Co~pu'ter~ ""d the

Human.itie~, Vol . .10, l'P. 265-274.,PERG,AMON PRESS, 1.976. Pdnted. in the U.S.A.

Computational Linguistics and the Humanist



WHEN LINGUISTS examine the languageprocessing 'activities of scholars who prefer to call themselves "humanists," the chief product they find is almost ineVitably a concordance, that is, an alphabetized list of character strings, me-rely the outside of language. From the linguists'pomt of view, the interior is separated from those character strings by what they call duality of patterning. But it is in the interior that the theoretically rich and intellectually satisfying problems, for both humanists and linguists, are to be found Computational linguistics has plunged into the interior. Syntax lies there, and has been explored rather thoroughly. The key to the next door is semantics, but if that key is to' work we must know in advance something of the treasure - some fundamental theoretical problems, in les sciences de l'homme - that lies behind the door. We start with. a consideration of the nature of computational Ifnguistics followed by a consideration of how to formulate problems in the analysis of literary texts so that the techniques can be fruitfully employed. After that can come a review of 'current work in semantics and discourse analysis and in speech understanding, In conclusion a brief sketch can show how that work might be used in starting a Iong-term Shakespeare project ..

What is Computational


Chomsky earned everlasting credit by bringing the study of natural language into conjunction with the study of mathematics (abstract algebra, recursive function theory), a conjunction that has made possible powerful models of natural language, without which computatlonal linguistics would have no guidance in constructing algorithms .. Algorithms -

feasible prescriptions for solving problems - are the substance of computation. Chomsky has also g~ven us the distinction between competence and performance .. However obscure the exposition of that distinction may be (see Hockett 1968.; 73-79), there is deep sigaiflcanoe in it. To describe all assembled bicycle is one thing; to describe the assembly of bicycles is another. The assembler must know what is in the blueprint (competence), but further needs the skill to take the parts in order, place them deftly, fasten them neatly (performance) .. In actuality, of course, the assembler" may never have seen the blueprint, nor need the performance of a speaker or hearer 'include in any physical sense the grammar that the linguist offers as the blueprint of a language. The linguist who characterizes competence and the psycholinguist who characterizes performance may have almost unrelated things to say about the samelanguage. The computational linguist .and the psychelinguist are both specialists in Ianguage process. Recognizing this Similarity, George MiIler has urged thatpsychollngulstics should adopt computation as its leading theoretical' metaphor (1974). Computational linguistics would readily accept the honor, but it cannot yet entirely eliminate the habit of modeling performance by writing an algorithm (say, a parser; see Grishman 1976) that makes use of a competence grammar. This habit has reached the level of transformational grammar (see Friedman I97!). It is not, however, the only method known. The algorithm and the grammar are woven together differently in a series of studies initiated by William A, Woods, Jr .. (for a good statement, see Kaplan 1973). ' Of course, trying to figure out how the brain

WUltamBenzon. is a Ph.D. candidate in English literature and David G. Ha.ys i~ a professor oftinguistics, botn at the Stare Unfll~1"8jty_ t)t New Yilrk, B:Uffalo.· 26S



processes language is quite different from programming digital computers to process language. All the neurological and psychological evidence indicates that the brain is quite a different computing machine from the digital computer. But the link between computational linguistics and the general. purpose electroni c digital compu ter is, in the language of philosophy, a matter of chance or contingency, not of necessity. We write algorithms for that kind of computer because that is the kind we have to work with. Nothing forbids us to devise algorithms for a different kind of computer, such as the human brain, and even to go beyond that to the simulation of that different kind of computer on a digital computer. If computational psycholinguistics emerges as a subspecialty, the psycholinguist will be responsible for performing the experiments from which a model of human speech processing can be derived, and the computational linguist will be responsible for providing the computational procedures to be used :in simula ring tha t model. Grammar has been much studied oflate, but for all its popularity, 11 remains a rather dry subject. The words of a text are linked together by a few very widely used devices: word order, affixes, function words such as prepositions and conjunctions, and contours of stress and pitch. (In writing punctuation partially supplies the missing contours.) Further devices of a more complex kind, called transformations, match different grammatical patterns so that the content of a whole sentence, for example as a relative clause, can be woven tightly into another sentence. To learn how modern linguistics organizes these materials into elaborate theories, the scholar need only turn to one of the many recent textbooks: to Iearn how computational linguistics makes processes Tun in grammatical ways, he can consult the works cited above. Computational linguists have rather completely exploited the leads discovered about fifteen years ago· Sheil (1976) shows that within a broad basic class of grammatical processors any choice must be made on extrinsic grounds, since intrinsically they are much of a muehness, But it is the hope for progress in semantics that linguists are presently more excited about. Humanistic scholarship, which has paid little attention to the ferment in grammar, may now draw closer to computational linguistics, for the real business of the scholar is far deeper into language and thought than can be reached with analyses of word order and prepositions. By reading and by borrowing from Iinguist colleagues, the scholar can obtain now what he needs in the way of

.grarnmatical tools; if tile tools are not quite good enough, he can collaborate with a computational linguist who knows the technology and refine the tools a bit. Having grammatical tools. tile scholar can look somewhat further to see his own work laid out.

