Some experiments in speech act prediction

Abstract

In this paper, we present a statistical approach for speech act prediction in the dialogue component of the speech-to-speech translation system Verbmobil. The prediction algorithm is based on work known from language modelling and uses N-gram information computed from a training corpus. We demonstrate the performance of this method with 10 experiments. These experiments vary in two dimensions, namely whether the N-gram information is updated while processing, and whether deviations from the standard dialogue structure are processed. Six of the experiments use complete dialogues, while four process only the speech acts of one dialogue partner. It is shown that the predictions are best when using the update feature and deviations are not processed. Even the processing of incomplete dialogues then yields acceptable results. Another experiment shows that a training corpus size of about 40 dialogues is sufficient for the prediction task, and that the structure of the dialogues of the Verbmobil corpus we use differs remarkably with respect to the predictions.