Learning a Radically Lexical Grammar

Learning a Radically Lexical Grammar Danny Solomon Mary McGee Wood Department of Computer Science University of Manchester Manchester M13 9PL, UK { danny, mary } @cs. man. ac. uk Abstract We describe a prototype system which induces a categorial grammar from a simple text corpus of children's reading books. Unlike previous attempts at grammar induction, (I) there are no rules of grammar, only a richly structured lexicon; (2) we rely both on an informing linguistic theory and on statistical methods applied to a corpus. 1 Introduction - Learning a 'Grammar' The work we describe was originally motivated by dissatisfaction with attempts to induce a rule-based grammar from a corpus (eg Berwick 1985), and the suspicion that a word- based grammar - at the logical extreme, a categorial grammar (Wood 1993) - might be easier to 'learn'. With the recent rapid growth of 'empiricism' in NLP, the same work now manifests 122 our dissatisfaction with the view that throwing a large computer at a large corpus to produce a large set of sets of numbers is in itself 'linguistics': we are committed to the honest description of 'real 'data, but only within an independently motivated theoretical framework. Neither of these are widely held views at present. For the first, even Brent's (1993) 'learning of lexical syntax' takes a distinct component of grammar rules as its starting point in learning a lexicon. For the second, as an example, the tension which pervaded the EACL meeting in Utrecht in April 1993, stoked by invited talks from Church (1993) and Sag (1993), made clear the deep division - even antagonism - between statisticians and theorists on the front line. However the success so far of our implementation gives some evidence for the viability of a totally lexical approach to the automatic learning of grammar, responsive to both principles and facts. The next section describes further some of the theoretical and practical motivations behind our work. Section 3 briefly sets out the exact categorial grammar used. In Section 4 we describe L - a prototype grammar induction system which illustrates the direction we believe will be productive. A discussion and a look towards some prospects complete the paper. 2 Motivation 2.1 Theo~tical It is hard to imagine an automatic (or even semi-automatic) grammar or lexicon induction procedure that could manipulate traditional grammatical entities such that human-meaningful results might obtain. Brute-force methods (ie those that exploit the massive raw computing power currently available cheaply) may well produce some useful results (eg Brown et al 1993). However, ff any linguistic insight is to appear as a result, we believe an underpinning in linguistic theory is essential - overall success will result from the combination and integration of that at which computers and human linguists excel. We believe this is only possible if both partners work in the same basic framework - and we also believe that training linguists to read disk sectors will probably be unproductive. Our premise is that: 1) giving grammatical entities structure is useful …

Content truncated to 500 words