摘要
【目的/意义】针对冲突事件领域数据集较少,现有事件抽取方法强依赖于大量数据,在低资源场景上面临着训练效果差、泛化能力不足的问题,本文提出了一种基于先验提示模板的小样本事件抽取方法PriorPromptEE。【方法/过程】本文首先基于“政治、军事”维度,完善冲突事件表示框架,采用人机协同方式,构建冲突事件小样本数据集。其次,本文为提示模板设计构造策略,在模型编码层嵌入基于先验知识的提示模板,输出层采用双指针实现事件元素的预测。【结果/结论】实验结果表明,PriorPromptEE能够在冲突事件小样本数据集上取得较好效果。对比序列模型,提升42%至45%,对比传统事件抽取模型,提升19%至45%,对比提示学习零样本抽取模型,提升19%,PriorPromptEE达到0.85,验证了该方法的有效性。【创新/局限】本文从“政治、军事”维度完善了冲突事件表示框架,采用提示学习的方式为冲突事件的核心元素抽取提供了新的思路,未来可从经济、舆情等维度完善表示框架,强化冲突事件抽取模型框架。
【Purpose/significance】In the field of conflict events,there is a lack of datasets,and existing event extraction methods heavily rely on large amounts of data,leading to poor training effectiveness and insufficient generalization ability in low-resource scenarios.To address these issues,this paper proposes a few-shot event extraction method,PriorPromptEE,based on prior prompt templates.【Method/process】Firstly,this paper improves the conflict event representation framework based on the political and military dimension,and constructs a few-shot conflict event extraction dataset using a semi-automatic labeling method.Secondly,a pipeline is used to design and construct strategies for prompt template generation.The prompt templates are embedded in the model's encoding layer based on prior knowledge,and the output layer employs a double-pointer mechanism to extract entity start and end positions,thus achieving event argument prediction.【Result/conclusion】Experimental results show that the proposed event extraction method,which incorporates prior prompt templates,performs well on few-shot conflict event dataset.Compared to sequence models,the score is improved by 42%to 45%,compared to traditional event extraction models,the score is improved by 19%to 45%,and compared to zeroshot event extraction models,the score is improved by 19%.The score of the proposed model,PriorPromptEE,reaches up to 0.85,validating the effectiveness of this approach.【Innovation/limitation】This paper improves the framework for representing conflict events from the perspective of international relations of political and military,and provides a new approach for identifying conflict types and extracting core elements based on prior prompt templates.In the future,the framework for representing conflict events could be further enhanced from economic,public opinion,and other dimensions to strengthen the model framework for conflict event extraction.
作者
陆伟
冯子琨
程齐凯
石湘
熊资
LU Wei;FENG Zikun;CHENG Qikai;SHI Xiang;XIONG Zi(School of Information Management,Wuhan University,Wuhan 430072,China;Wuhan University Information Retrieval and Knowledge Mining Laboratory,Wuhan University,Wuhan 430072,China)
出处
《情报科学》
CSSCI
北大核心
2024年第4期1-8,共8页
Information Science
基金
国家自然科学基金面上项目“基于机器阅读理解的科学命题文本论证逻辑识别”(72174157)。
关键词
冲突事件
事件抽取
小样本
提示学习
提示模板构造
conflict event
event extraction
few-shot
prompt learning
prompt template construction