Relation based models for passage retrieval in open domain question answering

Abstract

Most current passage retrieval systems for open domain question answering use statistical co-occurrence or lexical distance to model the relations between query terms. However, we know that such statistical measure provide only an approximation to the a??reala?? relations between terms. In this thesis, we propose the use of relation- based models for passage ranking and query expansion. We will propose two models, one using syntactic dependency relation and the other using semantic relation. Experimental results show that our syntactic dependency relation-based models significantly outperform density-based passage ranking with co-occurrence-based query expansion by up to 68% in mean reciprocal rank. Our semantic relation-based model also outperforms the density-based model by 4%. Based on this result, we combine the two models and propose a framework for passage retrieval for open domain question answering using relationship graph.