基于FinBERT和Bi-GRU的篇章级事件检测模型
Document level event detection model based on FinBERT and Bi-GRU
廖涛 1王凯 1张顺香1
作者信息
- 1. 安徽理工大学 计算机科学与工程学院,安徽 淮南,232001
- 折叠
摘要
现有的篇章级事件检测方法一般分为候选实体抽取和事件检测两个子任务,候选实体抽取任务的精度影响着事件检测任务的结果.针对现有方法候选实体抽取精度不足的问题,提出了基于FinBERT和Bi-GRU的篇章级事件检测模型.首先,采用FinBERT预训练模型提升词嵌入向量对金融语义的表示能力,从而增强模型对金融语义的感知.其次,将上一层获取到的语义表示输入Bi-GRU模型和多头注意力机制中捕获全局和局部特征,通过CRF进行解码并标注出候选实体和实体类型.最后,根据抽取的候选实体信息判断篇章中存在的预定义事件.实验结果表明,本文方法提升了候选实体抽取任务的精度,取得了较好的篇章级事件检测结果.
Abstract
The existing document level event detection methods are generally used to perform two sub-tasks:candidate enti-ty extraction and event detection,and the precision of candidate entity extraction affects the results of event detection.A docu-ment level event detection model based on FinBERT and Bi-GRU is proposed against the insufficient accuracy of the existing methods of candidate entity extraction.Firstly,FinBERT pre-training model is used to improve the representation ability of word embedding vector to financial semantics,so as to enhance the perception of financial semantics of the model.Then,the semantic representation obtained from the previous layer is input into the Bi-GRU model and multi-head attention mechanism to capture global and local features,decode and mark the candidate entities and entity types through CRF.Finally,according to the extract-ed candidate entity information,the predefined events in the document are judged.The experimental results show that the model improves the precision of the candidate entity extraction task and obtains better document level event detection results.
关键词
篇章级事件检测/候选实体/FinBERT/Bi-GRUKey words
document level event detection/candidate entity/FinBERT/Bi-GRU引用本文复制引用
基金项目
国家自然科学基金面上项目(62076006)
安徽省属高校协同创新项目(GXXT-2021-008)
安徽省自然科学基金面上项目(1908085MF189)
出版年
2024
NETL
NSTL
万方数据