Ranking Like Human: Global-View Matching via Reinforcement Learning for Answer Selection

Answer Selection (AS) is of great importance for open-domain Question Answering (QA). Previous approaches typically model each pair of the question and the candidate answers independently. However, when selecting correct answers from the candidate set, the question is usually too brief to provide enough matching information for the right decision. In this paper, we propose a reinforcement learning framework that utilizes the rich overlapping information among answer candidates to help judge the correctness of each candidate. In particular, we design a policy network, whose state aggregates both the question-candidate matching information and the candidate-candidate matching information through a global-view encoder. Experiments on the benchmark of WikiQA and SelQA demonstrate that our RL framework substantially improves the ranking performance.