Using DIA-MOLE for unsupervised learning of domain specific dialogue acts from spontaneous language

Abstract. This report introduces DIA-MOLE, a tool that supports an engineering-oriented approach towards dialogue modelling for a spoken-language interface. Our approach is applied to the domain of appointment scheduling. A major step towards dialogue models is to know about the basic units that are used to construct a dialogue model. DIA-MOLE does not employ theory-based dialogue units because they are subject to human interpretation and often cannot be recognized from data available in a spoken-language system. We pursue a data-driven approach and apply unsupervised learning to a sample set of spontaneous dialogues using multiple knowledge sources, i.e. domain and task knowledge, word recognition and prosodic information. Using these data, DIA-MOLE supports segmentation of turns and interpretation of their illocutionary force based on a model of the task. For this purpose we had to develop a model of interactive problem solving in the domain of appointment scheduling. As a result of learning we obtain domain- and task-specific dialogue acts (DDA). A first validation of the set of learned DDAs shows that they are prominent for this domain and task. Some DDAs show significant correspondence to specific nodes in an RST-structure. Automatic DDA labeling was compared with human dialogue act labeling according to a predefined labeling scheme. Dialogue act prediction was also employed to evaluate our approach