Retrieve, Rethink, and Review: Cross-Granularity Retrieval for Fact Verification

ACL ARR 2025 February Submission7401 Authors

Abstract: The dissemination of misinformation on online platforms has necessitated the development of automatic fact verification systems. Recent studies leverage semantic features of both claims and evidence to make predictions. However, these methods hypothesize that evidence is always available and accessible, which is undoubtedly impossible in real-world circumstances. Recent studies attempt to use retrieval-augmented approaches to retrieve relevant evidence to conduct fact verification tasks. However, these methods typically use the entire statement as a query to retrieve evidence, which may lead to missing relevant results. Besides, some studies utilize decomposed claims as queries, but they omit the filtering process, which may retrieve redundant information. Thus, to solve these challenges, we propose a novel Cross-granularity Retrieval-Augmented Network (CRAN) for open-domain fact verification. Specifically, we first utilize an LLM-based decomposer to divide the claims into atomic facts, facilitating sufficient retrieval. Besides, we leverage a novel reranking method to filter the redundant evidence. Then, we design a bipartite graph to fuse claim-evidence representations and make predictions. The experimental results on four common-used datasets demonstrate the effectiveness and superiority of our model.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: fact checking, rumor/misinformation detection

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7401