Ranking Answers and Web Passages for Non-factoid Question Answering: Emory University at TREC LiveQA

This paper describes a question answering system built by a team from Emory University to participate in TREC LiveQA’15 shared task. The goal of this task was to automatically answer questions posted to Yahoo! Answers community question answering website in real-time. My system combines candidates extracted from answers to similar questions previously posted to Yahoo! Answers and web passages from documents retrieved using web search. The candidates are ranked by a trained linear model and the top candidate is returned as the final answer. The ranking model is trained on question and answer (QnA) pairs from Yahoo! Answers archive using pairwise ranking criterion. Candidates are represented with a set of features, which includes statistics about candidate text, question term matches and retrieval scores, associations between question and candidate text terms and the score returned by a Long Short-Term Memory (LSTM) neural network model. Our system ranked top 5 by answer precision, and took 7th place according to the average answer score. In this paper I will describe our approach in detail, present the results and analysis of the system.