Computing Ies Sciences de "Homme: Literary Texts Literary theory and practical criticism occur in a universe of discourse which is quite different from that of computational linguistics. So, for that matter, do biological theory and practical medicine; the biologist and physician use texts, but they would not say that they are students of language. Although literature consists of text, theory and criticism in literary scholarship are generally not about language. Literary critics are interested in the general theory of literary meaning and in the meaning of texts, but their notion of meaning has little in common with what linguistic semanticists have taken to studying. To say what criticism is about is all too easy: It is about ideas. The problem is to connect up what criticism studies with what semanticists analyze. The problem is not to be taken lightly. Though literary theoreticians and critics are quite experienced in thinking about meaning, that thinking is. quite different in texture from the models of meaning currently implemented on computers. It is unlikely that the literary critic would even- recognize cognitive network theory (the most widely implemented computational semantics discussed below) as a theory of meaning. Nor, for that matter, would the computational linguist know what to do with, for example, E.D. Hirsch's Validity in Interpretation Or Octavia Paz's very different The Bow and the Lyre. The methods of scientific inquiry and humanistic Inquiry are quite different, both in purpose and in technique, To talk of simulating a formal model of literary performance is to talk of a rapprochement between the Two Cultures, science and humanism. The theme is a worthy ODe, and the problem is deep. It is not a matter of solving the problem before going on. Rather it is a matter of being aware of what is being attempted and having a place in one's mind for consideration of such fundamental questions, even if that place seems to be quite a distance from developing the theory, creating the algorithms, and writing the programs. Humanists rarely discuss in an explicit f0Im such





questions as "What is man's place in the scheme of of the operating equipment of the adult human things?" or "What is the meaning of life?", issues nervous system is a representation in cortical tissue of transactions between cortical and subcortical which implicitly animate all humanistic inquiry. And,although they are not questions of the sort structures, and of transactions between a semantic and a syntactic system, both of which are Located in which science is equipped to answer, there is, .nevertheless, a closely related question which my the cortex. He suggests a mechanism for handling scientific theory of man will have to deal with jf it personal pronouns and, indeed. makes this mechais to be a useful theory: "What is the nature of the nism primary; by means of it the self-representation human nervous system that it can support self- is bootstrapped into the system. While the adequacy of Benzon's proposals is an awareness?" Science cannot reasonably be expected open question - yet however wrong it be in detail to (attempt to) expound the meaning of life, but it can be expected to (attempt to) explain how or even in principle, it is clear. Clarity helps PLOindividuals of the species homo sapiens sapiens grammers, and is good for talk, too. It is, perhaps, come to pose such questions and are prone to the ultimate reason for bringing the computer into suicide in the absence of a satisfactory answer. the humanities at alI. Such considerations are a necessary part of any The computer, however, does not bring clarity; it attempt to create a theory of literature which is only begs for it. Obfuscation by computer is possuitable for programming, and which is at the same sible, but offensive. To avoid it, many distinctions time powerful enough to yield compelling results. must he made sharp. The creative novelist or poet Man is self-aware. Traditional literary topics, such as concocts a world of his own; if we can understand him, we can perceive his answer to the question, irony, the lyric 'I' intersubjectivity, the masks or faces problem in dramatic theory, all entail discus- "What is man's place in the scheme of things?' sions of self-awareness. At least one critical school Language is a good way of transmitting concocted has made self-awareness one of the central points of worlds from one mind to another, for close reading are necessary to the debate, only they talk about it under the guise of and delicate interpretation achi evernent of accurate perception, and good 'deconstructing the subject' (Macksey and Donato 1970). Consequently, it seems likely that any interpreters can disagree. But the computer model attempt to develop a computable model of literary at best a simulation of the novelist or poet as texts will have to include .3 simulation of self- creator, deals with the concocted world at one awareness if it aspires to deep results. remove, a.s a content of thought or communication. In fact, one can reasonably argue that a semantic Here are two universes of discourse, one embedded theory which cannot provide a model of self- within the other; a model of the outside universe awareness cannot deal adequately, in any type of does not determine the operation of the internal discourse, literary or nonliterary, with first- and universe. Benson's analysis of "Th'Expenee of second-person pronouns, which designate the Spirit," concerned with the semantic structure of speaker and addressee of the utterance. The idea of the poem, is not an. interpretation of that poem in a computer simulation of self-consciousness is mindany recognizable sense. The analysis attempts to boggling, but there does not seem to be any way explain how the poem works in the mind; it is not out of it (except perhaps by noting the difference an attempt to translate the significance of the poem between _a simulation and a good simulation). into expository prose. Yorick Wllks recently defended a concept of This is not to say that computer simulation machine privacy analogous to the privacy of human could not be used in investigating the interpretation minds and argues from it to a mind-body distinction of literary texts. It can be so used. But the method for computers (1975b). We neither accept nor reject would have to be indirect. Wilks's argument, but he did make it. it is worth To investigate psychoanalytic interpretations thinking about, and attempts at literary computing requires programming a simulation of a psychomust consider such problems. analytic theory on a computer and having that At least one model of self-awareness has been simulation operate on a literary text. Similar investimade in a form which, by using techniques gations are possible for a phenomenological interpredescribed below, could be simulated on a computer. tive strategy (such as that of Georges Poulet), or a Benzon (1976) gives a partial analysis of Shake- Marxist strategy. a N6W Critical strategy, a neospeare's sonnet "Th'Expence of Spirit" which hinges structuralist strategy, etc. The various simulations on an account of self-awareness. He argues that part could even argue with one another on the merits of




