置信度引导迭代检索与证据图推理的事实核查方法OA

To balance efficiency and reasoning depth in automated fact-checking,a framework of couples iterative retrieval with deep evidence graph reasoning is proposed.Anchored in a confidence-driven iterative mechanism,the confidence of input claims is first assessed by leveraging the intrinsic knowledge of large language models in the framework.For claims assigned low confidence,precise search queries are formulated to retrieve initial evidence and an early termination is employed to avoid redundancy.Then,a three-layer evidence graph is constructed to hierarchically integrate and reason over the retrieved evidence,replacing the single internal knowledge judgment to enhance verification accuracy in low-confidence scenarios.Finally,the trade-off between reasoning and cost is balanced by dynamically controlling retrieval depth.Experimental evaluations on two public datasets demonstrate that the proposed framework achieves a F1 score improvement of 1 to 6 percentage points over the FIRE baseline and reduces search times by up to 86.5%against those of SAFE baseline.In summary,this framework demonstrates feasibility in terms of reduced retrieval costs alongside consistent high ac-curacy.