基于提示学习的篇章级事件论元抽取方法研究  

Research on Document Level Event Argument Extraction Method Based on Prompt Learning

作　　者：薛继伟[1] 胡馨元 薛鹏杰 XUE Ji-wei;HU Xin-yuan;XUE Peng-jie(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163000,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163000

出　　处：《计算机技术与发展》2024年第6期125-131,共7页Computer Technology and Development

基　　金：黑龙江省省属本科高校基本科研业务费(2022TSTD-03)。

摘　　要：事件论元抽取是指在自然语言文本中识别出事件论元及其对应的角色,是事件抽取的关键。传统事件论元抽取方法将抽取范围局限在单个句子中,在面对长文本中论元分散的情况时表现不佳。近年来,有研究者提出基于提示学习的篇章级事件论元抽取方法,能根据提示信息在输入文本中获取事件论元,实现事件论元抽取。然而现有基于提示学习的方法大多是由人工手动构建提示模板,模板结构固定容易导致论元抽取错误。针对以上不足,该文在以往基于提示学习研究的基础上,提出以文本触发词为关键实现模板自动构建的方法,并在输入文本中融入事件角色语义信息,使模型能更好地捕获文本语义特征,提高事件论元抽取准确率。在篇章级数据集RAMS上的实验结果表明,该模型在事件论元识别和事件论元分类的F1值分别达到54.3%和48.1%,相比最优的基准方法分别提升了1.8百分点和1.2百分点,验证了模型的有效性。Event argument extraction refers to the recognition of event arguments and their corresponding roles in natural language texts,which is the key to event extraction.The traditional event argument extraction method limits the extraction scope to a single sentence and performs poorly when faced with scattered arguments in long texts.In recent years,researchers have proposed a discourse level event argument extraction method based on prompt learning,which can obtain event arguments from input text based on prompt information and achieve event argument extraction.However,most existing methods based on prompt learning rely on manual construction of prompt templates,and fixed template structures can easily lead to argument extraction errors.In response to the above shortcomings,we propose a method for automatically constructing templates based on text trigger words based on previous research on prompt learning,and integrate event role semantic information into the input text,enabling the model to better capture text semantic features and improve the accuracy of event argument extraction.The experimental results on the discourse level dataset RAMS show that the F1 values of the proposed model in event argument recognition and event argument classification reach 54.3%and 48.1%,respectively,which are 1.8 and 1.2 percentage points higher than the optimal benchmark method,respectively,verifying the effectiveness of the model.

关 键 词：论元抽取 提示学习 触发词 跨度选择器 预训练语言模型 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP391.4[自动化与计算机技术—控制科学与工程]