Abstract
Relation triple extraction, which outputs a set of triples from longsentences, plays a vital role in knowledge acquisition. Large language modelscan accurately extract triples from simple sentences through few-shot learningor fine-tuning when given appropriate instructions. However, they often missout when extracting from complex sentences. In this paper, we design anevaluation-filtering framework that integrates large language models with smallmodels for relational triple extraction tasks. The framework includes anevaluation model that can extract related entity pairs with high precision. Wepropose a simple labeling principle and a deep neural network to build themodel, embedding the outputs as prompts into the extraction process of thelarge model. We conduct extensive experiments to demonstrate that the proposedmethod can assist large language models in obtaining more accurate extractionresults, especially from complex sentences containing multiple relationaltriples. Our evaluation model can also be embedded into traditional extractionmodels to enhance their extraction precision from complex sentences.