CACRE: A Weakly Supervised Method for Cross-Attention Contrastive Relation Extraction With Large Language Models

In the relation extraction (RE) task, large language models (LLMs) have shown remarkable capabilities in predicting unknown relations, offering significant improvements in efficiency and flexibility over traditional methods. However, the probabilistic nature of the generation process in LLMs may lead to the occurrence of hallucinations, causing inaccurate relation triples be generated. To mitigate this problem, this paper proposes a novel weakly supervised method, Cross-Attention Contrastive Relation Extraction (CACRE), which aims at detecting erroneous relation triples generated by LLMs and then effectively distinguishing valid ones. The CACRE leverages contrastive learning and cross-attention mechanisms. Specifically, contrastive learning is applied to distinguish between positive and negative relation triples, enhancing the model’s feature extraction capability by learning discriminative features. Subsequently, a cross-attention mechanism is employed to capture the semantic associations between texts and triples, thereby improving the model’s ability to understand and extract information from the input content. Experiment results on the DuIE2.0 dataset and the TACRED dataset demonstrate that CACRE significantly outperforms baseline LLMs, with average improvements of 12% and 8% in precision, respectively.