Natural Language Processing

PANEL ON NATURAL LANGUAGE PROCESSING

Roger Schank (chairman)"; Eugene Charniak; Yorick Wilks; Terry Winograd; William Woods

Four questions were posed to the panelists. Some panelists chose not to answer each question. The questions were:

1. To what extent is natural language processing a separate field? Would it he wrong to say that AI and NLP are identical fields? Can prohlem solving he separated from NLP, for example?
2. Has frame theory heen a significant contribu­ tion to NLP? Is the organization of world knowledge the fundamental proolem in NLP?
3. Is syntax a dead issue? What low level langu­ age prohlems remain to he solved?
4. Why are new programming or representation lan­ guages necessary for NLP?
5. Ql: To what extent is natural language processing a separate field? Would it be wrong to say that AI and NLP are identical fields? Can prohlem solving he separated from NLP, for example?
6. EC: Natural language processing has been a suo­ branch of Artificial Intelligence since the early days of the field. For most of this time however, it has heen slightly on the periphery, partially due to its close relation to linguistics, a relation not shared by other sub-branches, and partially due to the fact that artificial !ntelligence was seen as nearly synonymous with problem solving and heuristic search. This has changed as of late and it now seems clear that language process­ ing is a central topic within AI. That is to say, many NLP researchers have been finding that AI proolems occur in NLP as well.

Minsky (1976) tacitly assumes a close connec­ tion between the problem of visual recognition (of, say, a room) and language "recognition" (as in recognizing that certain story circum­ stances are instances of a birthday party). This assumption has been extended by Bobrow and Winograd (1976) who see practically all of AI in- terms of the recognition problem. Yet other researchers have commented on the role of problem solving (Rieger 1976) (Schank and Abelson 1975) or search (Charniak 1977) in NLP. There are still aspects of NLP which are unique to it within AI (grammars) and problems in other domains which do not enter NLP (line finding), out it seems safe to say that NLP is well situated in the heart of AI.

Whether NLP is deservingly situated in the heart of AI is a more difficult question, as it depends on whether one sees the extension of NLP into questions of knowledge representation and problem solving as well motivated. That it is well motivated can be supported by one or more of the following premises.

1. These problems are important whether or not they are NLP in a strict sense. Furthermore, since many of these issues are most easily approached from the viewpoint of NLP i~ makes sense to "extend" NLP to these issues.
2. If one is concerned with not merely "analy­ sing" a sentence, but really "understanding" it, I do not see how these problems can be avoided. Strictly speaking this would depend on one's definition of "understand", but it is surely not coincidence that all of the stan­ dard tests for natural language comprehension assume that one has a large body of common sense knowledge, and that one can do problem solving with it.
3. Some more obviously "linguistic" processes like ambiguity and reference resolution depend on these processes.

These premises are not mutually exclusive, and I would support all of them, but any one would do to support the current view of NLP held in AI.

• RS: Natural Language Processing has been only a peripheral part of AI until the last few years. In the 1969 IJCAI only 10% of the papers were on NLP. In 1975 that percentage was 23%. When I was looking for a job in 1968, I talked to the S.tanford AI ": lab ahout writing a program that would talk to their hand-eye system in English. There was no in­ terest in the project by the lab.

Recently things have changed. Part of the change was due to the success of Winograd's program that did what the Stanford AI lab said was uninteresting. The change has come I think from something larger than any one program, however. Lately, researchers in AI have started talking about human process. and programs that simulate those processes, much more frequently than they ever did. Resear­ chers .4re starting to understand that tour­ de-forces "in programming are interesting but non-extendable.

To go back to the example I cited earlier, the AI lab at Stanford was interested in Winograd's tour-de-force. On the other hand, it "doesn't seem Winograd was interested in it as much as Stanford was. In any case he didn't pursue it. He (as well as the majori­ ty of members of this panel) has become more interested in what the general issues are. How people do cognitive processing is much more a part of AI than it ever was before.

For that reason AI and NLP are more and more becoming nearly identical fields, as the AI people recognize that how people use and re­ present knowledge is the key issue in the field and as the NLP people realize that that same issue is at the root of the problems of language processing.

Researchers in NLP have become less and less concerned with language issues per se. We are more interested in inferencing and memory models for example. We seem lately to be coming against the same problems that everyone else in AI has; knowledge representation; pro­ cessing of goals; planning and so on. I be­ lieve however that our solutions to these pro­ blems will continue to be different than those already proposed in AI. Whether AI shifts over to our solutions, or whether we continue to go on our merry way, it seems that a radical change from theinitial conception in AI of NLP as being some peripheral (albeit difficult) field has occurred. AI and NLP are becoming fields that aim to model human cognition. The more this becomes the case, the more the par­ ticular input and output devices being used to gather and transmit input to a system that un­ derstands will become side issues.

