大模型知识引导的复合多注意力文档级关系抽取方法

竹志超,李建强,齐宏智,赵青,高齐,李思颖,蔡嘉怡,沈金炎

Large model knowledge-guided composite multi-attention method for document-level relation extraction

Zhichao ZHU,Jianqiang LI,Hongzhi QI,Qing ZHAO,Qi GAO,Siying LI,Jiayi CAI,Jinyan SHEN

表 1 与先进模型的性能对比结果

Tab.1 Performance comparison results with advanced models

类别 模型 P /% R /% F1/% Sequence-based RoBERTa-base 79.92±0.72 78.60±1.02 79.25±0.75 SSAN 80.61±1.22 81.06±0.81 80.83±0.99 Graph-based GCGCN 80.98±0.46 81.54±0.77 81.26±0.51 GLRE 82.42±0.80 82.27±0.63 82.34±0.70 GRACR 83.01±1.34 83.13±0.61 83.57±0.68 Knowledge-based DISCO 83.66±0.25 84.28±0.46 83.97±0.42 K-BiOnt 85.14±0.39 84.36±0.50 84.75±0.39 KIRE 84.90±0.44 85.23±0.38 85.06±0.41 KRC 85.06±0.22 86.00±0.08 85.53±0.09 GECANet 86.31±0.15 85.55±0.14 85.93±0.14 LLM-based ChatGLM2-6B 37.42±4.32 41.21±3.48 39.22±3.90 LLaMA3-8B 40.56±2.79 43.75±3.50 42.09±2.83 Qwen-32B 45.76±3.68 48.20±4.61 46.95±4.46 — LKCM 87.26±0.16 87.69±0.09 87.47±0.12