Multi-feature error detection in spoken dialogue systems

The present paper evaluates the role selected features and feature combinations play for error detection in spoken dialogue systems. We investigate the relevance of various, readily available features extracted from a corpus of dialogues with a train timetable information system, using RIPPER, a rule-inducing machine learning algorithm. The learning task consists of the identification of communication problems arising in either the previous turn or the current turn of the dialogue. Previous experiments with our corpus have shown that combining dialogue history and word-graph features is beneficial for detecting errors (in particular in the previous turn). Other researchers have reported that combining prosodic and ASR characteristics is helpful (primarily in the current turn). In this paper, we investigate the usefulness of large-scale combinations of these features for the above two tasks. We show that we are unable to reproduce the benefits of prosodic features for learning problematic situations, even though the overall prosodic trends in our corpus are similar to those earlier reported on. Moreover, the best results are obtained using just minimal combinations of two sources of information