• YW: The superficial answer is that NLP must remain a separate field, if only because requirements of text processing, like spelling correction, will have no analogue in, say, visual process­ ing.

However, under the question are, I think, two deeper questions:

1. Will different AI activities (language, vision, physical manipulation) ultimately re­ quire access to the same knowledge base?
2. If knowledge bases for different activities are not independent, is one more fundamental than another?

On (a) I remain unconvinced by the "they must have the· same base" arguments--it is not ~ all clear to me that my physical ability to drive a car and my linguistic knowledge (or knowledge accessed in talk about cars, rather) are, or needs be, the same (Viz: one's diffi­ culty in describing easy physical tasks-)-.---­ Evolutionary arguments (Gregory: language came later than vision--therefore visual structures were used for language) and those from dogs, say, that see but do not talk, seem to me very weak, and to lead to no clear, undisputa­ ble, consequences for constructing AI systmes.

On (b) I believe that knowledge is dependent on language, rather than vice-versa (as seems the norm in AI beliefs, see McCarthy). This is hard to justify without bringing in the whole of linguistic philosphy, but a.s a very high-level psychological assumption it does, I believe, have empitical consequences for how we should construct AI systems.

TW: The fact that these questions are posed for a panel like this is a good indication of how immature our science still is. Imagine a time centuries ago when a similar panel gathered at the conference on celestial me­ chanics to discuss questions like: To what extent is astronomy a separate field? Would it be wrong to say that physics and astrono­ my are identical fields? Can angular momen­ tum be separated from astronomy, for example?

The answers to the questions (in both forms) are "To a large extent," "res," and "It de­ pends on what you mean bY 'separate'." AI is the general study of those aspects of cogni­ tion which are common to all physical symbol systems, including humans and computers. As such, it covers a wide range of cognitive processes, each of which embodies the general principles, and each of which has special features of its own. AI and NLP are not identical any more than AI and vision, or AI and game playing, or AI and medical diagnosis. Human language poses a special set of prob­ lems, having to do with creating and inter­ preting structured symbolic objects which can be conveyed over a limited channel (sequence of speech sounds or marks), and which are in­ tended to communicate from one thinking being to another. In this aspect, it is quite dif­ ferent from any of the other AI topics listed above. The structure of language is a result of its function, and by reducing it to what it has in common with other AI areas, we lose sight of its unique features.

The fact that the questions are posed as they are reflects a historical situation in the development of AI. In fact, most of the dif­ ficult issues with which researchers in NLP must grapple .are not language issues at all, but more fundamental issues of representation and cognitive processing. These include issues such as : the role of primitives in re­ presentation; the use of frame-like constructs in reasoning; the nature of plans and their relation to actions; the structuring of know­ ledge for problem solving and deduction; the amount of inferencing done with new knowledge; etc. etc. A quick scan through the litera­ ture in AI natural language work will show a high proportion of the verbiage devoted to these rather than the issues of language.

This is a necessary step. It would be as im­ possible to develop a satisfactory theory of language without having an understanding of general cognitive processing as it would be to have a satisfactory theory of celest~l motions without basing it on theories of me­ chanics. We who do natural language research are in a good position to examine these ques­ tions. By having a specific set of problems to deal with, we are put into contact with the broader issues in a way which would be impos­ sible through sheer top-down musing. It will almost certainly remain true for many years to come that research in natural language will be intertwined with research in broader issues of cognition and representation. But in doing this kind of mixed research, we should not lose sight of the fact that we are really wearing two hats, and that language is only a partial reflection of the range of issues in AI.

Q2: Has frame theory been a significant contribu­ tion to NLP? Is the organization of world knowledge the fundamental problem in NLP?

RS: Frame theory may not be the particular solu­ tion to the problems of NLP, but its signifi­ cance seems clear. Natural Language Process­ ing will be successful precisely when an ade­ quate theory of the organization and repre­ sentation of knowledge has been fully worked out. Frames are one suggestion for how to organize world knowledge. They will probably, when fully worked out, only help to solve one aspect of the organization of knowledge pro­ blem, but that is a significant aspect indeed.

