基于关系层级嵌入与多智能体协同的复杂关系抽取OA

Sentence-level relation extraction aims to predict semantic relationships between different entities within sentences.Current research faces two major limitations.Existing methods generally overlook hierarchical structural features of relations,restricting models'capability to represent complex semantic relationships,while in relation-intensive scenarios,the escalating complexity of prompt engineering for large language models(LLMs)leads to significant degradation in reasoning efficiency.These deficiencies critically impair the accuracy of complex medical relation extraction.To address these challenges,we propose a hierarchical relation classification approach and develop a collaborative framework integrating both large and small models.Additionally,we design a few-shot learning paradigm that not only constrains LLMs'output formats but also provides demonstration samples.To mitigate error propagation in joint modeling,we introduce a multi-agent collaborative approach to enhance model reliability.Experimental results demonstrate that the proposed relational level classification(RLC)method significantly outperforms existing baseline models.On the CHIP dataset,the ChatGLM model with RLC achieves25.11 percentage points improvement over the original model.Compared to the best-performing relation semantic enhancement method,the RLC method shows superior performance with5.61 and 4.78 percentage points increases in F1-score on the CHIP and BioRel datasets,respectively.Further research indicates that the multi-agent collaborative approach exhibits enhanced representational capabilities,delivering additional improvements of 1.68 and2.78 percentage points in F1-score on the CHIP and BioRel datasets,respectively,when combined with the fine-tuned RLC method.