their respective approaches, each a different concacted world in which dwell not fictitious characters but novelists and poets. This sort of thing is certainly well beyond the present state of the art, but we have no reason to believe that it might not .one day become possible. As for the desirability of doing such a thing, that is yet another question. There is only one point in this discussion: The analysis you get out of the computer is only as deep as the theory you put in. The linguistic key works only for those who already know something about the treasure behind the door. If you wish. to analyze literary texts, then you need a theory' of human literary performance With which to conduct the analysis. Any such theory that can be formulated now Is going to be extremely etude, but crudity is not the same as triviality. What is crude can be refined; to rise from triviality to significance is much more difficult. The time is ripe for crude but nontrivial theories. Linguistic Semantics and Discourse Theory

In the lively free-for-all debate now filling the literature of (compu ta tional) linguistics and artificial intelligence, we give our attention mainly to the several versions of semantic or cognitive network theory because 11 aims at constructing an object, the network, to which texts are related by linguistic processes. In the network some wish to simulate reasoning (see papers in Gregg 1974, 801so 1973); the ideas of a novelist or poet can, we think, be simulated in a network. Much of the work in artiflcial intelligence related to language has been reviewed and evaluated by Yorick WHks (l975d, 1976a). Psychology, once said to desire no theory but only a black box between stimulus and response, has begun the attempt to put a network inside the black box (Collins and Loftus I 975, Frederiksen 1975, Kintsch 1974, Norman et al. 1975, Solso 1973, 1975, Tulving and Donaldson 1972). Anderson (1975) uses a network in a crude but not triv:ial model of language acquisition. These psychologists adopt the network not because it is convenient for computation but because it i~ for the moment the b est kind of theory of human information processing. As a model of thought, the structure of ideas, knowledge, etc., UJ_ecognitive network turns up in diverse applications. Using a simulation of Propp's model, Klein et al. (1974) generate single-move folktales. Hays (1976) analyzes some concepts of

alienation employed in social science. White (1975) analyzes the belief structure of a millenarian cornmunity. Others study the visual recognition of scenes (Winston 1975). The range of applicability of the theory seems limitless; only the capacity of the theory, with plenty of room to grow up to its responsibilities, remains inadequate. In the 19605, Marvin Minsky and a group of students (Bertram Raphael, Daniel G. Bobrow, M. Ross Quillian, Thomas G. Evans, and Fischer Black) wrote models of knowledge and understanding (Minsky 1968). The fifteen years or less since then is not long enough to bring the whole field into agreement on more than the main principles of cognitive structure, but the following paragraphs suggest some principles and cite enough literature to permit the reader to discover the details of some of the current competing theories. One principle is hierarchical classification. Similar concepts are collected under a generalization; thus tree, vine, grass, and bush are similar concepts, each a variety of the more general green plant. Specific instances are linked to more or less general concepts. Syllogistic reasoning is represented in this part of the cognitive network: "All men are mortal; Socrates is a man; therefore Socrates is mortal" is certified by a sequence of links from the concept of Socrates, an instance, to the concept of man, to the concept of mortal entities. Another principle is the formation of new concepts by interlinkage of others. By linked concepts of rationality and animality, man, the rational animal, is defined. A tree seen at a distance is composed of foliage and trunk. Dancing is rhythmic motion of trunk and limbs. Just how many ways of linking concepts together are necessary is an important question. A short list proposed by Hays (1975) may be better or worse than others; the important point is that many now believe that new concepts are made from old ones in only a few distinct ways. Concepts can be localized in spacetime, At this place in my world! a maple tree stands, its lifetime spanning the twentieth century. At that place, on a certain day, Nureyev danced. In spacetime, concepts are linked by proximity: in space, they are above and below, north and south of one another, and in time they are simultaneous or sequential. Generalizing gives repeatability; a program for doing one thing after another is a generalized localization called a frame by Minsky (1975) and a script by Schank (1975b; see also Meehan 1975). The work of Scragg (1975) and Hendrix: (1975) should also be consulted.





Localizations in spacetime can be qualified. Although we tend to think first of real spacetime, we also concoct new worlds in fiction.dn dreams in hope and in fear. "John wants to find a banana and eat it" creates a new world in which the concept of his finding a banana is localized. The author's work is to concoct a world; the critic's, to explicate the concoction. In 'the formation of new concepts, two methods have to be distinguished. For a bird, a creature with wings, its body and the added wings are equally concrete or substantial, In other cases, something substantial is molded by a pattern. Charity, an abstract concept, is defined by a pattern: Charity exists when, without thought of reward, a person does something nice for someone who needs it. To dissect the wings from the bird is one kind of analysis; to match the pattern of charity to a localized concept of giving is also an analysis, but quite different. Such pattern matching can be repeated recursively, for reward is an abstract concept used in the definition of charity. Understanding, we believe, is in one sense the recursive recognition of patterns ill phenomena until the phenomenon to be understood fits a single pattern. Any belief about man '8 place in the scheme of things is representable by recursive linkages among abstract concepts. To analyze the semantic structure of a work of literature is to specify abstractions and their connections in such a way that if the creator had all of them in his cognitive network he could then create the piece. As far as anyone knows, it is inherent In. the formal nature of such an analysis that there be no unique solution; the same work could issue from different creators having different systems of abstract concepts. Not at all obvious, however, is how this limitation on formal analysis relates to current critical controversies about the capacity of a text to support multiple interpretations, Since the formal limitation exists in one universe of discourse and the critical controversy in another, making inferences from one universe into another is difficult and uncertain. The effect of a network representation of meaning is that the meaning of any given concept is a function of its place in the entire network. Imagine that your favorite unabridged dictionary has your favorite encyclopedia attached to it, so that, for example, the dictionary entry for "Indian Wars" leads you to an encyclopedia article. But every word in the article, like all those in the encyclopedia, is attached to the dictionary, and so back to the encyclopedia again. IT you follow all these links,