Scripts, our particular' instantiation of frame theory, will in no sense solve all of the pro­ blems in NLP. But it seems clear that a good deal of human functioning is script-based, and we have gone fairly far using scripts to un­ derstand newspaper stories.

YW: I don't think it has yet, but it well may do so. There has been a rush of ingenious sug­ gestions about, and implementations of, frame systems; but not yet enough thought about what claims are being made, and whether or not they are true. My instinct is to answer "no" to the second question, simply because I can imagine someone having organized world know­ ledge perfectly, but having got no farther with NLP. Conversely, we are all pretty good at NLP, but have pretty bad organization of world knowledge in many respects. I cannot see that "organization of world knowledge" is a concept that makes much sense, as such, and wholly independent of particular tasks and purposes.

However, it remains true that world knowledge organization is a growth point in NLP at the moment, and will probably remain so for some time. My own hunch is that the use of world knowledge in NLP (at high levels, that is) is going to pay off in connexion with robust systems that run on texts where they don't already know all the word senses etc. And the question here that I am most unsure of the answer to is, what is the role of high­ level knowledge structures, like frames, in actually parsing input text?

• TW: The second half of this question is another form of the questions above. It is both true and patently false, depending on how "problem in NLP" is interpreted. If we are interested in building systems or models which reflect the processes of language production or com­ prehension, then the answer is "Yes." When a person uses language, he or she is not sim­ ply using a "language faculty", but is making use of a full range of mental operations, of which only some are directly linguistic. If we wish to model this behavior, we must model all of these operations, and in doing so the organization of world knowledge provides the largest stumbling block.

On the other hand, if we view "NLP" as des­ cribing a scientific enterprise, then the organization of world knowledge is not a pro­ blem within its domain at all. It is a pro­ blem in the more general domain of AI, and one whose solutions will form a basis for attacking problems in NLP, which deal with the structure of language and communication.

The first half of the question raises two is­ sues. The phrase "contribution to NLP" is subject to all of the problems discussion above, and I won't go into them again. The phrase "frame theory" raises the question "What is a frame theory?" As far as I can see, none of the work which goes by this name is really a "theory" at all. There are two levels at which it has been couched: a general intuition, and a set of specific ~e­ chanisms. At the intuitive level, it has bee been of great importance. The attention of researchers has been focussed on a different set of issues from those which were predomi­ nant in the older, more atomistic ways of thinking about knowledge and meaning. Issues of memory structure, chunking, accessibility, and pattern-related control structure are be­ ginning to make inroads into a literature which previously focussed on predicates, pri­ mitives, parsers, and uniformly applicable algorithms. At the detailed level, I think we are far from having s.atisfactory notions of how to apply the intuitions. There are a number of simple ideas, such as the use of explicit cross linkages to switch between frames (Minsky), the notion of scripts as linear sequences of stereotypical events

(Schank), the association of specific proce­ dures with parts of a declarative frame struc­ ture (Winograd), and the need to view a single object as an instantiation of multiple frames (Bobrow and Winograd), but all of these are over-simplified and lack the kind of depth which will make them the center of a "theory". They are all useful, but do not constitute a theory any more than a set of specific tech­ niques for differentiation is a theory of cal- . culus.

• Q3: Is syntax a dead issue? What low level lan­ guage problems remain to be solved?

EC: Syntax is by no means a dead issue, but it is by no means clear what the issue is. Nobody claims that one can do without syntax, it is too easy to give sentences where it comes in handy, and nobody claims that one can rely en­ tirely on syntax to solve one's parsing pro­ blems, that we easily understand ungrammatical, or a-grammatical, sentences shows the contrary. ( 11 Pars-ing" here is simply the process of going from surface structure to "semantic represen­ tation". I am not restricting the term to grammatical parsing.) The real substantive issues as I see it are two: "how much" syntax is needed (both for analysis and generation), and how should this information be interfaced with the rest of the system. Further discus­ sion will show, I think, that we have ~o good answers to either of these questions, and so we should add a third, more pragmatic question: What is the best · way · to proceed until answers · are in on the first two?

