HopRetriever: Retrieve Hops over Wikipedia to Answer Complex Questions

Collecting supporting evidence from large corpora of text (e.g., Wikipedia) is great challenge for open-domain Question Answering (QA). Especially, multi-hop QA, scattered pieces are required to be gathered together support the answer extraction. In this paper, we propose a new retrieval target, hop, collect hidden reasoning Wikipedia complex question answering. Specifically, hop in paper defined as combination hyperlink and corresponding outbound link document. The encoded mention embedding which models structured knowledge how entity mentioned textual context, document representing unstructured within it. Accordingly, build HopRetriever retrieves hops over questions. Experiments on HotpotQA dataset demonstrate that outperforms previously published methods by margins. Moreover, our approach also yields quantifiable interpretations collection process.