Evaluating Answer Extraction for Why-QA using RST-annotated Wikipedia texts

In this paper the research focus is on the task of answer extraction for why-questions. As opposed to techniques for factoid QA, finding answers to whyquestions involves exploiting text structure. Therefore, we approach the answer extraction problem as a discourse analysis task, using Rhetorical Structure Theory (RST) as framework. We evaluated this method using a set of why-questions that have been asked to the online question answering system answers.com with a corpus of answer fragments from Wikipedia, manually annotated with RST structures. The maximum recall that can be obtained by our answer extraction procedure is about 60%. We suggest paragraph retrieval as supplementary and alternative approach to RST-based answer extraction.