In asking "how much" syntax is needed one usu­ ally takes as a basis for comparison the yet hypothetical set of rules which would allow one to generate all and only the grammatical sentences of English. Restricting ourselves to problems of analysis (since synthesis has played a comparatively unimportant role in AI), we can then ask at least two new questions. Are there any rules in the hypothetical basic set which are never needed to parse English, and secondly, are there rules which are so sel­ dom needed that we should not apply them "un­ less we have evidence that they are needed" (these last are scare quotes only). As for the first of these, since nobody has offered any candidates for such a rule, the current answer must be "don't know". My personal guess how-­ ever is that there are few if any such rules. Take, for example, the question of verb subject agreement. As everybody knows, and I have in­ advertantly shown all too often (in French), one can be understood in spite of the fact that one's verbs have the wrong ending. Perhaps then various of the verb ending rules are not needed. There are, however, occasional senten­ ces where they are crucial (Hunting dogs is/are forbidden), hence they do not qualify as being completely superfluous. Nevertheless, perhaps, they qualify as rules which are so seldom needed that we should only use them if we know they are necessary.

This is plausible, but there are technical reasons why this is impossible with current parsing systems. Virtually all AI parsers are depth first in that they try to produce one complete parsing, hopefully the correct one, first, and only try another if the first is rejected. In the example given above, by the time the parser gets to the verb it will have made a choice about the structure of the noun phrase. If the parser were to ignore subject - verb agreement, it would reach the end of the sentence with a successful, but possibly wrong, parse. That is to say, it would have no way to "know" that in this example the rule of subject agreement is needed. One could, naturally, have one'siParser discover both readings of the noun phrase, and hence have some indication that subject agreement is necessary, but whether in fact this is worth doing is a complicated trade-off question - more time handling noun -phrases against less time for verb phrases. My intuition tells me it is a bad trade-off, but only because sub­ ject verb agreement is such a simple question. In fact, nobody really knows the answers to questions like this.

One example may be suggestive, but it is hard­ ly an argument. And the situation with re­ spect to the second question I asked at the outset, ·~ow should syntactic information be interfaced with the rest of the system" is just as bad. So let me move on to the prag­ matic question of what we should do next. My personal belief is that one should simply take one of the existing 1arsers "off the shelf" and use it as a "front end". This view is rather old fashioned, but my reasons are quite simple. Few, if any of us, are really worry­ ing about ungrammatical texts, hence a gramma­ tical parse will not be unduly destructive. Furthermore it can be quite useful in establi­ shing the functional structure of the sentence. (That is, answering questions like, which is the direct object of the verb.) This is not say of course, that syntax is necessarily the only way, but given a free syntactic parser, why not use it. Except from arguments stemming from our ability to understand ungrammatical sentences, the only arguments one sees against syntactic parsers are of the form, 'vhy do a syntactic parse if it is not necessary". They are, that is, efficiency arguments.

Even if these statements are correct (and I for one have seen no hard figures, or arguments, to back them up) we know so little about language comprehension at the present time that ques­ tions of efficiency (aside from problems of ex­ ponential growth) are completely beside the point.

This is not to say that syntax is no longer a problem. If it is true that we can syntactically parse a much broader range of sentences than our programs can "understand" it is be­ cause of our ignorance of the latter process, and not because we understand the former.

RS: Although I doubt that many people agree with me, syntax has always been a dead issue. Our operating view was always semantics first, syntax later. Later it became Knowledge Re­ presentation first, syntax later. Syntax is the last thing that children get right when they learn how to talk. If that strategy works for them it should work for us. Note that I am not saying that syntax doesn't exist, only that more progress on understand­ ing issues will be made by ignoring it than by studying it. A great deal of effort has already been put into syntax. Most of these efforts do little to solve the problem be­ cause they beg the issue of what will be done with the syntactic trees they produce. By concentrating on understanding issues, syntax will be reduced to the secondary role it so richly deserves.

• YW: No, it's not a dead issue, but a number of people have agreed to put it on ice, and turn to something more interesting for a while.

Although there are now good off-the-peg syn­ tax parsers like Woods', I do not think syn~ tax parsing is settled. Those ATN's are clearly much better than the Harvard Analyzer, say, but still far too fragile with regard to semantic problems.

Conversely, I do not think those who have ar­ gued the superiority. of semantics-driven par­ sers have proved their case (I include my­ self.) I think they will only do so when they produce a semantics-driven parser as ro­ bust and portable as Woods' syntax one, and in as perspicuous a formalism (i.e. ATN's or production systems, or something like that). This will require a great deal more detail than has been forthcoming so far about how the semantic structures and rules determine relations as low-level as concord, for exam­ ple, and where exactly the conventional syn­ tactic generalizations are expressed in the system (if they are not, that too will be highly interesting). I think the whole busi­ ness is unproven at the moment, and settling it will have interesting consequences for linguistics.

• TW: Syntax is far from a dead issue unless we take a narrow (though popular) view that it covers the parsing of simple "grammatical" sentences into trees. Live questions in­ clude:

What is the relationship between standard syn­ tactic forms and the "phrasal lexicon"? Many idioms are syntactic structures (e.g. "the X'er the Y'er") which cannot be treated either through normal grammars or as lexical items.

How are syntactic cues used to communicate in­ formation about the message structure -- focus of attention: what is important, what the speaker considers new information, etc. There is a good deal of work in this area among lin­ guistics, and we are far from having satisfac­ tory answers from the point of view of build­ ing successful systems for communicating with human users.

How can syntactic analysis be fit into a mul­ ti-knowledge-source framework, to handle real language (with its stops and starts, ungram­ ~aticalities, etc.). This will demand inte­ grating ideas of syntax and grammar with those of frames and prototypes.

WW: During the pas.t decade or so, many advances have been made in natural language processing. Among these have been the development of for­ mal grammar models that provide efficient systematic frameworks for implementing gram­ mars of a complexity and sophistication mat­ ching the most advanced work in linguistics (e.g. Woods, 1970).

The most powerful of these grammars contain augments that allow one to associate condi­ tions with the grammar that refer to general semantic information and world knowledge, thus providing a formal interface between the syntactic knowledge embedded in the grammar i.tself and world knowledge that may be stored elsewhere and in some other fcrm. This has permitted the beginnings of a systematic in­ vestigation of the relationship between syn­ tactic knowledge and general semantic and world knowledge. However, our present state of progress in this area is quite immature -­ much of it is based on understanding single sentences in isolation, and the role of world knowledge is largely limited to rejecting in­ terpretations that do not satisfy certain se­ mantic selectional restrictions. The ultimate challenge in this area is to be able to choose between alternative readings of a sen­ tence based on sophisticated evaluation of the plausibility of the alternatives in con­ text.

From the other direction, many people have tried to attack the language understanding problem directly from the world knowledge, with minor if any interest in syntax. These systems to date, while they frequently illus­ trate suggestive approximations to various aspects of human performance, have not begun to develop a formal mechanism capable of handling a general range of language under­ standing phenomena. In many cases, the de-

I

.

vices underlying such approaches, when viewed as formal automata, lack the power necessary for a general treatment of the phenomena they purport to solve. In many other cases, the actual mechanisms underlying their performance are not even sufficiently clear to make such judgments. Much more care in formally defin­ ing representational conventions and the al­ gorithms that operate on them is required in the work in this area. Woods (1975) discusses· a number of representational issues that are frequently left vague in knowledge represen­ tation schemes, and the work of Cercone and Schubert(1975) and Brachman (1977) are begin­ ning to make some progress in this area.

One of the major difficulties in investigating the less syntactic characteristics of natural language understanding is that the fundamental problems of factual inference and plausible reasoning become a critical factor at this level. Since these are active research areas in their own right and are far from understood investigations into knowledge-based language understanding that are to make progress at this point must be making contributions to the study of factual inferencesand plausible rea­ soning as well. Unfortunately, many people who adopt the so-called knowledge-based ap­ proach do so without consideration of much re­ levant work that has gone on in the area of formal inference, dismissing the relevance of such work on the grounds that human reasoning is not necessarily logical or complete. In doing so, they lose the benefit of a great deal of understanding of fundamental problems and frequently propose inadequate treatments of problems that are well-known and well un­ derstood in the formal reasoning camps.

While many of the goals of researchers in for­ mal reasoning differ from those of natural language understanding and artificial intelli­ gence, the kinds of discipline involved in formally specifying the details of one's theo­ ry and rigorously assessing its capabilities that is characteristic of the formal inferen­ cing work is one that we should emulate. Suc­ cessful demonstration of a few chosen examples is not a sufficient benchmark -- especially the high-level description of the procedure that accomplishes the demonstration seems in­ tuitively satisfying, but the details of how it works are obscure. In this situation, one is too easily led to believe that something has been accomplished when in fact the under­ lying mechanism may be inadequate (sometimes either trivial or totally unsystematic and ad hoc).

In summary, while there remain a number of outstanding and troublesome problems of syn­ tax in natural language understanding, especi­ ally in the areas of coordinate and subordi­ nate conjunctions, mass terms, nominal com­ pounds, comparatives and superlatives, and discourse structure, it is clear that many of the most intractable ones involve the inter­ action of syntactic·.and semantic information and require access to plausible inference ca­ pabilities ·for the evaluation of alternative hypotheses. Thus, the problems of factual in­ ference and plausible reasoning are becoming a central problem in natural language under­ standing, as they have in a variety of other artificial intelligence areas. They are not the only problems, however, and their role in natural language understanding is intricately intertwined with the use of syntactic and semantic knowledge.

