Multimodal user state recognition in a modern dialogue system

Abstract

A new direction in improving automatic dialogue systems is to make a human-machine dialogue more similar to a human-human dialogue. A modern system should be able to recognize the semantic content of spoken utterances but also to interpret some paralinguistic or non-verbal information - as indicators of the internal user state - in order to detect success or trouble in communication. A common problem in a human-machine dialogue, where information about a users internal state of mind may give a clue, is, for instance, the recurrent misunderstanding of the user by the system. This can be prevented if we detect the anger in the users voice. In contrast to anger, a joyful face combined with a pleased voice may indicate a satisfied user, who wants to go on with the current dialogue behavior, while a hesitant searching gesture of the user reveals his unsureness. This paper explores the possibility of recognizing a user's internal state by using facial expression classification with eigenfaces and a prosodic classifier based on artificial neural networks combined with a discrete Hidden Markov Model (HMM) for gesture analysis in parallel. Our experiments show that all the three input modalities can be used to identify a users internal state. However, a user state is not always indicated by all three modalities at the same time; thus a fusion of the different modalities seems to be necessary. Different ways of modality fusion are discussed.