you can arrive at a complete definition of Indian Wars, but you will have consumed a very large part of the encyclopedia in doing so. You will still not yet know all about the Indian Wars; you will have acquired only what is explicit in the system. There remains an infinity of facts that can be inferred from what you have. For example, assume that the following pieces of information are stored directly in the system: (1) Sam Spade was a corporal under Custer at Little Big Horn, (2) Custer and his men were wiped out at Little Big Hom, and (3) the battle of Little Big Horn was fought on June 25, 1876. Although the date of Sam's death is not given directly in those statements, we can infer that he died on June 25, 1876. That inference is fairly trivial, but language understanding systems have to make such inferences, and a powerful system must make nontrivial inferences. Logic, the study of methods of inference, has an old tradition and a contemporary mathematization, but many have asserted that it is not yet sufficient or, to the extent that it does suffice, convenient for models of th.ought. Shapiro is among those building computer systems for cognitive networks in which the basic notions of implication (Shapiro and Wand 1976) and even conjunction (Bechtel and Shapiro 1976) are revised. Inference making is dangerous work, since it leads from a few assumptions to a vast horde of conclusions. Joshi and Rosenschein (1976) are examining tile interaction of inference with level of abstraction; most problems are posed at a high level of abstraction, and most systems of inference go at once to the bottom, where everything is much expanded. Their system keeps its material at as high a level of abstraction as possible using some tests to recognize situations in which a more concrete analysis is demanded. A system consists of a cognitive network, logical devices for operating within the network, and linguistic processes for moving back and forth between the network and language. Among the most richly developed systems is that of Reger Schank and his colleagues, which speaks of conceptual dependency (see Schank and Rieger 1974, Schank et al. 1975a,b, Cullingford 1975, Goldman 1975, Lehnert 1975, Meehan 1975, Rieger 1974). As their work has progressed, they have come to deal more and more directly with a small number of semantic primitives, and they have recently introduced scripts to hold knowledge together in large packages. En this system, a script is a causally organized conceptual structure .representing actions performed



in stereotyped situations, such as eating at a restaurant or taking a bus ride. The Script Applier Mechanism (SAM) is designed to understand stories which closely follow scripts (Schank et a1. 1975b). Each script is organized in relation to a goal which implies mutual obligations among participants in the script. The goal of the restaurant script is INGEST. The customer takes food, but he must payor suffer the consequences. The script formalism allows some variation, as between the fast-food stand and the gourmet showplace, and a sequence of scenes. SAM can make a long story short, or it can infer details and make a short story long. It can translate, and it can answer questions (Lehnert 1975). To answer a question it fust identifies the question as belonging to one of five types for which different kinds of answers are appropriate, then searches and makes inferences to obtain an answer. It can be set to give terse or full resl?onses. In Meehan's system to generate stories (1975), a person makes some decisions about the plot and the system uses planning structures to develop the action. The idea of simulation is perhaps unattractive to many humanists, who may, even more than scientists, crave contact with reality. But a group of psychologists in San Diego (Norman, Rumelhart, and the LNR Research Group, 1975) contend that cognitive networks resemble the content of human thought in more than gross outline. They have a computer system, and they have human experimental subjects. The computer cannot store a picture of a building, at least as LNR uses it, but it can store a concept of a building, with parts in specified arrangement. Experiment shows that human subjects make mistakes that would be hard to explain if they remembered the building by means of a picture: why, for example, should the facade of a broad, low building grow narrower and taller in pictorial storage? Again, why should memory for pieces on a playing board depend on the rules of the game? Their system is fun and rich, their reports rewarding to anyone who hopes that cognitive networks will help understand understanding. In the search for primitives, Wilks has been vigorous (1975a,c,d, 1976a,b). He has amassed some 70, and uses them in a system for analysis of text and for translation. Ultimately, the primitives of human thought no daub have to be related to the elementary capacities of the person for perception and reflection; Locke told us to look for simple ideas, and we are still trying.

Speech Processing Systems Let us now return to language, having taken note at least of the problems of structures of ideas, For many language is speech, and the interpretation of poetry, and probably of some prose, is fixed by oral and not by written representation, Of the substantial work in progress toward the analysis of spoken language reports can be found in special issues of IEEE Transactions of Acoustics, Speech, and Signal Processing (Vol. 23, No.1, February 1975) and Proceedings of the IEEE (Vol. 64, No.4, April 1976), and in Reddy (1975), who has also (1976) reviewed the field in an essay that we



A speech understanding system has a semantic or cognitive component for two reasons. One is that the object of analyzing the speech signal is to connect it with the body of knowledge so that a question can be answered or a command executed. The other reason is that without such knowledge of what might be said it is practically impossible to decide What has been said. If one types text into a computer, every character is clear and the boundaries marked by spaces furnish good help in further processing. The input from a microphone is not so satisfactory: the audible signal is murky, the identity of segments not sharply defined, the boundaries lost in overlapping. Words have different sounds in different contexts of occurrence and gaps within a word may be more pronounced than gaps between words. The strategy adopted in most places to overcome these deficiencies is to use factual, semantic, and syntactic knowledge all at once, each as a source of hypotheses about what is in the acoustic signal. Three systems are highly developed in the United States, SPEECHLlS~ the product of Bolt Beranek and Newman, Inc., operates with knowledge of lunar geology and of travel budgets (Bates 1975, 1976, Nash-Webber 1975, Rovner et al, 1975, Woods 1975, Woods & Makhoul ]974)_ VDMS, the joint product of Stanford Research Institute and System Development Corporation, operates with knowledge of submarines (Becker and Poza 1975, Hendrix 1975, Paxton 1975, Silva 1975, Walker 19:75). And REARSAY is the product of Carnegie. Mellon University; version 1 operated on chess, version n on news stories (Erman et al, 1976, Reddy and Newell 1974, Reddy et al. 1976). Different in Important details, these systems are alike in general strategy. Each knows about a small







universe of discourse, and -any incoming signal either solid-state devices are almost ready to receive our blueprints and purchase orders. is about this universe or it is incomprehensible. Keeping the vocabulary and syntax small is one reason for treating a small world, but keeping the Pro spero: Computing Shakespeare facts and principles of fact construction small is more urgent. Each system has a number of indepenLet us create a fantasy, a system with a semantics dent components (semantics, syntax, acoustics, so rich that it can read all of Shakespeare and help discourse structure, prosody) that receives informain investigating the processes and structures that tion about the input and possible analyses and each comprise poetic knowledge .. We desire, in short, to adds its own hypotheses and rejectsfhose from reconstruct Shakespeare the poet .in a computer. other sources: "That interpretation maybe good Call the system Prospera. syntax, but it is impolite discourse." A control How would we go about building it? Prospera is component gives each of the other components 11. certainly wen beyond the state of the art. The share of the computer's time and weighs the computers we have are not large enough to do the opinions issuing from all components. One is job and their architecture makes them awkward for reminded of the blind men trying to come to OUf purpose. But we are thinking about Prospera agreement about something they have encountered; now, and inviting any who will to do the same, Arbib (1975) suggests that different portions of the because the blueprints have to be made before the nervous system have the same problem. machine can be built. First we must have a critical To bring everything together, the control compopurpose for Prospero to execute, then an arehitecnent of SPEECHilS has an overall strategy for ture for its hardware and software, then, perhaps, co-ordinating information from many sources into a the machine. single Interpretation (Rovner et el, 1975). VDMS With the machines we have now, we can formuhas a best-first parser. In HEARSAY a recognition late and test on a very small scale some ideas about overlord, .ROVER, does the oo-ordlnatron. Asorigihow the wholeaffrur is to be conducted. We can nally implemented, each module was -a separate job analyze '8 sonnet or two, and some passages from in a timesharing system, the modules scheduled by the plays ~ perhaps the soliloquies .. We can separate the monitor. Now a c1oselycoupled network of the plot structure from the play and unalyze the small rcomputers shariiitg a sfngle memory is being plots in outline form without burdening ourselves put together (Erman et al, 1975). In the new with the weight of their detailed realization 'in the. implementation, the small computers can work drama. simultaneously. The general idea is to represent the requisite world knowledge - what the poet had :in his head Nothing beyond the availability of sequential digital computers and the lack of any other kind ~ and then investigate the structure of the paths which are taken through that world view as we forces us to perform our calculations sequentially, The architecture of the standard computer is not move through the object text, resolving the meaning what we need (Lipovski & Su 1975), for the brain of the text into the structure of conceptual interis not sequential, but ridiculously parallel (Arbih relationships which is the semantic network. Thus 1975). To do one task and then another requires the Prospera project includes the making of a. semantic network to represent Shakespeare's version time; to do two tasks at once requires equipment. At present the speed of operation of computers in of the Elizabethan world view. The creation of the network for one sonnet or attempting to understand speech is not very good, one passage frorn- a play must start from scratch, a nor, for that matter, is the speed of logic ina cognitive network high enough. For the future, we sort of analysis that has never been done before. To can imagine an architecture like that of a cognitive analyze another sonnet or another passage, more network: each concept is a tiny computer, and information must be added to the world-view concepts are linked so as to iTIherit from one model, but if the new passage is chosen so that another, to define one another, etc. In such an there is some semantic overlap with the old, then it active network, knowledge of acoustics, syntax, and will not be necessary to start from scratch. The many other matters could be stored. Since we have growth of the world-view model beingcumulative, it some inkling of how to make such a computer, it is is not necessary to create a new one for each Dew time to make our theories clear, for the makers of text.




The cognitive mode! of Shakespeare's world view, or as much of it as is needed for one sonnet, is only the background. A program, our model of Shakespeare's poetic competence, must move through the cognitive model and produce fourteen lines of text. After so many have 'asked about the nature of poetic competence, it may seem foolish to try again, but of course the project is not to be entirely original; the program can formalize any good notion about poetic competence, new or familiar. The advantage of Prospera is that it takes the cognitive model as. given - clearly and precisely - and the poetic act as a motion through the model. Instead of asking how the words are related to one another, we ask how the words are related to an organized collection of ideas, and the organization of the poem is determined, then, by world view and poetics in unison. The sources of hypotheses about poetic competence are the whole of literary criticism, psycholinguistics, and the creative intuition of the builders of Prospera as they try to make it produce one sonnet, and then another, and more after that. We take it, since all three must be exploited to the full, that it is time to do some serious thinking about the possibility 'of experiments in literary psycholinguistics, and that Prospero will arrive sooner if many students of computers and the humanities ac quaint themselves in detail with cogni tive networks' and. the linguistic processes that connect them with text. The general strategy for investigating passages from the plays is the same as for the sonnets: pick your passages, model the requisite portion of the world view, and run it. But do not pick passages at random, choose them with care - Ulysses on degree from Troilus and Cressida, «What a piece of work is man" from Hamlet, one of Benedick's soliloquies from Much Ado about Nothing. Better still, choose passages spoken by different characters in one play. If, as Thomas McFarland has argued, Othello and 1ago are two sides of one character (1966, pp. 72 ff.), then it would be appropriate to choose passages spoken by each in isolation and work toward a model of the linguistic style peculiar to each, a model of each one's character. Then one could see what these models tell about passages in which they speak to each other. The character of Othello occupies a certain type of space in the world view, the character of 1ago occupies a rather different type of space. How do you characterize those spaces? Plot analysis requires a different strategy. We

want to work with the bare bones of the play, the framework. Initially. that skeleton will have to be extracted from the play by intuitive methods for doing it by machine would require Prospero, which will not exist for a while yet. Read Propp on the folktale, Levi-Strauss on myth, Northrop Frye on criticism (The Anatomy) and on Shakespeare (Fools of Time, A Natural Perspective) and anything else which might help. Then create a formal theory of plot structure and describe the plots which interest you in terms of that theory. There are at least two lines of investigation; (1) assuming that Shakespeare borrowed his stories from others, but changed most of them considerably, you can use your model to investigate the nature of the changes; or (2) you can compare the plats of different plays, such as Much Ado about Nothing, a comedy; Othello, a tragedy; and The Winter's Tale, a romance, all of which involve a man's suspicion that a woman he loves is unfaithful. Each of these men is mistaken. In the comedy, this suspicion falls between betrothal and the wedding (in the Claudio-Hero subplot). Othello does not begin to doubt Desdemona until after the wedding, but he doubts her before their marriage has been consummated. Hermione has given Leontes a son and is pregnant with another child before her honor is suspected. Then note that the relationship between the man deceived and the deceiver is remote in Much Ad@, close in Othello (lago is Othello's trusted aide), and identical in the Tale (Leontes deceives himself). Do we have a pattern here? The state of the art will support initial efforts in any of these directions. Experience gained there will make the next step clearer. If the work is carefully planned, knowledge will grow in a useful way. The world-view model will become larger and more sophisticated, character models will be developed and knowledge of plot structure will grow. The blueprint of new hardware can emerge and lead to the building of a machine adequate for the analysis of an entire play. That job can begin by taking a model of the plot structure, models of the characters, and the world-view model and refitting them so they can work in one large system. We have no idea how long it will take to reach Prospera. Fifteen years ago one group of Investigators claimed that practical automatic translation was just around the comer and another group was promising us a computer that could play a highquality game of chess. We know more now than we did then and neither of those marvels is just around the present comer. Nor is Prospera. But there is a





difference. J'csell a translation made by machine, one must first have the machine. Humanistic scholars are not salesmen, and each generation has its own excitement. In some remote future may lie the excitement of using Prospera asa tool, but just at hand is the excitement of using Prospero as a distant beacon. We ourselves and our immediate successors have the opportunity to 'clarify the mechanisrris of 'artistic treationand interpretation. We may well value that opportunity, which ban come but once in intellectual history. But when, or if, we or our descendants actually possess a Prospera, what exactly will it be? A system capable of writing new plays, perhaps, but what would it take to write a playas good as Shakespeare's worst? We' can work on the representation of world views, and we can work also on linguistics and even on poetics, the control of interplay between sauna ana sense, but no one can say which is harder .. Prospero may be an extremely soplustieated research tool; as a model of the human mind its foremost purpose is to help In the creation ofa still better model of mind and its evolution, Those who see a threat to playwright, scientist, or philosopher in the machine that can simulate thought must set a limit on the mind which is not implicit in any system with recursive abstractions. If we ever obtain a machine that can apply patterns of thought to the phenomena of life, we shall most likely ponder on the machine and ourselves for a while, and see then, as we cannot until the machine begins to help us, what a narrow sector, what a shallow stratum of_ the human essence is captured in the design of the machine.

(Note: American Journal of Computational Linguistics, or AJCL, is the primary English language journal in the field. It is published quarterly on microfiches and contains articles of some length as well as. current bibliography. IEEE Transactions in Acoustics, Speech, and Signal Processing, Communications of the ACM, and the International Journal [or Man-Machine Studies frequently carry articles on natural language processing, Directories of persons active in the field can be found from time to time in AJCL. Information about the preprints of the 1976 International Conference on Computational Linguistics, rCCL, can be obtained from Martin Kay, Xerox Palo Alto Research Center, Palo Alto, CA 94304_)

Anderson, John R. Computer simulation of language acqutsiticn: A first report. In Salsa ,ed., Information Processing and Cognition: The Loyola Symposium: New York, John Wiley and Sons (975), pp, 295-349. Arbib, Michael A. Artificial Lntelllgence and brain theory; IJnities and diversitj.,s. Computer and Information Science Department, University of Massachusetts, Technieal Report 7SC-8 (September 1975). Bates, Madeleine. The use 0. f syntaxina speech. understanding system, IEEE TI'(lIJS(I.cnol1sn Acoustics, Speech, o and Signal Processing 29, I (Fe bruary 1975), 112-117. Bates, Madeleine. Syntax in automatic speech understanding. AJCL 4S (I ~n6). Bechtel, Robert and Stuart c..Shapiro. A lQgie for semantic networks. Computer Science Department, Indiana Unlversttv. Technical Report No. 4'7 (March 1976). Becker, Rich!trd·d Fausto Poza. Acoustic research. in speech understanding. IEEE Transactions ot! ACOU$tics, Speech, and Signal Processing 23 (t975), 416.426. Benzo.n, WilIiam. Cognitive networks and literary semantics. Modem Language Notes (The Centennial Issue, Summer 1976). Collins, Allan M. and E(izabeth F. Loftus. A spreadingactiv atio n theory of se m 911 t ic PTa cessi n g. Psycho. logical Review 82 (J 975), 407-428. Cullingford , R.E. The application of script-based world knowledge in an integra leti story understand] ng system, Abstract Jn AIeL 46 (1916)\ 31. Preprints of lCCL 1976. Erman, Lee D., Richard D. Fennell, Victor R. Lesser and D. Raj Reddy. System organization for speech understanding: Implications of network and multiprocessor c 0 mp u r er arch ite cture, IEEE Tral'lSl!l!:tions on Computers 25 [Aprt] 1916), 414-421. Frederiksen, Car! H. Representing I·ogical and semantic structure ofknowJedge acquired from discoutse. Cognitive Psychology 7 (l975), .37 J -48 S. Friedman, Joyce with Thomas H. Bredt., Robert W. Doran, Ball' W. Pollack and lheodore S. Mariner. A Computer Model of Transformationat GrammQ-r. New York; American Elsevier (197 t). Goldman. tleil M. Sentence paraphrasing from a €onceptual base. Communications oj the A eM 1 S (l975), 96. Gregg; Lee W. Knowledge and Cognition. Potomac. MD: Lawrence Erlbaurn Assocla tes (I 974). -Grlsh man, R al ph. A surve y 0 f syn tact ic ana lysis pro ce d ure s for naturatIanguages. AJCL 47 (197 6).. Hays, David G. Cognltlve Structures. Manuscript (J 975). Hays, David G. On 'alienation'; An essay in the psycholinguistics of science. In Alienation, ed, Felix. Geyer, in press. Lelden: H.E. Stenfert Kroese .B.V. (1976). Hendrix; Gary G. Semantic processing for speech understanding. AIeL 34 (1975), 34-.48. Hockett, Charles F. The State q{ the Art. The Hague: Mouton (1968). Josh i, Aravlnd K. and Stanley 1. Rosenschein, Problems of inferenclng and its relation to decompcstnon. Abstract in AJeL 48 (1976), 38-41. To appear in SMIL. Kaplan, Ronald M. A general syntactic processor. In Randall Rustin, ed, Natural Language-Processing; New York: Algorithmics Press (1973), pp .. 193-241. Klein, Sheldon et at, ModeUing Propp and Levi-Strauss in meta-symbolic simulation system. Computer Sciences Department, University of Wisconsin. Technical Report No. 226 (October 1974). Kin tse h., Walter, The Representation of Meaning .111 Memory. Hillsdale. NJ; Lawrence Erlbaurn Associates (1974).


