Kevin B. Korb
Herbert Simon has long been a leading advocate of a dominant role for symbol processing within artificial intelligence, going so far as to posit, in his and Alan Newell's "physical symbol system hypothesis," that being a physical symbol system is both a necessary and sufficient condition for being intelligent (cf. Newell and Simon, 1976). In his current paper Simon applies his perspective to the analysis of meaning in literature and, hence, to the proper methodology for literary criticism. In attempting to open a dialogue between the study of literature and the study of cognition I think Simon is wholly to be commended. Such a dialogue offers some promise of breaking down arbitrary institutional barriers to the conceptual enrichment of both endeavors. In contrast to Simon's noble aims and the potential significance of his efforts, all that I can offer at the moment are the minor virtues of the gadfly, visiting one or two weak spots in Simon's discussion.
Simon's basic thesis is that the meaning of a word, phrase, or sentence is to be found in the pattern of activations between neurons (rather, "symbol structures" at some aggregate level of analysis) that reading or hearing the expression induces. From the side of the originator, the meaning is that pattern of activations which induced the production of that particular expression. This point of view has some interesting consequences for interpretation and pedagogy that perhaps have been underappreciated. It especially makes clear that the linguistic, historical, and environmental context within which a phrase is received has a direct impact on the meaning of the phrase; furthermore, what follows the phrase is also part of the context, also helping to fix its meaning (disambiguate its interpretation). Although I think research within cognitive science has done much to clarify such a view of meaning-particularly work with semantic networks and neural networks-I suspect Simon overstates his case when he suggests that this vision of meaning is quite new, or that the resultant concept of meaning has "gained a clarity" otherwise unavailable. The history of the discussion of meaning is rich with associationist ideas, beginning at least with John Locke.
Again, there appear to be some substantial unclarities lurking within this account of meaning. For one thing, Simon acknowledges that in addition to whatever activation pattern arises on a specific occasion, we should consider the potential activation patterns that could have arisen, but did not in fact arise. We are told that the entire framework of associations, connections that might be activated by the expression, constitute its "potential meaning," whereas the actual activation pattern is its "actual meaning." Surely there must be some kind of meaning that extends beyond the "actual" meaning, for the actual pattern of associations that may arise on any given occasion is more than idiosyncratic, it is peculiar not just to the individual but to the individual in some unique circumstance. These are not the kinds of meanings that we communicate and share with others, or even with our subsequent selves through our memories. But Simon's "potential meaning" apparently does the job of carrying communicative meaning no better: what might get hooked up to any given expression will turn out to be anything whatsoever, given a sufficiently rich environment (context) to build the associational bridge.
Trying to approach meaning primarily through whatever is actually conjured up on any particular occasion appears to be fundamentally flawed. Intuitively, should we hear a phrase that reminds us of some very distant, tenuously related event, we would not count that remembering as a part of meaning of the phrase, but instead as a coincidental aspect of having heard it. In any case, actual patterns of association are connected with every case of perceiving anything-not just hearing words, but observing airplanes, birds, and clouds-and, whereas these all have meanings in some sense, they do not have meanings in the sense that Simon set out to analyze. Nor would it be enough merely to point out that one set of patterns has arisen in response to words whereas the other arose in response to, say, clouds-for very much more must distinguish these varieties of meaning. For example, the meanings of some words combine to form propositions that may be true or false; associative patterns in response to clouds, wind and rain do not. An account of meaning that fails to distinguish between the meaning of linguistic expressions and the meaning of clouds on the horizon is one that is radically incomplete.
It sounds perfectly fine to say that the meaning of an expression is the pattern of activation with which it is received (or produced). Furthermore, we can make reference to actual programs that implement just such varieties of activation and can do some limited varieties of story understanding or generation. So it seems that we have an operational definition of meaning that ought to make issues about semantics precisely treatable. But I suggest we in fact have nothing more than a vague wave of the hands in the general direction of associative memory, rather than a significant account of meaning. Neither actual nor potential meaning provide a serious explication of the concept, and we are simply left puzzled about how much-and how-such patterns of activation contribute to meanings as we know them.
I do not wish to deny that activation patterns do not come into a proper understanding of meaning, but I do not see that they do so quite as simply as Simon would have us believe. Perhaps a more plausible approach than to tolerantly accept any activation that arises, or may arise, would be to look for those activations that we can expect to arise for normal members of the linguistic community under some range of circumstances.
