Leverage Web-based Answer Retrieval and Hierarchical Answer Selection to Improve the Performance of Live Question Answering

This paper describes the system we submitted to the Live QA Track in TREC 2015 which focused on live question answering for real-user questions. The real user questions extracted from Yahoo! Answers will be sent to the participant systems and the participants are required to provide an answer in real time (less than one minute). To address this task, we constructed a pipeline system which includes four components, i.e., question expansion, answer document retrieval, candidate answers selection and answer re-ranking. The purpose of the last three components aims at identifying accurate answers in a hierarchical manner. Then the answers with less than 1, 000 characters are returned to the server in less than one minute. Our system outperforms the averaged scores of all submitted runs on a 4-level Likert scale and ranks the 2th out of 14 teams.