Modelling Incremental Self-Repair Processing in Dialogue.

Abstract

Self-repairs, where speakers repeat themselves, reformulate or restart what they are saying, are pervasive in human dialogue. These phenomena provide a window into real-time human language processing. For explanatory adequacy, a model of dialogue must include mechanisms that account for them. Artificial dialogue agents also need this capability for more natural interaction with human users. This thesis investigates the structure of self-repair and its function in the incremental construction of meaning in interaction. A corpus study shows how the range of self-repairs seen in dialogue cannot be accounted for by looking at surface form alone. More particularly it analyses a string-alignment approach and shows how it is insufficient, provides requirements for a suitable model of incremental context and an ontology of self-repair function. An information-theoretic model is developed which addresses these issues along with a system that automatically detects self-repairs and edit terms on transcripts incrementally with minimal latency, achieving state-of-the-art results. Additionally it is shown to have practical use in the psychiatric domain. The thesis goes on to present a dialogue model to interpret and generate repaired utterances incrementally. When processing repaired rather than fluent utterances, it achieves the same degree of incremental interpretation and incremental representation. Practical implementation methods are presented for an existing dialogue system. Finally, a more pragmatically oriented approach is presented to model self-repairs in a psycholinguistically plausible way. This is achieved through extending the dialogue model to include a probabilistic semantic framework to perform incremental inference in a reference resolution domain. The thesis concludes that at least as fine-grained a model of context as word-by-word is required for realistic models of self-repair, and context must include linguistic action sequences and information update effects. The way dialogue participants process self-repairs to make inferences in real time, rather than filter out their disfluency effects, has been modelled formally and in practical systems.