Online adaptation of language models in spoken dialogue systems
Bernd Souvignier, A. Kellner · 1998
The robust estimation of language models for new applications of spoken dialogue systems often suffers from a shortcoming of training material. An alternative to training a language model is to improve an initial language model using material obtained while running the new system, thus adapting it to the new task. In this paper we investigate different methods for onlineadaptation of language models. Apart from the standard techniques of supervised and unsupervised adaptation, we look at two refined approaches: the first allows multiple hypotheses from N-best lists as adaptation material and the second uses confidence measures to exclude unreliably recognized sentences from adaptation. We apply adaptation to both the language model used by the speech recognizer to focus the beam search and to the stochastic language understanding grammar. It turns out that the understanding grammar can be improved quite significantly using N-best lists or confidence measures, whereas unsupervised adaptation may even result in a deterioration of the system. The language model used by the speech recognizer is improved very satisfactorily by each of the chosen approaches. 1.