In S. Garrod & M. Pickering (Eds.), Language processing (pp. 233-279). Hove, U.K.: Psychology Press.

Connectionism and Natural Language Processing



Nick Chater & Morten H. Christiansen


Introduction

Many of the chapters of this book are concerned with topics in language processing. This chapter is concerned, by contrast, with a particular method, connectionist computational modelling, which has been applied to a wide range of topics. It is, furthermore, a controversial method: some have argued that natural language processing from phonology to semantics can be understood in connectionist terms; others have argued that no aspects of natural language can be captured by connectionist methods. And the controversy is particularly heated because of the revisionist claims of some connectionists: for many, connectionism is not just an additional method for studying language processing, but it offers an alternative to traditional theories, which describe language and language processing in symbolic terms. Indeed, Rumelhart & McClelland (1987: p. 196) suggest ``that implicit knowledge of language may be stored among simple processing units organized into networks. While the behaviour of such networks may be describable (at least approximately) as conforming to some system of rules, we suggest that an account of the fine structure of the phenomena of language and language acquisition can best be formulated in models that make reference to the characteristics of the underlying networks.'' We shall see that the degree to which connectionism supplants, rather than complements, existing approaches to language is itself a matter of debate. Finally, the controversy over connectionist approaches to language is an important test case for the validity of connectionist methods in other areas of psychology.

In the next section, Background we describe the historical and intellectual roots of connectionism, then introduce the elements of modern connectionism, how it has been applied to natural language processing, and outline some of the theoretical claims that have been made for and against it. We then consider four central topics in connectionist research on language processing: Word naming and visual word recognition, Lexical processing during speech, Morphological processing and Syntax. These illustrate the range of connectionist research on language, give an opportunity to assess its strengths and weaknesses across this range, and allows the general debate concerning the validity of connectionist methods to be illustrated in specific contexts. We would argue that debates in each of these areas, while interrelated, should each be considered on their own merits: it may be that connectionist approaches are valuable in modelling some aspects of language processing, but not in others. Finally, in the Conclusions we sum up and consider the prospects for future connectionist research, and its relation to other approaches to understanding language processing and language structure.




Click to request a copy of this paper.

rule.gif (155 bytes)

Home | People | Research | Links | Contact | Publications | Presentations
Cognitive Neuroscience Laboratory

Please email suggestions/errors to mhc27@cornell.edu