GRAAL: Graph-Based Retrieval for Collecting Related Passages across Multiple Documents †

Finding passages related to a sentence over a large collection of text documents is a fundamental task for claim verification and open-domain question answering. For instance, a common approach for verifying a claim is to extract short snippets of relevant text from a collection of reference documents and provide them as input to a natural language inference machine that determines whether the claim can be deduced or refuted. Available approaches struggle when several pieces of evidence from different documents need to be combined to make an inference, as individual documents often have a low relevance with the input and are therefore excluded. We propose GRAAL (GRAph-based retrievAL), a novel graph-based approach that outlines the relevant evidence as a subgraph of a large graph that summarizes the whole corpus. We assess the validity of this approach by building a large graph that represents co-occurring entity mentions on a corpus of Wikipedia pages and using this graph to identify candidate text relevant to a claim across multiple pages. Our experiments on a subset of FEVER, a popular benchmark, show that the proposed approach is effective in identifying short passages related to a claim from multiple documents.