Multi-grained unsupervised evidence retrieval for question answering

Evidence retrieval is a crucial step in question answering (QA) tasks, which can filter original context to provide supporting evidence for reading comprehension and reduce the time of answer inference. Most research interests in evidence retrieval have been paid to supervised methods, while unsupervised evidence retrieval received limited attention. However, most existing works for unsupervised evidence retrieval regard sentence-level and passage-level evidence retrieval as two independent processes. In order to fuse these two levels of information, we propose an efficient and unsupervised multi-grained evidence retrieval method for QA, which considers multiple interactions, including query-sentence, query-passage, and passage-passage. Specifically, we use sentence-level retrieval to obtain an evidence framework. Then we propose a score fusion mechanism to model the unilateral guiding relationship between sentence-level and passage-level retrieval. On the basis of score fusion, we propose a gated selection mechanism to retrieve evidence passages, which can supplement reasoning information for the evidence framework. Extensive experiments on open-domain and multi-hop QA datasets show that our method has fast retrieval speed while achieving better retrieval and QA performance than baselines.