摘要
事件抽取是自然语言处理的重要任务,而事件检测是事件抽取的关键步骤之一,其目标是检测事件的发生并对其进行分类。目前基于触发器识别的中文事件检测方法存在一词多义、词与触发词不匹配的问题,影响了事件检测模型的精度。针对此问题,提出基于双重注意力的无触发词事件检测模型(Event Detection Without Triggers based on Dual Attention,EDWTDA),该模型可跳过触发词识别过程,实现在无触发词标记情况下直接判断事件类型。EDWTDA利用ALBERT改善词嵌入向量的语义表示能力,缓解一词多义问题,提高模型预测能力;采用局部注意力融合事件类型捕捉句中关键语义信息并模拟隐藏的事件触发词,解决词与触发词不匹配的问题;借助全局注意力挖掘文档中的语境信息,解决一词多义问题;最后将事件检测转化成二分类任务,解决多标签问题。同时,采用Focal loss损失函数解决转化成二分类后产生的样本不均衡问题。在ACE2005中文语料库上的实验结果表明,所提模型相比最佳基线模型JMCEE在精确率、召回率和F1-score评价指标上分别提高了3.40%,3.90%,3.67%。
Event extraction is an essential task of natural language processing,and event detection is one of the critical steps of event extraction,whose goal is to detect the occurrence of events and classify them.Currently,Chinese event detection has problems of polysemous words and mismatches between words and triggers,which affect the accuracy of event detection models.We propose the event detection without triggers based on dual attention(EDWTDA),which skips the process of trigger word recognition and directly determines event types without trigger word tags.First,the ALBERT model is applied to improve the semantic representation ability of word embedding vectors.Second,we fusion local attention and event types to capture key semantic information and simulate hidden event triggers to solve the problem of mismatch between words and triggers.Third,the global attention is introduced to mine contextual information in documents to solve the problem of polysemous words.Further,the event detection task is converted into a binary classification task for solving multi-label problem.Finally,the focal loss function is used to address the sample imbalance after conversion.Experimental results on the ACE2005 Chinese corpus show that compared with the best baseline model JMCEE,the accuracy rate,recall rate,and F1-score of the proposed model increases by 3.40%,3.90% and 3.67%,respectively.
作者
程永
毛莺池
万旭
王龙宝
朱敏
CHENG Yong;MAO Yingchi;WAN Xu;WANG Longbao;ZHU Min(Key Laboratory of Water Big Data Technology of Ministry of Water Resources,Nanjing 210098,China;College of Computer and Information,Hohai University,Nanjing 211100,China)
出处
《计算机科学》
CSCD
北大核心
2023年第1期276-284,共9页
Computer Science
基金
江苏省重点研发计划(BE2020729)
中国华能集团关键技术(HNKJ19-H12,HNK20-H64)。