A different sort of objection is directed at the school of thought from which Herbert Simon is surveying the literary landscape. In my view, it is just a major weakness of Simon's account that he assumes the physical symbol system hypothesis, implying that symbolic representations suffice to capture all of the semantic content that is accessible to us. There are at least two respects in which this account is unsatisfactory: it fails to supply any crucial role for referential meaning in meaning generally; and it underappreciates the potential role of quantitative reasoning in the generation and processing of meaning.
First, it fails to find any difference between expressions that are referential from those that are not. Although Simon acknowledges the distinction himself-in effect, noting that stories of Pegasus are not about something in the same sense that a story of George Washington might be-his theory does not make that same acknowledgement: since the range of phenomena upon which it draws (mental associative links) is circumscribed by the skull, it cannot draw distinctions based upon the success or failure of mental contents to refer to objects beyond that boundary. In short, Simon's discussion suffers from a lack of any account of the role of referential meaning in constructing meanings generally. Yet this is a subject that has achieved some prominence within cognitive science since Hilary Putnam's notable discussion of it in "The Meaning of 'Meaning'" (Putnam, 1975)-which argues strongly that meanings are not all "in the head." This is not to deny that there is precedent within cognitive science for ignoring the external world of referents, particularly after Jerry Fodor popularized such a posture as "methodological solipsism" (Fodor, 1980). But it is one thing to strike that pose methodologically-as a useful fiction while we conduct our computational experiments-and quite another to adopt it when putatively examining the mysteries of meaning, wherein the mysteries of referential meaning plausibly play some part. (Referential meaning appears to play some part in answering Searle's challenges to AI, for example; cf. Harnad, 1989, and Korb, 1991.)
The second insufficiency is Simon's exclusive reliance upon symbol processing, under the natural assumption that this implies the exclusion of quantitative (real-valued) processing such as is found in neural networks (equivalently, connectionist networks), genetic algorithms, Bayesian reasoning and information-theoretic classification. Hardly anyone denies that symbolic processing is important for AI, even connectionists (see, for example, Hinton, 1991 for recent efforts to achieve significant symbol processing using connectionist means). Many symbolicists, on the other hand, do deny that quantitative processing has a role to play in AI, commonly urging that logic programming can address every issue, even including scientific induction. This is, I think, a dubious notion (for particulars, see Korb, 1993). It is not really clear whether this thought is shared by Simon: after all, the kinds of associational structures which he proposes for carrying the semantic content of our concepts are necessarily modeled with associative weights that are real-valued, and so in the relevant sense non-symbolic. Thus, Simon notes that after reading some text "activation gradually dies away"; this dying away is clearly a quantitative process. On the other hand, Simon continually reveals anti-numerical beliefs; in this particular paper, for example, he asserts the unvarnished untruth: "numbers seldom enter at all into the programs that have been written to simulate human thinking." This reveals a continued allegiance to an intolerant ideology in artificial intelligence, one that has no place for quantitative methods.
What Simon concludes here of the divergent schools of literary criticism surely does not lose its meaning-or correctness-when applied reflexively to the divergent schools of cognitive science:
An outsider, viewing the internecine battles that seem to go on constantly within the community of critics and theorists of criticism, wonders why they cannot be settled easily and pacifically. . . .If the claims of uniqueness and exclusive correctness were abandoned, as they surely must be, peaceful coexistence could be wholly restored. Of course that might not be as much fun as the current noisy combat. (24)
Mr. Simon is multiply right.
1. Admittedly, Simon writes that symbols have meanings "especially because some of them can denote. . . .things, relations, and events outside the head." In the immediately preceding sentence, however, Simon asserts that "meanings are held in heads." How denotation can be a non-semantic relation and, if so, how it can be especially important for the origination of meaning are left utterly mysterious.
2. Some might argue that this assumption is unnatural since, after all, numerals are symbols. But the reals cannot be brought into 1-to-1 correspondence with any set of numerals; the real numbers themselves cannot be treated as symbols. What is at issue is more a matter of attitude in any case: it is a sociological fact that those who have most prominently advocated symbol processing as the exclusive means for AI programming have ignored, or worse rejected outright, the quantitative methods I refer to in the text.