基于分阶段语义感知的事件抽取大语言模型框架

李延松,陈宁,刘锋光,陈盼,黄晓峰,葛慧丽

Large language model framework for event extraction based on staged semantic perception

Yansong LI,Ning CHEN,Fengguang LIU,Pan CHEN,Xiaofeng HUANG,Huili GE

表 2 在ACE05和CASIE数据集上的模型性能对比

Tab.2 Model performance comparison between ACE05 and CASIE dataset

方法 主干网络 模型参数规模 CASIE ACE05 F1e/% F1a/% F1e/% F1a/% Bert-base BERT-base 110×106 68.98 60.37 72.50 59.50 EEQA 2×BERT-base 220×106 — — 72.40 53.30 UIE T5-v1.1-base 770×106 69.33 61.30 73.36 54.79 USM RoBERTa-Large 355×106 71.73 63.26 72.41 55.83 InstructUIE FlanT5-11B 11×109 67.80 63.53 77.13 72.94 Schema-Aware-EE ChatGPT > 20×109 70.28 56.28 73.68 49.56 LC4EE GPT-4 > 175×109 — — 77.20 54.90 TALOR-EE GPT-3.5-turbo > 20×109 — — 70.50 47.70 本文方法 ChatGLM3 6×109 90.93 63.71 76.56 72.07 本文方法 GLM4-9B-0414 9×109 93.40 66.87 77.81 76.52