知识与句法融合的因果关系抽取网络  

Event causality identification network based on knowledge and syntactic structure

作　　者：汪诗蕊 解博涵 丁玲 陈建廷 向阳[1] WANG Shirui;XIE Bohan;DING Ling;CHEN Jianting;XIANG Yang(School of Electronic and Information Engineering,TongJi University,Shanghai 200000,China)

机构地区：[1]同济大学电子与信息工程学院,上海200000

出　　处：《大数据》2024年第3期82-92,共11页Big Data Research

基　　金：国家自然科学基金项目(No.72071145)。

摘　　要：因果关系抽取作为关系抽取的一个重要任务,近年来得到了广泛关注。现有的因果关系抽取方法大多将句法结构和背景知识割裂开进行研究,早期的因果关系抽取方法偏重于从句法结构层面进行分析,随着深度学习技术的发展,预训练模型结合背景知识的方法成为主流。然而上述两种方法均未完全融合句内信息和外部知识,带来了不同程度的信息缺失。为了解决这一问题,提出了结合句法结构和背景知识的因果关系抽取模型。该模型将句子解析为同时包含句法和知识的知识句法图结构,使用图卷积网络进行信息融合。模型同时考虑了句法和知识两部分信息,从而进一步丰富了实体嵌入,达到了良好的因果关系抽取效果。本模型在EventStoryLine数据集上取得了良好效果,F1值达到0.445,与现有方法相比提高了2.3%。Event causality identification is an important task of relationship extraction,which has received much attention recent years.Most of the existing methods separate syntactic structure from the background knowledge information.The early causality extraction methods focus on the analysis of syntactic structure level.With the development of deep learning,the methods that use the pre-training model combined with background knowledge has become the mainstream.However,neither of the above two kinds of methods fully integrates the sentence information and external knowledge,resulting in different degrees of information loss.To address this problem,we proposed a novel model of event causality identification combining syntactic structure and background knowledge.Our model parses sentences into knowledge syntactic graph structures that contain both syntax and knowledge,and uses the graph convolution network for information fusion.It considers both syntax and knowledge information,which further enriches the event representation and performs effectively.In experiments on the widely-used dataset EventStoryLine,the F1 score of our model achieves 0.445,a 2.3%improvement over existing methods.

关 键 词：因果关系抽取 预训练模型 图卷积网络 自然语言处理 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]