基于BERT语义增强的因果关系抽取模型  

Causal relationship extraction model based on BERT semantic enhancement

作　　者：孙争艳 张顺香[2,3] 陈磊 朱广丽 魏苏波[2,3] SUN Zheng-yan;ZHANG Shun-xiang;CHEN Lei;ZHU Guang-li;WEI Su-bo(School of Computer Science,Huainan Normal University,Huainan 232038,China;School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China;Anhui Artificial Intelligence Laboratory,Artificial Intelligence Research Institute of Hefei Comprehensive National Science Center,Hefei 230000,China)

机构地区：[1]淮南师范学院计算机学院,安徽淮南232038 [2]安徽理工大学计算机科学与工程学院,安徽淮南232001 [3]合肥综合性国家科学中心人工智能研究院安徽省人工智能实验室,安徽合肥230000

出　　处：《计算机工程与设计》2024年第12期3758-3763,共6页Computer Engineering and Design

基　　金：安徽省科研计划编制项目重点基金项目(2024AH051731);国家自然科学基金面上基金项目(62076006);安徽高校协同创新基金项目(GXXT-2021-008);安徽省重点研发计划国际科技合作专项基金项目(202004b11020029);国家重点实验室开放基金项目(COGOS-2023HE02);淮南市指导性科技计划基金项目(4302)。

摘　　要：在金融领域,由于专业术语的复杂性,以及句子之间的依赖性,导致因果关系抽取的准确率较低。针对这一问题,提出一种基于BERT语义增强的因果抽取模型,建立基本模型和增强模型,以获取丰富的文本特征实现语义深度提取。使用BERT预训练模型得到上下文特征,通过对抗神经网络的对抗学习进一步学习高区分度特征,以此提高因果关系抽取的准确性。实验结果表明,该模型能够提高因果关系抽取的准确性。In the financial field,the complexity of professional terminology and the dependence between sentences lead to a lower accuracy of causality extraction.To address this issue,a causality extraction model based on BERT semantic enhancement was proposed,including a basic model and an enhanced model,to obtain rich text features and achieve semantic deep extraction.The BERT pre-training model was used to obtain contextual features,and the adversarial learning of the adversarial neural network was employed to further learn high-discriminative features,thus improving the accuracy of causality extraction.Experimental results demonstrate that the proposed model can improve the accuracy of causality extraction.

关 键 词：因果关系抽取 信息抽取 金融领域 对抗神经网络 对抗学习 基本模型 增强模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]