BEN .zON AND HAYS Schank, Roger C. et 31. Conceptual Information


Lehnert" Wendy, Question answering in a story uudersrendmg system. Department of Computer Science, Yale University, Research Report No. 57 (December 1975). Lipovski, G" Jack and Stanley Y.W. Suo On non-numeric archltecture, SIGIR Porum 10, 1 (Summer 1975), 5.:20. Macksey, Richard and Eugenio Donato, The Langllages Of Criticism and the Scienc/!;$ Qf Man.' The Structurailst Controversy. Baltimore: Johns Hopkins (1970). McFlI11anil, Thomas. Tragic Meanings In ,Shakespeare. New York: Random Bouse (L9'66). Meehan, James R. Using plBnning structures to generate stories. AJCL 33 (1975), 78-94. MiUer" George A. Toward a third metaphor for psycholinguistics: 1n Weime,r ami Palermo, eds. Cognition and tile BJimbolfc Processes. Hillsdale, NJ: Lawrence Erlbaum Associates (1974), pp. 397-413. Miru;ky, Marvin, Semalltitl lnfi;mrw.tion Pl'ocessi1}g, Cambridge, 'rite MIT Press (r96S), Minsky, .Marvin, A framework for representing knowledge. In P.B., Winston, ed, The Psychology of Computer Vision. New York: McGr.Bw-HiII (197S), pp. ZU-277. Nssh-Webberv Bonnie. Semantic support for a speech under,sta:ndiIlgSYSlem, IEEE Transactions 011 AcouStics, Speecll, and Bigflll/ ProcesSing 23, 1 (Febru!lry l'97S), 124-12'9. Norman, Donald A., David E. Rumalhart and the LNR Research Group. Explop.atiol1s in COgrlitiolJ. SaD Francisco: W.H, Freeman (J 975). Paxton, William H. A best-first parser. JEEE Transactions on ACOlIStfl$, Speech, ana 8igrJa/ Processing 23 (1975), 426-432. Reddy, 0. RaJ. Speech Recog!{l.ition; Invited Papers Of -rile IEEE Symposium. New York: Academic Press (l975). Reddy, D .. Raj. Speech recognition by .maehrne: It reView. Proceedings of the rltEE64 (April 1976), 501-531. Reddy, D. Raj, Lee D. Erman, Richard D. F ..nnell and Richard B. Neely. The Hearsay-I speech understanding system: An example of the recogntnon pro. cess. IEEE Transactions 011 Cempute~$25 (April 19'76), 422-431, Reddy, D. Raj and Allen NewelJ. Knowledge and its representatlon in a speech understandjng system, In Lee W. Gregg, ed., K.1I'Ol\iIetjge and Cognition. Lawrence Erlbaurn Assoclates (i 974), pp. 253.285. Rieger, CilJlCk. Understandtng by conceptual inference. AIeL 13 (1974), Rovner, Paul, Bunnie Nash.Webber and William A. Woods. Control can cepts in II speech. uaderstandlng system. IEEE Tnmlj'ltctio'ts on Acoustics, Speech, an(J Signal . Processing 23 (February 1975:), 1'36-140. Schank, Roger C. and C.J. Rieger rn. Inference and the computer understanding of natural language, A'rti!icfal1ntelligtfu;e 5. (1974)" 3'73-412.

understander, Department of Computer Science, Yale University. Research Report 43 (August 1975b). SCl"agg, GI'eg W. Frames, planes and nets: A syn'lllesill. - - Worl<ing papers of the Instttute for Semantic and CogniHve 'Studles, 'No. 19 (1975). Shaptro, St,uart C. and Mitchell Wand. The relevance of' - relevance. Computer Science Department, Indiana University, Bloomington. Technical Report No. 46 (March 1976). Sheil, B..A .. On context-free parsing. Abstract in AfCL 48 (1976), 5.8-6.3. Toappear in SAUL. Silvao, GeorgeUe. On automatic speeeh-understandtng systems ... Computers arid the Humanities 9 (September 197 S), 237·24·4. 80180, Robert L., ed. Contemporary Issue« inCognft/.ve Fsyclt%g)!'. Washington, DC: V.H. Winston & Sons (1973). Bollia, Robert L., ed. Information Processing and'Cogl,1ition. Hillsdale, NJ: Lawrence Erlbaum Associates (197S). .. Tulving, Entiet 'and Wayne Donaldson, eds, Organization of Memory. New York: Academic Press (l972~. Walker, Denald. E. The SRI speech understanding system, fESE Tl'pnsacti.onsol1 Acoustics, Speech. arid Sigrlal Processing :2 3 (1975), 39'7-416. White" Mary J, Cognitive networks and world view; The metaphysical terminofogy of a millenarian commUIlHY. Dissertation, SUNY Buffalo (1975). Wilks, Yo.tiek, An intelligent analyzer and understander of English. Communications of 'the A CM 1 S (197 Sa), 264.274, Wilks, Yorick .. Putnam and Clarke and mind and bOdy .. British Journal for the Pili/wop II)' of Saiel~ce '26 (19'7Sb), 213-225, . Wilks, Yorick. A preferenti;tl, pattern-seeking, semantics fOT Datura! language inferen ell. Artfrwia/ Inte,lligence 6 (1975c), 53-'74. Wilks, Yo.rick.Seven theses on arttflclal intelligence and, natural language. Institute for Semantic and Cognitive Studies, Working Paper 17 (197 5d). Wilks) Yorick. Natural language understanding systems within the A.I. paradigm: A survey and some comparisons. AIeL 4{) (1976"1). Wilks, Yorick. Frames, stories" scripts, and fantasies, Abstract in DCL 48 (1916b), :U. I'reprints of ICCL 1976. Winston, Patrick H., ed. TIle Psycholqgy of Computer Vision. New York: McGrllw-Hill (1975). Woo ds, WilHam A. Mo tlvatlon and overview of SPEECHLIS: An experimental prototype for speech understanding research. IEEE Transactions on Acoustics. Speech, and Signal Frocessfng23, 1 (Februa ..y .I 975), 2-11. Woods, William A. and J. Makhoul. MecboIl.icol mference problems in continuous speech understan din g'• Artt!ici'al1iltelligenae 5,. 1 (]974), 73-91.

New York: American Elsevier (1975a). Schank, Roger C. and the Yale AI Project. SAM - A story

• I .,

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.