Classifying Discourse Relations

Classifying Discourse Relations Mridhula Raghupathy & Hena Mehta [email protected] | [email protected] Faculty Advisors: Dr. Aravind Joshi, Dr. Ani Nenkova, & Dr. Alan Lee Abstract The goal of this project was to study properties of discourse relations as they appear in the Penn Discourse Tree Bank (PDTB), a large corpus of naturally occurring text whose discourse relations and their features have been annotated. We began with an examination of the PDTB in order to arrive at a systematic description of the structure and patterns of discourse relations in the text. We then moved on to the task of classifying these discourse relations into their semantic senses based on the features we had discovered in the PDTB. This led us to the exploration of two specific areas – the question of connective ambiguity, trying to tell whether explicit markers of discourse relations can unambiguously indicate particular semantic senses, and the differences between the classification of explicit relations marked by these visible connectives and implicit relations where relations are derived by other non-obvious cues. While it appeared that the issue of connective ambiguity was not very severe, the differences between explicit and implicit relations were rather drastic, making the classification task easier for explicit relations on the one hand, but much more difficult for implicit relations.