Of Minds and Language (53 page)

The perceptual strategies differ from language to language: we found that by age 4, children acquire processing strategies adaptive to the statistical regularities in the structure of their own language (Slobin and Bever 1982). Thus, in English what develops is sensitivity to word order, in Turkish, sensitivity to patient/object inflectional markers, in Italian and Serbo-Croatian, sensitivity to a mixture of the two kinds of linguistic signals. This reflects the fact that each language has its own CFC, which children learn.

The following points are in large part the result of email discussions with Noam.

The idea that the child acquires knowledge of syntax by way of compiling statistical generalizations and then analyzing them with its available syntactic capacities is but another instance of learning by hypothesis-testing. For example, it is technically an expansion on the TOTE model proposed by Miller et al. (1960). An initial condition (statistically grounded pattern) triggers a TEST meaning, and an OPERATION (derivation) which triggers a new TEST meaning and then EXIT. Karmilov-Smith and Inhelder (1973) advanced a different version – cognition advances in spurts, triggered by exposure to critical instances which violate an otherwise supported generalization.

The dual nature of the acquisition process is also related to classical theories of problem solving (e.g., Wertheimer 1925, 1945). On such models, the initial stage of problem organization involves noting a conceptual conflict – for example, “find a solution that includes both X and Y: if the answer is X then Y is impossible, but if Y then X is impossible”: characteristically the solution involves accessing a different form of representation which expresses the relation between X and Y in more abstract terms. In language the initial conflict expresses itself as the superficial identity of all the constructions in (12) which exhibit the canonical form constraint, while assigning different semantic relations; the resolution is to find a derivational structure for the set that shows how
the different surface constructions are both differentiated and related derivationally. Hence, not only is language-learning hereby interpreted in the context of a general set of learning principles, it is also interpreted as a special instance of a general problem solver. This also explains why language learning is fun, and hence intrinsically motivating: the gestalt-based model suggests that language-learning children can enjoy the “aha” insight experience, an intrinsically enjoyable sensation which may provide critical motivation to learn the derivational intricacies of language (cf. Weir's 1962 demonstration that children play with their language paradigms when they are alone).

Note that the terms “motivation” and “fun” are technical terms based in aesthetic theory, not the everyday notion of conscious desire, nor any notion of “reinforcement.” Elsewhere, I have developed analyses of what makes objects and activities intrinsically enjoyable (Bever 1987). The analysis draws on the classic aesthetic definition: stimulation of a representational conflict which is then resolved by accessing a different form or level of knowledge. The formal similarity of this definition to the gestalt model of learning affords an explanation of why aesthetic objects are enjoyable: they are mini-“problems” involving conflicting representational solutions, resolved by accessing a level which creates a productive relation between those solutions, thereby eliciting a subconscious “aha.” This kind of analysis is ordinarily applied to serial arts such as drama or music, in which the representational conflict and its resolution can be made explicit over time. But the analysis works for static objects, explaining the preference, for example, for the golden mean rectangle. In language, one kind of conflict is elicited by the thematic heterogeneity of superficially identical surface phrase structures: the child's resolution of that conflict requires access to an inner form of the sentences, via distinct derivational histories – a resolution which involves accessing a distinct level of representation. Thus, learning the structure of a language elicits a series of mini-ahas in the child, making it an activity which is intrinsically attractive.

The model also offers a partial answer to the frame problem (see Ford and Hayes 1993), the problem of how statistical generalizations are chosen out of the multiple possibilities afforded by any particular set of experiences. This problem was classically addressed by Peirce (1957) as the problem of abduction, who argued that there must be constraints on all kinds of hypotheses, even those ostensibly based on compilation of observations (cf. Chomsky 1959c, on the corresponding problem in S-R associative theory, and this volume). But the problem is also a moving target for the language-learning child. At any given age, the generalizations that are relevant to progress in learning are different: if the child has mastered simple declarative constructions, or some subpart of her language's inflectional system, this changes the import of further exposure
to the language. Thus, we must not only address constraints on the initial state of the child (see Mehler and Bever 1968 for discussion), we must address how constraints apply to each current state of knowledge, as the child matures and acquires more structural knowledge. That is, the abductive constraints themselves have a developmental course. By what process and dynamic? Another way of putting this is, what filters (aka “frames”) possible generalizations and how does the filter itself change as a function of current knowledge?

In the A×S scheme, there are two kinds of processes which filter generalizations. First is the set of salient regularities among elements that are available to the input: at a phonological level, infants have available perceptual categories that provide an initial organization of the input; this affords an innate categorization of sound sequences, available for formal derivational analysis. The other side of the filtering process is the set of computational devices available to provide a derivation. That is, those generalizations about sound sequences that endure are just those that can be explained by a set of possible computational phonological rules. Such rules must have natural domains (presumably innately determined) such as segmental features, syllabic structures, lexical templates. At the syntactic level, the corresponding problem is to isolate a natural segmentation of the potential compositional input. To put it in terms of the example we are focusing on, how does the system isolate “NP V (NP)” as a relevant kind of sequence over which to form a generalization? In the model proposed, the solution lies in the fact that the derivational component has its own natural units, namely clause-level computations. The result is that the derivational discovery component acts as a filter on the multiple possible statistical generalizations supported by any finite data set, picking out those that fit the derivational templates. Most important is that the properties of the derivational filter change as the knowledge base increases in refinement.

