Topic Continuity for Discourse and Dialogues

Abstract:

Topic continuity, one of the major linguistic devices for text cohesion, plays important roles in multi-sentential scenario. It integrates multiple sentences into a holistic unit --- e.g., a cohesive paragraph or a fluent dialogue --- distinguishing them from a set of random sentences. In this thesis, we study how topic continuity helps building computational models for written discourse and oral dialogues, two forms of multi-sentential texts.

First, we study how topic continuity helps written discourse understanding through the task of implicit discourse relation recognition in Penn Discourse Treebank. We find it is the prerequisite for discourse relations. Specifically, it licenses individual semantic characteristics to form corresponding discourse relations.

Second, we study the effectiveness of topic continuity in oral dialogue generation. We start with task-oriented dialogue system, where the topics to be continued across utterances are annotated as slots in dialogue research community. We propose a sequence-to-sequence model, called Sequicity, to capture such slots for both knowledge base search and response generation.

Then, we extend Sequicity into semi-supervised explicit dialogue state tracker (SEDST), a more general model. It captures topic continuity across utterances for both task-oriented and non-task-oriented dialogue systems, in both semi-supervised and unsupervised modes. For task-oriented dialogue systems, SEDST works even when slots are not (fully) annotated. For non-task-oriented dialogue systems, the discovered topics can be adopted as explicit signals for dialogue state tracking.