Go to Semiotic Review of Books Home Page
Go to SRB Highlights
Go to SRB Archives
This article appeared in Volume 3 (1) of The Semiotic Review of Books.
A Computational Model of Metaphor Interpretation. By James H. Martin, Boston/San Diego/New York: Academic Press 1990, 229 p., ISBN 0-12-474730-2
It is a widespread opinion that the field of Artificial Intelligence (AI) is a realm of myths and fairy-tales. The short history of Al is often viewed as a history of splendid announcements: who does not remember the high expectations raised by the idea of a General Problem Solver, and the disillusionment caused by the narrowness of its realization, the so called Expert Systems. Nevertheless, thoughtful positions like those of Weizenbaum (1976) or Dreyfus & Dreyfus (1986) are rare, even if frequently cited. Contemporary researchers in Al who are engaged with problems like Natural Language Processing ( NLP) also spread unbridled optimism. As everybody knows, one of the worrisome problems of NLP is the under standing of metaphors because they don't possess a fixed meaning. Usually, developer s of N L P system s either ignore metaphors since they cannot cope with them (then they say that metaphor is unimportant anyhow), or they claim to provide a fully satisfying algorithm to process all kinds of metaphors ( with little practical success).
At first sight, James H. Martins's Computational Model of Metaphor Interpretation does not share this effusiveness. His approach starts from a practical example: if a naive user wants to use the UNIX operating system, he or she can obtain advice from the UNIX Consultant system ( UC), which is a natural language system, so that the naive user can address it without knowing computer language. But UC doesn't understand metaphorical utterances which occur very often in technical domains. To close this gap, Martin has developed MIDAS (Metaphor Interpretation, Denotation and Acquisition System). MIDAS is meant to support the UC system by interpreting the learning com m on U N I X domain n metaphors as they are applied by the user. Thus the system should be able to understand questions like "How can I enter U N I X ? " or " How can I kill the process? " .
Martin' s book consists of eleven chapters which boil down to three focal points: the phenomenon of metaphor, the computational model of MIDAS, and a review of previous metaphor processing approaches. A quick reader, or one who is not well acquainted with Al but interested in metaphor theory, may focus his or her attention on Chapter 1, which introduces Martin's approach; Chapter 3, which deals with a theory of understanding metaphor; and Chapter 6, which is about learning new metaphors by extension of known ones. Chapter 2 and 10 give a review of related research; Chapter 4 and 5 show how conventional metaphors are processed by MIDAS. Chapter 7 to explain the system of interpreting and learning new metaphors; and Chapter 11 draws some conclusions and designates some topics for further research. Without going too far into the computational details, I will examine now some of the main points of Martin's approach from a semiotic and philosophical viewpoint.
Martin does not claim to know the Royal Road to processing any metaphor. What he intends is to develop a method of understanding conventional metaphors. It is a matter of fact, as Martin points out, that previous approaches to processing metaphors inevitably failed because they treated metaphors as exceptional cases. Martin views conventional metaphors rather as part of conventional language, and emphasizes that there's no need to treat them as exceptions. He sees everyday natural language to be usually interspersed with innumerable conventional metaphors which often are unseen or unspectacular, but nevertheless indispensable. Martin finds support for this view in research in cognitive sciences and psycholinguistics, especially in the approach of Lakoff and Johnson (1980) . They take metaphorical structures to be part of conventional language as well as the central component of the conceptual system underlying the structure of language and of experience.
This alone is enough to make the book remarkable. In away, its fundamental assumptions on metaphor overcome the barren confrontation between literal language as normal and regular, and metaphor as anomaly or unimportant rhetorical ornament. However, by treating metaphor as part of pure conventional language, Martin is led to some further far reaching, but questionable background assumptions: he views conventional and explicit language knowledge as being enough to understand metaphors. He even treats novel metaphors in this way: "It is this explicit knowledge of metaphor that accounts for both the conventional and novel metaphor that we encounter" (p. 2). By referring in this topic exclusively to the research of Lakoff ( 1987 ) and Turner/Lakoff ( 1988 ), he tries to justify the claim that novel metaphors can in principle be understood on the basis of explicit conventional knowledge ( see 110).
At this point one begins to suspect that Martin's conventionalism might be a bit exaggerated. The suspicion is fostered by the fact that Martin does not take the broad discussion about metaphor in the humanities into consideration at all. Therefore it is useful to elucidate the crucial difference between Martin's background assumptions and most metaphor theories in the humanities (see e.g. Black 1962, Goodman 1968, Ricoeur 1975, and Kittay 1987): In contrast to Martin they view metaphor as unconventional use of ordinary language and metaphorical meaning as a result of a creative process, which is based upon implicit rather than explicit knowledge. Since they take metaphor to be part of ordinary -- but not necessarily of conventional -- language, they do not confuse ordinary language with conventional use of language and explicit knowledge.
