| 摘 要: 文档级事件抽取作为信息抽取领域的一项颇具挑战性任务,其关键在于如何精确识别并抽取文档中蕴含的多个事件信息。针对这一难题,提出了一种双阶段事件抽取方法。在事件类型检测阶段,将问题转化为基于对比学习的多标签分类任务,并通过创新性的事件子空间的动态负例选择策略,有效增强了不同类型事件之间的区分度。在事件论元抽取阶段,利用大模型进行多阶段推理,完成事件论元的识别。实验结果表明,所提方法F1值提升了0.6个百分点,验证了本文方法的有效性和可行性。 |
| 关键词: 对比学习 大语言模型 事件抽取 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 新疆维吾尔自治区高校科研基金资助项目(XJEDU2024094) |
|
| A Two-Stage Event Extraction Method Based on Contrastive Learning and Large Language Models |
|
ZHOU Peiyao1,2, HU Zhikui1,2, ZHANG Dawei2, ZI Kangli2
|
(1.School of Computer, Jiangsu University of Science and Technology, Zhenjiang 212000, China;
2.Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China)
zpy1690934380@163.com; h3hzz@outlook.com; zhangdawei_just@126.com; czikangli@126.com
|
| Abstract: Documen-t level event extraction, as a challenging task in the field of information extraction, hinges on the accurate identification and extraction of multiple even-t related pieces of information within a document. To address this challenge, this paper proposes a two-stage event extraction method. In the event type detection stage, the problem is transformed into a mult-i label classification task based on contrastive learning. By innovatively employing a dynamic negative sampling strategy within event subspaces, the distinction between different event types is significantly enhanced. In the event argument extraction stage, large language models are utilized for mult-i step reasoning to identify event arguments. Experimental results demonstrate that the proposed method achieves a 0.6 percentage points improvement in F1-score, verifying its effectiveness and feasibility. |
| Keywords: contrastive learning Large Language Model event extraction |
