Multi-perspective Denoising Reader for Multi-paragraph Reading Comprehension

Fengcheng Yuan, Yanfu Xu, Zheng Lin, Weiping Wang, Gang Shi

Abstract: Multi-paragraph reading comprehension (MPRC) aims to answer questions based on a number of paragraphs. The background paragraphs in MPRC are usually collected by information retrieval (IR) systems. Most of the MPRC models indiscriminately read all the background paragraphs, and regard paragraphs which contain the answer string as ground truths. Among these paragraphs, some are not semantically related to the question, therefore, the noisy data will distract the model from finding the right answer. Besides, most of the MPRC methods only use the paragraph-question relevance to predict the answer span. However, the utilization of the paragraph-paragraph relevance that provides enhanced and complementary evidence has only been explored in limited methods. To address these issues, we propose a hierarchical model with a multi-level attention mechanism, which can leverage both the inter-relation (paragraph-question relevance) and the intra-relation (paragraph-paragraph relevance), to filter out noisy data and extract the final answer. We conduct experiments on two challenging public datasets Quasar-T and SearchQA. The results demonstrate that our model outperforms recent MPRC baselines.