The question whether Martin is aware of this conflict, and if so, how he resolves it, requires careful examination.
Martin's main claim is that representation as well as interpretation of metaphors, and even learning of new metaphors, can be achieved by means of MIDAS. This knowledge-based computer system is supposed to process explicit and declarative knowledge about conventional metaphors. If it really worked, it would not only be a great success in Al, but also a prolific contribution to semiotic, linguistic and philosophical research on language: it would reduce worries about metaphor. The question is, whether MIDAS keeps its promise.
Martin takes metaphorical meaning to be not single word meaning, but as being a result of the association between two concepts: if we say for example "The fire has eaten up the whole wood," then of course the fire hasn't really eaten any wood at all. But the activity of the fire, destroying by burning, is being viewed as a process of consuming a meal. The 'viewing-as' means that we interpret target concepts ( the fire and the wood) in the light of a source concept (the eating). This point of view is close to modern metaphor theory in linguistics and philosophy, particularly with regard to the so called 'interaction view of metaphor', introduced by Richards ( 1936) and Black ( 1962) . There, the metaphorical process is looked upon as creating new meaning through an open-ended interaction between the frame of a sentence ( i.e. target) and the metaphorical focus ( i.e. source). The result of the interaction depends on contextual factors: speaker's and hearer's intentions, implicit knowledge, situational and communicational conditions. These factors cannot be known adequately and sufficiently in advance. For this reason Danto ( 198~ designates metaphor as intensional context, as it contains in every application the manner of its meaning within its specific context.
On the contrary, however, Martin regards this process only as combination of predefined elements: he holds that understanding metaphor can be achieved by association between concepts which represent explicit ( hence predefined ) knowledge. This evokes a second suspicion: in the rules of natural language, but the constraints of computational language are the basis and the background of Martin's metaphor theory. Only if metaphor is reduced to a combination of predefined, explicit and declarative knowledge-elements can it be processed by a knowledge representation language. In the book the theoretical chapters, 3 (on metaphor understanding) and 6 (on metaphor extensions come before the chapters in which the theory is applied to the computational model We can thus presume that this theory is not the basis, but the result of the latter. Martin neither mentions nor denies this possibility, and perhaps he is not conscious of it Without further methodological reflection, he claims to arrive at his metaphor theory only by observing typical kinds of metaphor and by exploring the systematicities of metaphorical language (see 44 and 53). Finally, the untroubled self evidence with which Martin's metaphor theory is developed and then combined with the computational model strengthens the suspicion.
MIDAS is based upon a knowledge representation language called KODIAK, an extended sem antic network system in the tradition of the Al language KL-ONE, which is said to be very sophificated. The KODIAK network links knowledge elements by means of an inhertiance mechanism and a hierarchical organization of concepts. Conventional metaphors, then, can be represented as sets of of associations, or relations, between source and target concepts. To be represented in this way, they must be categorized in a system of abstraction hierarchies. In this hierarchical arrangement, common background metaphors, called "core metaphors", hand down their sets of associations to "extended metaphors" which can possess new associations, as long as these are coherent with the core metaphor. If several extended metaphors stem from the same core metaphor, then they are called "core-related".
To explain with one of Martin's examples (p.42 ): in the sentence "Mary has a cold", "to have" serves as a core metaphor, as it views infection as possession. Extended metaphors then have to include the "infection-as-possession" status, as in the case of "John got his cold from Mary": to become infected, the infection must be possession, only then is the extension "becoming infected-as-getting" possible. A sentence like "Mary gave John a cold" is core-related to "John got his cold from Mary" because both stem from the same core metaphor "infection-aspossession".
If metaphors are classified in abstraction hierarchies like this, similarities and differences among metaphors can be grasped. Similarities are represented in shared explicit relations and concepts of the target domain. The sentences "Mary gave John a cold" and "Mary gave John an idea" ( seep.48) have the same source but different targets. What makes them similar is that both "idea" and "cold" share the abstract concept of being viewed as an object that can be possessed and also be transferred from a sender to a recipient.
Again, it seems that computational constraints, and not the structures of natural language, are the basis for this arrangement and particularly for its strict requirements, like coherence, derivability and preservation of the hierarchy. But, alas, natural language is not prearranged in such clear-cut and logical hierarchies: the broad discussion in the philosophy of language, especially since Wittgenstein's 'pragmatic turn', has shown that it is rather questionable whether natural language meets such formal requirements at all. (In contrast to Martin, the developer of KLONE clearly recognized these problems: see Brachmann 1985.)
