Bill Keller - Research Interests

My research interests include:

Some selected papers are also listed below.

 

Language-Based Interfaces in  Ubiquitous Computing Environments

In project NatHab In joint work with David Weir and Ian Wakeman I am investigatinglanguage-based interfaces to ubiquitous computing environments. The aim is to allow non-technical users in the home and office to personalise their environment using natural language "policies" for service configuration (see Weeds et al 2004)

Grammatical Inference/Language Modelling

I am interested in the application of machine learning techniques to problems in language learning/grammatical inference. Practical techniques for grammatical inference have great application to current work in NLP and speech recognition, as well as wider application to syntactic pattern recognition.

In joint work with Rudi Lutz I have explored the use of genetic algorithms to infer stochastic context-free grammars from language data. The problem addressed in this work is that of inferring a suitable grammar for a target language given a stochastic sample of the sentences in the target (i.e. no `negative information' or `teacher'). We adopt a Bayesian approach using the genetic algorithm to `evolve' the most probable grammar given a corpus. To avoid over-fitting of the data the Minimum Description Length Principle is used to provide an informative prior for candidate grammars. (See: Keller & Lutz, 1996, Keller & Lutz, 1997, Keller and Lutz 2004). We have also explored the use of priors in conjunction with the EM algorithm in the context of learning Hidden Markov Models (Keller & Lutz, 2002).

I was the organiser of a workshop on Automated Acquistion of Syntax and Parsing at the 10th European Summer School in Logic, Language and Information (ESSLLI-98).

Formalisms for Linguistic KR

A good deal of my early research was concerned with the formal foundations and computational properties of formalisms in computational linguistics (see e.g. Keller 1992a)

My D.Phil work dealt with unification-based grammar formalisms and the properties of logical languages (so-called feature logics) for expressing constraints on linguistic objects. I investigated a generalization of one of these logics, Rounds-Kasper logic, which incorporates the device of functional uncertainty due to Kaplan and Zaenen (see Keller, 1992b, Keller, 1993)

Unification-based grammar formalisms tend to be very powerful mathematically (as powerful as general-purpose programming languages). Joint work with David Weir led to the development of a restricted unification-based formalism that is more powerful than Linear Indexed Grammar (LIG), but which can also be processed in polynomial time using techniques that are similar to those developed for LIG in Vijay-Shanker and Weir, (1993b). The formalism, (which we refer to as partially linear PATR) manipulates feature structures rather than stacks (see Keller and Weir, 1995)

I have also investigated the lexical knowledge representation language DATR, which was originally developed by Gerald Gazdar and Roger Evans  at the University of Sussex, and is probably the most widely used language for specifying natural language lexicons in the NLP community. I provided the first declarative semantics for the full DATR language in Keller, 1995, and presented an alternative, operational semantics, which axiomatises the evaluation of DATR expressions in Keller, 1996.

 

Selected Publications

Back to Contents