Towards Learning Dialogue Structures from Speech Data and Domain Knowledge: Challenges to Conceptual Clustering using Multiple and Complex Knowledge Source

This paper introduces an engineering-oriented approach towards dialogue modelling. While dialogue models in existing dialogue systems usually are manually coded, or at least the data on which they are based is manually labeled, we investigate the possibility of learning dialogue models from a large set of example dialogues using real data in a spoken dialogue system. We assume the dialogue system to be front end to a knowledge based system (kbs) . In this scenario, different modules like the kbs itself, lexicon, word recogniser, parser etc. provide information that may contribute to the recognition of dialogue acts and to the generation of system's reaction. Information in our scenario may be ambiguous, probalistic or have a complex structure, and our learning algorithm must be able to process it. Implications on the learning algorithm from these demands are discussed. Unlike other approaches that use machine learning for learning classification rules for a set of predefined classes from some discourse theory, we want to learn an application specific classification schema from speech data using conceptual clustering. The reason for this is that all theories have their strengths and weaknesses, and combining all of them seems to be overkill for the limited variability of dialogues in a specific application (the engineering viewpoint). Additionally, we expect some new insight to the importance of different discourse phenomena and perhaps results impose some changes to existing theories. 1 . Introduction In this introduction we motivate our approach, define a learning task and choose an appropriate learning algorithm. Section 2 investigates different knowledge sources in an spoken dialogue system and section 3 points out their implications on the learning algorithm. Finally there are some concluding remarks. Dialogue modelling for natuaral language interfaces has become a topic of substantial interest during the last decade. It combines work on discourse analysis and speech acts with domain tasks and human computer interaction. All of the work on dialogue presumes the existence of a set of units, often called dialogue acts which are building blocks for a dialogue model. Some work tries to define a set of dialoue acts from a theory of action, contributing to models of mutual knowledge [Coh87, Coh90, Faw89]. Bunt [Bun94] distinguishes dialogue acts contributing to different dimensions of context, like linguistic, semantic,