Hierarchical Answer Selection Framework for Multi-passage Machine Reading Comprehension

Abstract

Machine reading comprehension (MRC) on real web data, which means finding answers from a set of candidate passages for a question, is a quite arduous task in natural language processing. Most state-of-the-art approaches select answers from all passages or from only one single golden paragraph, which may cause the overlapping information and the lack of key information. To address these problems, this paper proposes a hierarchical answer selection framework that can select main content from a set of passages based on the question, and predict final answer within this main content. Specifically, three main parts are employed in this pipeline: First, the passage selection model uses a classification mechanism to select passages by passages content and title information which is not fully used in other models; Second, a key sentences sequence selection mechanism is modeled by Markov-Decision-Process (MDP) in order to gain as much as effectual answer information as possible; Finally, a match-LSTM model is employed to extract the final answer from the selected main content. These three modules that shared the same attention-based semantic network and we conduct experimental on DuReader search dataset. The results show that our framework outperforms the baseline by a large margin.