Consequently Martin now examines the problem of "missing" metaphor only in the light of these requirements, not in respect of the real uw d metaphorical utterances. Thus he asserts that the sentence "How can I give birth to a princess?" fails to meet the requirements and therefore must be refused as missing metaphor: The birth event has the restriction that the created thing and the mother are the same type of identity. This does not correspond to the target domain adequately. (The net result is that the connection between the core metaphor and the target concept of the creation of the process does not match the connection from the core to the source concept of birth in a consistent enough manner to allow for its use" (p. 951, my italics). So what? Natural language conventions do not interdict this use. Everybody will understand metaphors like this even if they might be ugly or give a lopsided view. Usually we can draw our conclusions about metaphorical meaning from contextual, situational and communicational factors as well as from language knowledge.
Martin, however, explicitly avoids the use of contextual information, mainly for the"practical reason that the UNIX Consultant testbed for the theory had no context wider than the immediate sentence in which the metaphor occurred" ( p.110) . In this way, he digresses more and more from the conditions of natural language. The abandonment of contextual factors forces him to replace this crucial source of information with the hierarchical arrangement and its requirements. Presumably it is only because of the small context of UC (which excludes other contexts) that his hierarchical system works at all. Martin has to pay for this with the exclusion of metaphors which have no fixed position in the hierarchy or which do not fulfil the requirements. Unquestionably, these "naughty," but living metaphors are more interesting, more surprising and more creative that those represented in a fixed hierarchy.
But yet, perhaps the hierarchies of MIDAS are extensive enough to cope with easy common metaphors? As we know from a marginal note on p. 105, "MIDAS is attached to a small prototype version of UC built to demonstrate the applicability of the metaphor system" (my italics). There is nothing more to be found about he capacity of MIDAS and the degree of its realization in the whole book. Is this a new example of the above mentioned splendid announcements of Al, or is it only thoughtlessness? I n any case, to give the reader a comprehensive and honest picture of this approach it would have been useful, if not essential, to take into consideration the context-dependency of natural language as well as the question of the system's capacity, particularly with regard to a possible combinatoric explosion (the stage where a system breaks down because of its own relational complexity).
Apart from this serious proviso, Martin's computational model seems to work quite well. To capture conventional metaphors in KODIAK, so called "metaphormaps" have to be established. They provide the "metaphor-sense" by containing all required associations linking source and target concepts. The question remains unanswered, however, how to find the adequate associations to represent a metaphor beyond the fulfilment of the formal requirements. This concerns the unsolved ( and unsolvable?) question of possible anthologies in language as well as in nature ( see Kuhn 1979) . Still, assuming that a sufficient number of concepts and associations between concepts is established in MIDAS, metaphor interpreting can start. This is to be done by a subsystem of MIDAS called MIS ( Metaphor Interpretation System) .
MIS proceeds in two main steps: the first step is a syntactic parse and preliminary semantic representation. This stage is called "Primal Content" as it "represents the meaning of an utterance that is derivable from knowledge of the conventions of a language, independent of context" ( p90). In a way it is related to literal meaning, but in contrast to the latter, it is only something like a first, unchecked hypothesis about possible meaning. The primal content, as an intermediate stage, praises a set of constraints ( like roles and respective role-fillers of the concepts) with which the meaning can be selected. The second step is the final interpretation which produces the "Actual Content" as final meaning. The n nal interpretation is based upon two inference processes which select the appropriate meaning. The first is the replacement of an abstract concept by a more specific one ("concretion") ; the second is the replacement of a given source concept with the corresponding target concept ("metaphoric unviewing'). The fundamental operation of both of these inference processes is the constraint checking.
The whole MIS preview takes the form of a six-step algorithm: it first parses the sentence and then lists the constraints of the involved concepts. After that it collects possible interpretations and validates them by testing them against the constraints. Finally, it applies and then returns all consistent interpretations. If it ends up with several ambiguous interpretations, the system uses a specific heuristic to choose among them. It reads as follows: "select the interpretation that most tightly matches the constraints posed by the input concepts (p.104). Again, for this heuristic the restrictions of MIDAS and not the regularities of natural language are basic. The problem of ambiguity can be solved by this heuristic, but there is no guaranty that it really grasps the intended meaning of a metaphor.