Recent discussions, and this conference, have clarified current linguistics as “biolinguistics,” the isolation and study of genetic endowment and boundary conditions on the faculty of language. The formal approaches to isolating and explaining universals via abstracted biological constraints on what language is, or by examining the data required to set parameters in an ideal learner, clarify the relevant abstract conditions on individuals learning language. Yet it is a collection of concrete individuals that learn and use language. Thus, these boundary conditions may profit from inclusion of the motivations and actions
of individual learners. I have given various kinds of examples of linguistic universals, showing how we can benefit by examining the dynamics of language learning in individuals. The extent to which individuals learn language by way of mechanisms not specific to language alone clarifies what we should take as the essential universals of language. The discussion in this paper of EPP is an example of this kind of argument.

Marc D. Hauser

The topic that I want to talk about today falls under the title “The illusion of biological variation.” Let's consider a canonical perceptual illusion, one in which the image is completely static, with nothing moving at all, except that your visual system thinks it is. Now, no matter how many times you tell the subject that the image is static, his or her visual system won't believe it; it can't. Illusions are interesting because, no matter how aware we are of them, they simply won't go away. Similarly, and by way of analogy, I will suggest today that much of the variation that we see in the natural world is in some sense an illusion because at a different level of granularity, there are some core invariant mechanisms driving the variation.

As in any talk that attempts to go beyond one's typical intellectual limits or comfort zone, I must first make a few apologies. The first one is to Chris Cherniak and other theoretical biologists, for my gross generalizations drawn from some of the very deep facts they have uncovered about the natural world. The second one is to Noam and other linguists because I am going to generate some wild speculations about language evolution from a very fragmentary bit of evidence. The third apology is to a class of philosophers, and in particular to John Rawls, for cutting out all the subtleties of argumentation that have gone on about utilitarianism on the one hand, and deontological principles on the other, so that I can cut to the chase and tell you about how the moral faculty works. And then a final apology to John Cage and many minimalists in music
and art, particularly for taking some grotesque liberties with their theories and painting a slightly different picture of what I think they really were after.

The first point to make is that when we look upon the natural world, we immediately see extraordinary variation in animal forms, what looks like limitless variation, not just in size (from extremely small animals to immensely huge animals), but in shapes, material properties, and so forth. Similarly, we see apparently limitless variation in the patterns of animal locomotion, including, most noticeably, those observed in the air, on land, and in the sea. Somebody raised a question earlier about the immune system
¹
– again, a system with limitless variation in the kinds of responses that it generates to different kinds of problems in the environment. I want to call all of this observed variation, the “illusion of biological variation.” It is an illusion, at least in part, because when biologists have looked deeply into the sources of variation in these different domains, as Cherniak's talk in this conference illuminated (see
Chapter 8
), we find something different – a common set of core mechanisms that generates the variation.

Let me put this into a historical context by quoting from two biologists who confronted the nature of biological variation. The first is Sewell Wright, who may be known to many of you. He was a distinguished evolutionary biologist who, following on from Darwin, talked about the nature of adaptation and in particular the notion of an adaptive landscape. Here is what Wright pointed out in the 1930s, which I think is very telling in terms of the story I want to paint today (Wright 1932). He says that the older writers on evolution were often staggered by the seeming necessity of accounting for the evolution of fine details. He then adds that structure is never inherited as such, but merely types of structure under particular conditions. Now, at the time Wright was discussing these matters, there were major revolutions afoot in genetics and molecular biology. If we fast-forward the story to today, here is an almost verbatim quote from Mark Kirschner, a systems biologist who makes very much the same point but takes it a little bit further and takes it in a direction that will hopefully have great appeal especially to the linguists in the audience who are interested in certain kinds of structural properties. In essence, Kirschner (Kirschner and Gerhart 2005) says that novelty in the organism's physiology, anatomy, or behavior arises mostly by the use of conserved processes in new combinations at different times, in different places and amounts, rather than by the invention of new processes. This is very much in line with some of the things that Gabby Dover says in his contribution here (see
Chapter 6
). Kirschner stresses that, in the 4–5 billion years of cellular life on Earth, there have been four core processes
leading to variation: rearrangement, repetition, magnification, and division. For those of you who have been tracking what has been happening in the minimalist program, you will see a kind of family resemblance to these four core processes.

The idea that I want to push today – a project that Noam and I have been working on a bit over the last year or so, but that I will take full responsibility for in terms of errors – is to invoke three principles that extend the minimalist program in linguistics to the mind and other domains of knowledge more generally. The first is that any time we observe an open-ended, limitlessly expressive, powerful system, it will be based on a fixed set of principles or mechanisms for generating the observed variation – that is, some kind of generative, combinatorial system. Secondly, these generative mechanisms must in some sense interface with system-internal and -external processes, with nature potentially finding the optimal solution given the current conditions. This allows for lots of accidental variation that happened before, but I will talk specifically about how it allows solutions to the current conditions. And then lastly, each variant we observe will be determined by some kind of process of pruning, where the local experience tunes up the biologically given options.