Utterance Content Planner for Generating Conversational Query Expression : Deciding Utterance Content According to Characteristics of Query Content

Abstract

Daily conversation contains a variety of query expressions that can not be classified into general query expressions. For example, the request form may be used as a query, or additional information may be attached to a query expression. This paper proposes an utterance content planner (UCP) that decides the appropriate utterance content for generating a conversational query expression. As input for the UCP, we define five parameters that characterize the topic - which type of information is being requested, how unpredictable the query is, and how special is the knowledge needed to answer the query, etc. The UCP decides the utterance content according to the values of the parameters. The UCP outputs combinations of two kinds of utterance content units (UC unit), which are Kernel UC unit such as Yes/No interrogative or WH interrogative and Auxiliary UC unit which supplies additional information like "Explanation about topic" or "Situation explanation". We define six Kernel UC units and six Auxiliary UC units. In order to select the appropriate UC units, some heuristic rules and statistical models are used. The selected UC units are combined using combination rules and combination constraints. An evaluation of the UCP showed that it was able to predict real conversational query expressions quite well.