September 30 2008, 14.00 hrs

Suzan Verberne, University of Nijmegen

Title:

Using Structural Information for Improving Why-Question Answering

Abstract:

My PhD research project "In Search of the Why" aims at developing a system for answering why-questions. Today I will present my recent work on extending a simple passage retrieval approach with structural information. The starting point is Lemur's TFIDF, which retrieves a relevant answer in the top 150 for 79% of the test questions. However, only 45% of the questions is answered in the top 10. We aim to improve the ranking by adding a reranking module. For re-ranking we consider a set of 31 features representing structural information of the question and answer candidate: syntactic structure as well as document structure. We find a significant improvement over the baseline for both MRR and Success@10, which is now 55%. The most important features for re-ranking are TFIDF (the baseline score), the presence of cue words, the question's main verb, and the relation between question focus and document title.

	Home > ... > Past colloquia > September 30 2008, 14.00 hrs
<< EWI Home Staff Research Teaching Publications Intranet Colloquia DB Vacancies News Search	September 30 2008, 14.00 hrs Suzan Verberne, University of Nijmegen Title: Using Structural Information for Improving Why-Question Answering Abstract: My PhD research project "In Search of the Why" aims at developing a system for answering why-questions. Today I will present my recent work on extending a simple passage retrieval approach with structural information. The starting point is Lemur's TFIDF, which retrieves a relevant answer in the top 150 for 79% of the test questions. However, only 45% of the questions is answered in the top 10. We aim to improve the ranking by adding a reranking module. For re-ranking we consider a set of 31 features representing structural information of the question and answer candidate: syntactic structure as well as document structure. We find a significant improvement over the baseline for both MRR and Success@10, which is now 55%. The most important features for re-ranking are TFIDF (the baseline score), the presence of cue words, the question's main verb, and the relation between question focus and document title.