Abstract

We investigate the effectiveness of lexical, topic and structural similarities on the semantic relevance between a question and a passage which may contain the answer. We propose a web-based method to measure the lexical similarity between a question and a passage based on the semantic similarity between words or phrases. The topic similarity between a question and a passage is evaluated using a probabilistic language model. For structural similarity, we solve it as a simple relation constraint satisfaction problem. We then integrate the three different similarity metrics into a weighted average metrics for evaluation of the relevance of a passage to a question. The experiment results show that our approach outperforms the three baselines by up to 76.83% in MRR of top-1 results, 73.48% in MRR of top-5 results, and 77.13% in MAP.

Paper Details