Cluster-based aprroacheies to the statistical modelling of dialogue data in the British National Corpus

Spoken language dialogue systems (SLDSs) are becoming more and more widespread as interfaces between human users and automated systems. The speech recogniser is a key component of an SLDS, and a statistical language model is at the core of most successful recognisers. In this paper, we describe and compare several cluster-based strategies for modelling dialogue turns, applying them to pairs of successive short dialogue turns from the British National Corpus - the type of turn pairs likely to be common in the kind of highly interactive dialogues in which an SLDS would be expected to participate