 2. It has a false view of the counter-examples as rejectors.

role

of

3. It encourages talk of theoretical advance in a non-theoretical area, and downgrades the engineering aspects of AI, and thus the notions of tests and application, which are the only criteria of assessment we have or could have. 4. It distracts attention from the heuristics of misunderstanding which should be the key to further advance.

REFERENCES Minsky, M., "A framework for knowledge", MIT AI Memo NO,

representing 306, 1974.

Riesbeck, C., "Computational understanding", Memo from ISSCO No. ~, 1974. Wilks, Y., Derivations", Memo, 1968. Wilks, Y., Language",

"Computable Semantic Systems Development Corp.,

"Decidability Mind, 1971.

and

Natural

Wilks, Y., "One Small Head", Foundations Language, 1974.

of

Wilks, Y., "Philosophy of Language" in Notes for the Tutorial on ComDutatlonal Semantics, ISSCO, Castagnola, 1975a. Wilks , Y., "A preferential Pattern-matchlng Semantics for Natural Language Inference", Artificial Intelli~enoe, 1975b. Wilks, Y., "An intelligent analyzer and understander of English", Comm. A.C.M., 1975c. Winograd, Language,

T., Understandin~ Edinburgh, 1972.

Natural

1 i I

I

133