CARE: Contextual Augmentation with Retrieval Enhancement for Relation Extraction in Large Language Models

Danjie Han, Heyan Huang, Shumin Shi, Cunhan Guo, Xun Li, Yanghao Zhou, Changsen Yuan

Abstract: To address the issues of inconsistent relation format outputs and contextual degradation in large language models (LLMs) for relation extraction tasks, a cross-level context retrieval-enhanced framework is proposed. First, to mitigate label inconsistency, formatting errors, and semantic deviation during inference, a relation label correction mechanism is designed based on semantic similarity and syntactic structure. This mechanism calibrates the outputs of LLMs, thereby improving the accuracy and consistency of predicted relation types. Second, to meet the contextual modeling demands of different types of instance bags, a hierarchical context augmentation strategy is introduced. For multi-sentence instance bags, a graph-based retrieval enhancement mechanism is employed, combining intra-bag entity co-occurrence networks with document-level sentence relation graphs to enhance cross-sentence semantic understanding. For single-sentence instance bags, a TF-IDF-based semantically similar sentence expansion strategy is developed to enrich training contexts while preserving semantic consistency, alleviating the problem of insufficient contextual information. Finally, a low-rank adaptation (LoRA) mechanism is adopted to enable parameter-efficient fine-tuning of LLMs, significantly reducing training overhead while maintaining competitive performance, thus improving their practicality in relation extraction tasks.

External IDs:dblp:conf/nlpcc/HanHSGLZY25