Selecting Paragraphs to Answer Questions for Multi-passage Machine Reading Comprehension

摘要

This paper addresses the problem of question answering style multi-passage Machine Reading Comprehension (MRC) and suggests that paragraph-level segments are suitable to answer questions in real Web query scenario. We propose to combine a learning to rank framework with an attention-based neural network to select the best-matching paragraph for a specific question. To estimate the quality of a paragraph with respect to a given query, its largest ROUGE-L score compared against the annotated answers is used as the ranking indicator. Experimental results on a real-world dataset demonstrate that the proposed method obtains a significant improvement compared to the state-of-the-art baselines.