RRetFC: Leveraging Recursive Retrieval For LLM-Enhanced Complex Fact-Checking

Yuxuan Xie, Xiaoliang Liu, Peng Wu, Li Pan

Abstract: Fake claims spreading in social media are often carefully crafted and complex, so that multi-source evidence and reasoning are required for checking their authenticity. To this end, we propose RRetFC, a Recursively Retrieval enhanced method for complex Fact Checking with LLMs, which has a multi-step recursive retriever and an LLM-enhanced reader. Specifically, for the evidence retriever, as the long-range semantic dependencies between explicit entities of a complex claim may involve some other hidden entities, it is hard for traditional retrieval methods to get all necessary evidence for fact-checking in once retrieval. Thus RRetFC designs a recursive retriever which gets all necessary evidence recursively with the help of LLM. For the reader, as the evidence retrieved from some external knowledge sources may inevitably contain noise which may undermining the judgment, RRetFC firstly employs an LLM to filter out noise from the retrieved evidence and carefully reason to generate an analysis report. Subsequently, RRetFC fine-tunes a SLMs based on the analysis text to produce the final prediction label. Our experiments on two public fact-checking datasets show that RRetFC outperforms state-of-the-art baseline methods and exhibits superior evidence retrieval capabilities.

External IDs:dblp:conf/icann/XieLWP25