Representing Discourse Information for Spoken Dialogue Generation

Prosody and intonation convey important distinctions of “Information Structure”, marking portions of the utterance as standing in relations to the surrounding discourse such as “theme” and “rheme”, and marking relations of contrast between referring expressions and potential reference sets. The use of default intonation contours in standard “text-to-speech” applications can be quite successful, especially when the default pitch-accent assignments are moderated by “previous mention” information as an approximation to informational “given-ness” However, for some applications, particularly those involving dialogue rather than spoken monologue from written text, as well as ones involving systematic comparison among and coordination of alternatives, generation from richer meaningrepresentations can offer a more reliable alternative. Possible applications are inventory query, translating telephony for theme-marking languages like Japanese, personalised patient medical education, and animation of autonomous conversational agents.