Feature Sharing in the Generation and Interpretation of Nominals in Dialogue

Natural Language Generation in Spoken and Written Dialogue

Volume

Papers from the 2003 AAAI Spring Symposium

Contents

In a fully collaborative, mixed-initiative dialogue it is necessary to both interpret and generate nominal descriptions and so the question of the extent to which these processes can share knowledge is helpful for deciding what to include in the dialogue history and for gaining insight for generation into what enhances interpretation and vice versa. Although theoretical and symbolic models for the interpretation and generation of nominals share much in common, this is not necessarily the case with the statistical models that have been tried. There have been no studies to compare these feature sets and examine whether features that have provided good results for interpretation will do so for generation and vice versa. In this paper, we describe work in progress to do so. So far we have tested generation features for 3 models on the interpretation task and found that they could provide a significant contribution.