Journal of Frontiers of Computer Science and Technology ›› 2024, Vol. 18 ›› Issue (11): 3015-3026.DOI: 10.3778/j.issn.1673-9418.2312040

• Artificial Intelligence·Pattern Recognition • Previous Articles     Next Articles

Document-Level Event Detection Method Based on Information Aggregation and Data Augmentation

TAN Lijun, HU Yanli, CAO Jianwei, TAN Zhen   

  1. National Key Laboratory of Information Systems Engineering, National University of Defense Technology, Changsha 410073, China
  • Online:2024-11-01 Published:2024-10-31

基于信息融合和数据增强的篇章级事件检测方法

谭立君,胡艳丽,曹健威,谭真   

  1. 国防科技大学 信息系统工程全国重点实验室,长沙 410073

Abstract: Event detection is a key task in the field of natural language processing, aiming to identify event trigger words and correctly classify their event types. Sentence-level event detection methods fail to effectively utilize intra-sentence and inter-sentence event relevance information, facing numerous challenges such as polysemy and event co-occurrence. Additionally, neural network-based event detection models require a large amount of text data for training, but the scarcity of corpus data severely affects the accuracy of results and the stability of the model. To address these issues, this paper proposes a document-level event detection method based on information aggregation and data augmentation, called LGIA (local and global information aggregation). This method adopts an encoder-decoder framework, designing a sentence-level local information extraction module based on dilated convolutional networks and a document-level global information extraction module based on conditional layer normalization, to deeply explore the contextual semantic information and the event correlations of the entire document. Meanwhile, this paper employs a data augmentation strategy of synonym replacement to effectively expand the data samples, thereby alleviating the impact of data scarcity. Experimental results validate that the proposed LGIA method achieves good results on the ACE2005 dataset and significantly improves performance on the augmented TAC-KBP2017 dataset, with F1 scores reaching 77.6% and 65.3%, respectively, demonstrating superior performance compared with existing baseline methods.

Key words: event detection, information aggregation, data augmentation, encoder-decoder framework

摘要: 事件检测是自然语言处理领域的关键任务,旨在识别事件触发词并正确分类其事件类型。语句级事件检测方法未能有效利用文本中的句内和句间事件相关性信息,面临着一词多义、事件共现等众多难题。此外,基于神经网络的事件检测模型需要大量的文本数据作为训练支撑,但语料库的数据不足严重影响着结果的准确率及模型的稳定性。针对上述问题,提出了基于信息融合和数据增强的篇章级事件检测方法LGIA。该方法采用编-解码框架,设计了基于膨胀卷积网络的句子级局部信息抽取模块和基于条件层归一化的篇章级全局信息抽取模块,以深入挖掘整个文档的上下文语义信息和事件间的相关性。同时,采用了同义词替换的数据增强策略,有效扩充了数据样本,从而缓解了数据不足问题带来的影响。经实验验证,LGIA方法在ACE2005数据集上取得了较好的结果,并在数据增强后的TAC-KBP2017数据集上得到了显著的性能提升,F1值分别达到了77.6%和65.3%,相较于现有的基线方法展现出了更优越的性能表现。

关键词: 事件检测, 信息融合, 数据增强, 编码-解码框架