Ann Copestake
(University of Cambridge)
Computational Semantics of Natural Language:
what (if anything) does it have to do with logic?

Computational approaches to the semantics of natural language can be seen as having their roots in two different traditions. One is a logical approach, where natural language utterances are seen as somehow corresponding to a formal language. This work predominantly follows the ideas of Richard Montague. Montague thought there was "no important theoretical difference between natural languages and the artificial languages of logicians". The second tradition is related to the 'ordinary language' philosophy of Wittgenstein's later work (although most researchers are probably unaware of the link). The essential idea behind the computational models (e.g., vector space models of document meaning in IR and search, distributional semantics models of word and phrase meaning) is that language is modelled using language. For instance, in a distributional model, the meaning of a word is modelled by the contexts in which it appears in some corpus of natural language. Both in terms of practical results and volume of current research, this tradition has been far more successful than the logic-based approaches, even though the syntactic properties of natural languages are generally not modelled.

This talk will start with a broad (but personal) account of research in computational natural language semantics from these perspectives. I'll discuss why the strictly Montogovian approach has proved of only limited utility, and explain how we can, nevertheless, take ideas from it which are more broadly applicable. These allow us to describe meaning arising from the inflectional morphology and syntax of natural languages using meta-level models which correspond to sets of expressions in logical object languages. For most practical applications, however, we work directly with the meta-level model. I will also mention some recent research which suggests ways in which we can combine such models with distributional techniques.