Applications of Discourse Structure for Spoken Dialogue Systems

摘要

Language exhibits structure beyond the sentence level (e.g. the syntactic structure of a sentence). In particular, dialogues, either human-human or human-computer, have an inherent structure called the discourse structure. Models of discourse structure attempt to explain why a sequence of random utterances combines to form a dialogue or no dialogue at all. Due to the relatively simple structure of the dialogues that occur in the information-access domains of typical spoken dialogue systems (e.g. travel planning), discourse structure has often seen limited application in such systems. In this research, we investigate the utility of discourse structure for spoken dialogue systems in more complex domains, e.g. tutoring. This work was driven by two intuitions.First, we believed that the "position in the dialogue" is a critical information source for two tasks: performance analysis and characterization of dialogue phenomena. We define this concept using transitions in the discourse structure. For performance analysis, these transitions are used to create a number of novel factors which we show to be predictive of system performance. One of these factors informs a promising modification of our system which is implemented and compared with the original version of the system through a user study. Results show that the modification leads to objective improvements. For characterization of dialogue phenomena, we find statistical dependencies between discourse structure transitions and two dialogue phenomena which allow us to speculate where and why these dialogue phenomena occur and to better understand system behavior.Second, we believed that users will benefit from direct access to discourse structure information. We enable this through a graphical representation of discourse structure called the Navigation Map. We demonstrate the subjective and objective utility of the Navigation Map through two user studies.Overall, our work demonstrates that discourse structure is an important information source for designers of spoken dialogue systems.