Consider for example the question, "How can I kill the process?" MIDAS would check the established concepts and metaphor maps with regard to their constraints, like killing-as-slaying (a living thing), killing-as-terminating (proceedings), killing-as-dismissing (an arraignment), killing-as-eating (consume whole contents of), or killing-as-defeating (an opponent). Since the target concept "process" satisfies only the constraints of killing-as-terminating, the system would come to the (at first sight) correct interpretation that the question meant "how can I terminate the process?". But let us imagine a situation in which two users interrupt an exhausting UNIX session because they are thirsty. Greedily one of them utters a colloquial: "Well, let's kill this bottle" after the drink they go back to the UNlX to start again a process that did not work well before. Now the other says: "Now we're gonna kill this process". And then he addresses the computer: "How can I kill the process?" Knowing all these extra-linguistic contextual factors, we immediately understand that this metaphorically means to consume the whole process in the sense of to call it up and to make it run until finished. And it implicitly contains the meaning that, by making it run, the process shall be killed in the sense of being defeated, because it did not co-operate. It also may mean that they want to kill off the problems with the computer, as they are annoyed with it. What we encounter is a complex, but very usual case of metaphor understanding, in which several background metaphors (computer as consumable thing, as opponent, as annoyance etc.) constructs a complex cluster of meaning. This cluster of overlapping metaphorical concepts does not have to be coherent at all (see Lakoff/Johnson 1980). On the contrary, even conventional metaphors draw their creative and cognitive power, their resonance and their open-endedness from the stimulating effects of incoherent and inconsistent structures, and from the transgression of semantic and pragmatic rules (see Black 1977, R icoeur 1978 and Hesse 1984).
Following its heuristic, MIDAS nevertheless would interpret the above utterance as a question about termination. Again we can state the fact that the abandonment of contextual factors forces Martin to restrict MIDAS to small contexts within narrow bound sand with predefined concepts. This statement is not intended to disparage Martin's approach, but it seem s to be necessary to make clear what is really meant when he speaks of "metaphor interpretation.".
Sometimes (presumably very often) MIDAS encounters metaphors that it does not know, i.e. for which it has is no established metaphormaps. Martin's basic assumption here is that new metaphors can best be grasped by the systematic extension of an existing metaphor. As stated above, this is meant both for conventional metaphors which are new only for MIDAS and for novel metaphors which establish new meaning. To achieve this goal three kinds of extension inferences are provided: similarity extension proceeds on the assumption that "a metaphor with the same source concept being applied to a similar target will have a similar meaning" ( p.lll) and it follows the principle of analytical reasoning. As shown above, similarity of targets means sharing the same abstract concept. Core extension is founded on the presupposition that "conceptual core-relationship among words in the source domain will be at least partially previewed in the target domain ( p.ll7) and it follows the principle of core preservation. Combined extension makes use of both similar and correlated structures of a metaphor.
Metaphor extension can be performed by another subsystem of MIDAS, called MES (Metaphor Extension System). Again, the essential criterion for the determination of a new meaning stems from the arrangement of the abstraction hierarchy, its requirements and the heuristic: to find out the meaning of a new metaphor, MES searches and then evaluates a set of candidate metaphors by measuring the conceptual distance from new input metaphor to candidate metaphor. This is done either by working out the length of the core-relationship path from input source concept to candidate source concept, or by determining the hierarchical distance from input target concept to candidate target concept. The conceptually closest metaphor wins the race. The whole process then ends up applying the metaphor and storing the new metaphor-maps.
The idea of metaphor extension seems to be very promising. Yet only a certain type of metaphor is covered by the extension approach, namely that called epiphor. As Wheelwright (1962) points out, epiphor stands for the extension of meaning through analogical comparison. We can draw this comparison by reason of pregiven relations. An epiphor is not novel, but rather something like a new token of a known type, and it therefore refers to the broad field of conventional metaphors. On the other hand, we often encounter metaphors that create novel meaning. This kind of metaphor, called diaphor produces its meaning by juxtaposition and synthesis, without resort to pregiven analogy or shared properties: "The essential possibility of diaphor lies in the broad ontological fact that new qualities and new meanings can emerge, simply come into being, out of some hitherto ungrouped combination of elements" (Wheelwright 1962, p.85). The crucial criterion of diaphor is not similarity, but emotional congruity which, of course, depends on situation, context, intention, feeling and intuition etc.
Without much effort we can predict that the more a metaphor contains diaphoric elements, the less it can be captured by MIDAS. But even easy epiphors that do not follow the criterion d closest concepts will not be grasped, since this criterion (like the above-mentioned heuristic) arbitrarily favours the more obvious, but not necessarily the right interpretation. Thus, we must repeat our reservation with regard to the fundamental assumptions of this approach and particularly to the criterion of closest concepts.
