Transforming relation extraction into fact verification utilizing large language models

This paper introduces an innovative approach for relation extraction utilizing large language models, named FVRE, which redefine the conventional task of information extraction as a process of fact verification to bridge the gap between pretraining and inference of this task. By aligning the task more closely with the inherent generative strengths of large models, this method notably bridges the gap between relation extraction and the models' capabilities. It thereby potentially increases the precision of extracted relations. We empirically validated our approach through rigorous testing on two distinct datasets: SemEval 2010 Task 8 and the CCKS 2019 IPRE corpus. The experimental results indicate that our method not only holds up under zero-shot scenarios but also significantly outperforms the baseline of directly applying large models to information extraction tasks. This demonstrates the robustness and effectiveness of our proposed paradigm in enhancing relation extraction performance for large language models.