Unsupervised Alignment-based Iterative Evidence Retrieval for Multi-hop Question Answering

Evidence retrieval is a critical stage of question answering (QA), necessary\nnot only to improve performance, but also to explain the decisions of the\ncorresponding QA method. We introduce a simple, fast, and unsupervised\niterative evidence retrieval method, which relies on three ideas: (a) an\nunsupervised alignment approach to soft-align questions and answers with\njustification sentences using only GloVe embeddings, (b) an iterative process\nthat reformulates queries focusing on terms that are not covered by existing\njustifications, which (c) a stopping criterion that terminates retrieval when\nthe terms in the given question and candidate answers are covered by the\nretrieved justifications. Despite its simplicity, our approach outperforms all\nthe previous methods (including supervised methods) on the evidence selection\ntask on two datasets: MultiRC and QASC. When these evidence sentences are fed\ninto a RoBERTa answer classification component, we achieve state-of-the-art QA\nperformance on these two datasets.\n