• Q4: Why are new programming or representation lan­ guages necessary for NLP?
• RS: I have noticed that when students are stuck on a problem in AI that is too difficult for them they often suggest writing a new programming language. I have little doubt that such pro­ gramming languages would be of some value in facilitating program writing, so that is not why I don't let them do it.

I don't let them do it because it has always seemed to me that writing a new programming language is a way of avoiding tackling a pro­ blem you do not know how to program in the first place. If you do know what the solution to a problem is like, then it should be pos­ sible to program it in any language, though · of course some languages will make doing it easier than others. But if you don't know how to solve the problem, new programming langu­ ages won't help.

Don't misunderstand. I think that there is a strong possibility we might use such a langu­ age if strong theoretical biases that we dis­ agree with are not built into it. Nonetheless I cannot help but wonder if the real problems aren't somehow being avoided by working on 'l . new languages . · ~

• YW: I think the best reason is standardization, and the advantage that would come from down­ grading superficial differences between nota­ tions and idiosyncratic systems. This faci­ lity seems to me the main advantage offered by KRL, but it will be psychologically very dif­ ficult to get many researchers to adopt ~ such standard language. One reason is the obvious one that such high-level languages are bound to commit users, both as regards control structures and metaphysical questions of re­ presentation, in ways they do not want to be committed. If such a language does not com­ mit one (and KRL makes little or no commitment about inference and control structures) then, of course, it will not be a programming lan­ guage at all, and many of the problems will not arise. I have a hunch, and this may be

over-cynical, that advance on this front may not, in the end, be made by those working at it from our end, as it were. It may come from those working on programming languages proper, and their work may later be seen to be rele­ vant to AI (rather than from those tackling the problem of knowledge representation from AI directly). I feel that we should for now go on working on sloppy systems that actually work, and hope that the language constructors (as well as the "semantics of programs" peo­ ple) will come along later and make what we do seem perspicuous and sensible as well as ef­ fective. The converse view (that we need ad­ vance with the languages first) seems to me unappealing because I cannot, here and now, imagine what a new-wonder-programming-language that made NLP possible would be like. (But that may well be only my own lack of imagina­ tion.)

TW: New programming and representation languages will always be needed for research not only in NLP, but in all of AI. Research follows a cycle:

A researcher has some new intuitions about the structure of a problem or a mechanism, and decides to embed them in a program.

He or she writes that program in some existing language, which was designed to facilitate working within a previously understood set of intuitions .

Some of the new ideas prove to be of general utility.

Someone (possibly the same researcher) writes a new language which facilitates building systems which make use of these new ideas.

• ~ The whole thrust of higher-level languages
• , .· . is to look for things which are done commonly , ,. and in a standard way, and to make their de­ tail invisible to the user, so that he or she can concentrate on the next higher level of structure. Without this, research would rapidly bog down, since the complexity of any system is strictly limited by the mental ca­ pacities of people who program it. To write more complex systems, we need languages which hide "those complexi~ies we can afford not to think about. AI went through a clear case of this twenty years ago, when the complexities of list processing and storage allocation were hidden in the primitives of LISP.

I believe (as do many others today) that there is another level which is ripe for the same treatment. This level has new structure along both the declarative and procedural di­ mensions. Declaratively, it deals with the way in which descriptions (defined by seman­ tic notions of "description", "prototype",

"instance", etc.) are implemented in list structures. The AI languages (e.g. Micro­ Planner, Conniver, QA4) took a small step in this direction, but did not try to deal with the semantics of description in a systematic way. Current representation language research (KRL, MDS, OWL, partitioned semantic nets) is much more ambitious in its attempt to provide the user with a higher level of structure. Procedurally, new languages must provide means to reduce the complexities of building systems which multi-processing, resource allocation, and integrated goal-driven and data-driven processing. We have many different starting points in the current AI systems, (including production systems, standard hierarchical control, coroutines, etc.) but they have not been satisfactorily unified in a workable sys­ tem.

Although no one can predict the details, it seems clear that some language or languages at this higher level will eventually become a widely accepted standard for work in NLP (and AI in general), just as LISP has domina­ ted the more traditional ALGOL/FORTRAN lan­ guages for current AI work.