图书情报工作 ›› 2021, Vol. 65 ›› Issue (24): 89-99.DOI: 10.13266/j.issn.0252-3116.2021.24.010

• 知识组织 • 上一篇    下一篇

基于多层卷积神经网络的金融事件联合抽取方法

李旭晖1,2, 程威1, 唐小雅3, 于滔1, 陈壮3, 钱铁云3   

  1. 1. 武汉大学信息管理学院 武汉 430072;
    2. 武汉大学大数据研究院 武汉 430072;
    3. 武汉大学计算机学院 武汉 430072
  • 收稿日期:2021-06-15 修回日期:2021-09-27 出版日期:2021-12-20 发布日期:2021-12-29
  • 作者简介:李旭晖,副教授,硕士生导师,E-mail:lixuhui@whu.edu.cn;程威,硕士研究生;唐小雅,硕士研究生;于滔,硕士研究生;陈壮,博士研究生;钱铁云,教授,博士生导师。
  • 基金资助:
    本文系国家自然科学基金重大研究计划"大数据驱动的管理与决策研究"重点支持项目"基于知识关联的金融大数据价值分析、发现及协同创造机制"(项目编号:91646206)和深证信息联合研究计划课题"企业全生命周期关键事件识别和要素抽取"(项目编号:CHINFO201802)研究成果之一。

A Joint Extraction Method of Financial Events Based on Multi-Layer Convolutional Neural Networks

Li Xuhui1,2, Cheng Wei1, Tang Xiaoya3, Yu Tao1, Chen Zhuang3, Qian Tieyun3   

  1. 1. School of Information Management, Wuhan University, Wuhan 430072;
    2. Big Data Institute, Wuhan University, Wuhan 430072;
    3. School of Computer Science, Wuhan University, Wuhan 430072
  • Received:2021-06-15 Revised:2021-09-27 Online:2021-12-20 Published:2021-12-29

摘要: [目的/意义] 为进一步提升金融领域事件抽取的效果,增强事件抽取两个子任务之间的关联性。[方法/过程] 在中文金融文本上进行事件抽取相关研究,提出一种融合预训练模型与多层卷积神经网络的金融事件联合抽取方法,首先通过预训练模型BERT捕捉句子序列的综合语义信息,然后接入本文设计的多层卷积架构MultiCNN,分层提取局部窗口和高维空间语义信息,同时实现事件识别和要素抽取这两个任务,再通过引入对比损失,进一步强化两个任务之间的关联。[结果/结论] 在中文金融事件数据集上F1达到82.20%,比各个基准抽取模型均有一定提升。

关键词: 中文事件抽取, 卷积神经网络, 预训练模型, 联合学习

Abstract: [Purpose/significance] In order to further improve the effect of event extraction in the financial field, the correlation between the two subtasks of event extraction needs to be enhanced.[Method/process] This paper carried out related research about event extraction on Chinese financial texts,and proposed a joint extraction method of financial events that integrated the pre-training model and a multi-layer convolutional neural network. First, the pre-training model BERT captured the comprehensive semantic information of the sentence sequence, then accessed the multi-layer convolutional architecture designed in this paper——MultiCNN, hierarchically extracted local window and high-dimensional spatial semantic information, realized the two tasks of event recognition and element extraction at the same time, and then introduced contrast loss to further strengthen the association between the two tasks.[Result/conclusion] F1 has reached 82.20% on the Chinese financial event data set, which has a certain improvement over the benchmark extraction models.

Key words: Chinese event extraction, convolution neural network, pretraining model, joint learning

中图分类号: