The Architecture of Cognition: Rethinking Fodor and Pylyshyn's Systematicity Challenge

5. Systematicity in the Lexicon: On Having Your Cake and Eating It Too

1 Introduction Fodor and Pylyshyn’s (1988, henceforth F&P) seminal work marked an important step forward in the debate about what kind of architecture is required for human cognition. By focusing on core concepts of “systematicity ,” “symbolic,” and “compositionality,” F&P suggested new ways to consider the competing claims about connectionist versus symbolic systems in an explicit and precise way. Whether or not one agrees with their definitions (and there has been considerable debate both about the correctness as well as clarity and coherence of these definitions), the narrowing of the debate has led to stimulating discussion. Absent such discussion, for example, terms such as ‘systematic’ may be taken to refer to significantly different things. The equation y(t) = f(x, yt–1 ), for example, describes a dynamical system that is, in general terms, systematic . The problem is that such equations can be used to characterize both connectionist and (with some stretching) symbolic systems. F&P usefully narrow the definition of systematic in a way that allows us to ask important questions: Are connectionist models capable of systematicity? Are only symbolic models systematic? Is human cognition itself systematic ? The main point of F&P, of course, is that a cognitive architecture must be capable of systematicity in order to explain human cognition. A related formulation, which captures much of what F&P argued for, is that human cognition is algebraic, where that algebra requires symbols that are abstract and context free (see Marcus 2001). Although F&P’s definitions characterize both representations as well as rules, much of the ensuing debate has tended to focus on their claims regarding systematicity of rules. Representations, and in particular, lexical representations, have not figured prominently in the debate. This emphasis on rules is not surprising, given the zeitgeist of the time in which the 5 Systematicity in the Lexicon: On Having Your Cake and Eating It Too Jeffrey L. Elman 116 Jeffrey L. Elman paper was written. In that period, rules were seen as the major source of linguistic generativity. Because the lexicon was understood as involving a relatively stable inventory of entities with fixed meanings, pronunciations, and so on, the contribution of the lexicon to productive language use was felt to not be particularly interesting. In intervening years, however, the lexicon has come into its own. Lexical knowledge is now acknowledged to be quite rich. Given proposals that the lexicon includes abstract patterns (e.g., Goldberg 2003), it is now seen as a source of linguistic productivity. One consequence of this is that boundary between the lexicon and rule systems has become somewhat blurred. In this chapter, I consider ways in which systematicity, as F&P have defined it, might apply to the lexicon. Their treatment of the lexicon is light, which requires me to make some assumptions about how systematicity might apply to the lexicon. Whether or not these assumptions are reasonable is for the reader (and, presumably, F&P) to judge. A number of chapters in this book challenge the claims made by F&P with regard to rule systems. I will argue that similar problems arise in considering the lexicon. So I begin by discussing empirical data that are problematic for the view of the lexicon as an enumerative database of the sort that would be required by an F&P approach. That is, the option of significantly enlarging the format of lexical entries to accommodate these data—the obvious move—not only leads to an unwieldy combinatoric explosion of information, but more seriously compromises the theoretical assumptions that motivated placing some, but not all, information in the lexicon. I then go on to argue that a very different mechanism than the lexicon is required in order to capture the richness of lexical knowledge. To lend some concreteness to this conclusion, I conclude with a preliminary model that I believe possesses some of the characteristics that are desirable in such a mechanism. This model eliminates the lexicon, qua lexicon, while providing an alternative way for lexical knowledge to play a role in language comprehension. 2 The Lexicon as Dictionary The metaphor of the mental lexicon as a dictionary is pervasive and compelling . However, metaphors bring a lot of baggage with them, sometimes hidden from view. If the lexicon is to do real work for us, we need to go beyond metaphor and specify what the properties of the lexicon are. Systematicity in the Lexicon 117 There is considerable and important divergence on this issue, although commonalities exist across many theories. These are well captured in Jackendoff ’s (2002) description...