A Retriever-Reasoner Method for Multi-Document Reading Comprehension

Abstract

Machine reading comprehension (MRC) is one of the main research of question answering system tasks. In real-world scenarios, due to the large scale of document data, and many questions that need to be answered urgently, traditional extractive MRC method faces low efficiency, low accuracy, and has poor reasoning ability. In order to efficiently and accurately extract knowledge from massive unstructured text data, we propose a multi-document MRC model based on a two-stage approach. Through document retrieval and answer reasoning, the model can hierarchically match the relevant passages from coarse-to-fine. Experimental results show that our method outperforms baseline and improves Rouge-L and BLEU-4 by 2.5% and 1.9%, respectively. In addition, the reasoning process is interpretable and can provide support for information understanding.