Q-REFormer: Aligning Q-Former with LLM for Text-based data Relation Extraction

Relation extraction aims to identify semantic associations between entities in text, serving as a fundamental task for constructing knowledge graphs and enhancing information retrieval. Traditional methods that rely on local contextual analysis exhibit limitations in capturing implicit relationships between distantly located entities, particularly when entity pairs are separated by long syntactic dependencies. To address this challenge, this paper proposes Q-REFormer, a BLIP-2 framework-based relation extraction model featuring two core innovations: (1) The integration of learnable query vectors (Q-former) to align structured relation descriptions generated by large language models (LLMs) with sentence encodings, thereby enhancing cross-entity semantic reasoning capabilities; (2) A novel entity-aware attention mechanism that dynamically allocates interaction weights among entity pairs to resolve long-range dependency issues. Experimental evaluations on ACE 2005 and SemEval 2010 datasets demonstrate that Q-REFormer achieves F1 scores of 88.02% and 92.08% respectively, showing significant superiority over existing models. Ablation studies further validate the effectiveness of LLM-generated relation descriptions and the query learning module.