Can distributional approaches improve on Good Old-Fashioned Lexical Semantics?

Can distributional approaches improve on Good Old-Fashioned Lexical Semantics? Ann Copestake Computer Laboratory, University of Cambridge aac10@cl.cam.ac.uk Abstract In this position paper, I discuss some linguistic problems that computational work on lexical semantics has attempted to address in the past and the implications for alternative models which in- corporate distributional information. I concentrate in particular on phenomena involving count/mass distinctions, where older approaches attempted to use lexical semantics in their models of syntax. I outline methods by which the earlier models allowed the transmission of information between lexi- cal items (regular polysemy and inheritance) and address the possibility that similar techniques could usefully be incorporated into distributional models. 1 Introduction While there has been much recent discussion of techniques for developing compositional approaches to distributional semantics, especially with respect to particular categories of phrase (e.g., adjective-noun), as far as I am aware, there has been no attempt to discuss systematically all the roles that distributional se- mantic representations might play in the production of a model of a sentence. Indeed, from the viewpoint of researchers working on ‘traditional’ areas of computational linguistics, such as parsing and generation, and those primarily interested in modeling language for its own sake, rather than application-building, the extensive work on distributional semantics has been somewhat disappointing in failing to provide models which are integrated with existing work to help solve long-standing problems. In some respects, most work on distributional semantics lacks ambition compared to earlier research on lexical semantics, in that previous approaches at least attempted to provide accounts that were fully integrated with syntax and full-coverage compositional semantics: i.e., which used lexical semantics as part of the models that assigned syntactic structure or logical form.1 There are reasons to think that distributional approaches could well be more appropriate in such contexts, but a demonstration of this will involve looking at a broad range of phenomena. This paper is intended as a first step in outlining some of the issues that might be considered. I first want to distinguish the discussion of lexical meaning here from the various approaches to deriving distributional meaning from sentences investigated by Clark and Pulman (2007), Baroni and Zamparelli (2010), Mitchell and Lapata (2010), Guevara (2011) and others which in turn relates to pre- vious approaches to combining connectionist and symbolic approaches (e.g., Smolensky and Legendre, 2006). That line of work assumes that a syntactic representation (or perhaps a logical form) is available to guide the process of composition of distributions.2 This work is mostly orthogonal to the issue I wish 1Note that use ‘compositional semantics’ in its predominant sense to mean an approach in the tradition of Montague gram- mar, construed broadly, but including a treatment of quantification. 2I note the possibility of working with logical forms, since, although it is usual to work with syntactic relationships when working on compositional distributional semantics, the assumption is that these relationships are semantically meaningful. It thus seems possible that the models would, in principle, perform better if they were built on …

Content truncated to 500 words