Dialogue Structure and Pronoun Resolution

This paper presents an empirical evaluation of a pronoun resolution algorithm augmented with discourse segmentation information. Past work has shown that segmenting discourse can aid in pronoun resolution by making potentially erroneous candidates inaccessible to a pronoun’s search. However, implementing this in practice has been difficult given the complexities associated with deciding on a useful scheme and then generating the segmentation reliably. In our study, we investigate whether or not shallow schemes that are easy to generate will improve pronoun resolution in a task-oriented corpus of dialogues. Our results show that incorporating this shallow segmentation at best marginally improves pronoun resolution performance.