It is the reviewer's duty to speak also about editorial qualities. Put briefly, the book could have been edited more carefully. We can pass over the fact that in the introduction the description of the chapters' contents does not mention Chapter 7 and that the book contains much tiresome repetition (sometime empty and trivial) summaries. Really confusing is the fact that Chapter 2 and 3 are full of misnumbered examples: the illustrative sentences from examples (7) to (40) are referred to in the text as examples (1) to (34). To complete the bewildering situation, examples (28) to (30) seem to be marked with their correct numbers. The same is to be found with Figure 3.1 which ocurs in the text as Table 2.1. The reason for this may be that the overview Chapter 1 was written afterwards, but does not justify sloppy compilation.
Formal complaints aside, I take the view that Martin's book can be read as an example of a regrettable mutual ignorance between AI research and cognifive sciences on the one hand and the humanifes on the other. For one thing, this computational model is one of the most sophisticated approaches to processing metaphor in AI, and perhaps it really works well within the narrow context of UC. For another, it seems that Martin has stopped worrying about metaphor before he ever started. As if the humanities had nothing to contribute to this topic, no pertinent metaphor theory (except certain pscycholinguistic approaches) is to be found in this book. If we judge it from the viewpoint of the humanities, Martin's claim to develop a computational model of metaphor interpretation has failed because of its unfitting notion of metapor. MIDAS cannot really interpret metaphors -- not even conventional ones -- but only process predefined cases of lexicalized metaphorical meaning, constrained by the requirements of the system.
Like the avaricious King Midas of the Greek myths whose touch transformed everything into gold, MlDAS blindly tries to transform every metaphor into fixed meaning. But as the myth goes, King Midas starved, transmuting even his food into gold as he touched it. This symbolic loss of ability to differentiate seems to repeat itself with the 'computational gold' of MIDAS.
Midas' other fault was that he judged music although he was unmusical and he was given donkey's ears to punish his fault. In a similar way MIDAS can be called 'unmetaphorical'. It looks as though MIDAS can expect to receive donkey' s ears ...
Black, M. (1962), Models and Metaphors. New York: Ithaca
Black, M. (1977), "More about Metaphor", Dialectica 31 (1977),431-457
Brachmann, R.J. (1985): " 'I Lied about the Trees' or, Defaults and Definifons in Knowledge Representation " . The AI Magazine (Fall 1985), 80-83.
Canto, A.C. (1981) The Transfiguration of the Commonplace. Cambridge, Mass.: M.I.T. press
Dreyfus, H.L. and S.E. Dreyfus (1986): Mind over Machine. New York: Free Press
Goodman, N . (1968), Language of Art ,Indianapolis: Bobbs-Merrill
Hesse, M. (1984), "The cognitive claims of metaphor" JP van Noppen (ed.), Metaphor and Religion. Brussells: Free University of Brussells, 27-45
Kittay, E.F. (1987), Metaphor - Its Cognotive Force and its Linguistic Structure. Oxford: Clarendon Press
Kuhn, Th.S (1979) "Metaphor in Sciences' In A. Ortony, Metaphor and Thought Cambridge: Cambridge University Press, 409-419.
Lakoff, G. (1987) , Women, Fire and Dangerous Things . Chicago: University of Chicago Press
Lakoff, G. and M. Johnson (1980) , Metaphors We Live By. Chicago: University of Chicago Press
Lakoff, G. and M. Turner (1988), More Than Cool Reason; A Field Guide fo Poetic Metaphor. Chicago: University of Chicago Press
Richards, I .A. (1936) The Philosophy of Metaphor. Oxford: Oxford University Press
Weizenbaum, J. (1976), Computer Power and Human Reason. From Judgement to Calculation. San Francisco: Freeman & Co
Wheelwright Ph. (1962), Metaphor and Reality. Bloomington: Indiana University Press
Bernhard Debatin is research assistant in the Philosophy Department of the Berlin Technical University where he is a member of an interdisciplinary research project on sociocultural problems of using computer-aided design. There he examines problems of language and knowledge processing and of Human-computer interaction from the viewpoint of language-philosophy, communication theory and semiotics. He is also the author of various articles on metaphor theory and cognitive sciences, as with as on interpretation and mass communication. He was editor with Dieter Hirschfeld of Antinomien der Offentlichkeit (1989, ISBN 3-88619-613-5), and with Hans J. Wulff Das Telefon im Spielfilm (1991, ISBN 3-89166-146-0).