【作者】 李明亮；

【导师】 王勇；

【作者基本信息】 广东工业大学 ， 计算机技术（专业学位）， 2022， 硕士

【摘要】 随着互联网用户的增加,用户之间交互所产生的数据慢慢累积,电子文本信息在互联网中呈现爆炸式的发展。当突发事件发生时,会有许多相关的事件信息在互联网中发布,如何从海量的文本中抽取结构化的突发事件信息也成为网络舆情分析的重要方面。目前突发事件抽取方法主要存在两个问题:一是在触发词抽取方法中,未能充分利用文本的上下文语义信息的重要性进行抽取。二是在论元抽取方法中,未能充分利用句子中的依存句法信息进行抽取。针对目前方法的局限性,本文提出一种基于深度学习的中文突发事件抽取算法,为了便于将各个模块进行优化,采用流水线的方式进行事件抽取,将事件抽取分为基于多头注意力机制的触发词抽取模型以及融合依存句法信息的论元抽取模型。主要研究内容如下:(1)提出基于多头注意力机制的触发词抽取模型:使用预训练模型BERT-wwm作为文本表示,该模型针对中文特有的语法表达,采用全词覆盖的方式进行训练,能够学习到中文词语的语义信息。为了获得文本的全局特征,使用BiLSTM对文本的前向与后向语义信息进行抽取。引入多头注意力机制进行语义信息提炼,多头注意力机制可以将对触发词抽取更重要的特征进行增强,削弱不相关或不重要的特征,使得特征分布更合理,同时能够从不同语义空间中捕捉文本特征,增强文本表达。考虑相邻节点之间的相关性,通过CRF联合解码的方式进行结果标注。(2)提出融合依存句法信息的论元抽取模型:由于触发词与论元存在一定的关联,因此将触发词抽取模型所获得的触发词位置特征与预训练模型BERT-wwm相融合作为文本表示,然后与BiLSTM拼接进行上下文语义信息提取。为了获得句子中每个语言单位成分之间的依存关系,构建关于突发事件的依存句法树,将构建的依存树转换为邻接矩阵,并将此矩阵与图注意力网络相融合。图注意力网络对句法邻接矩阵进行建模,然后学习文本中的依存句法,并通过注意力机制关注对论元抽取影响更重要的依存句法信息,最后通过CRF联合解码的方式进行结果标注。本文在上海大学语义智能研究院所构建的中文突发事件语料库(CEC)进行实验,将研究结果与当前先进的模型进行对比实验,并对模型各模块进行消融分析。实验结果表明,本文提出的触发词抽取模型,论元抽取模型的抽取结果在精确率,召回率以及F1值均优于对比模型,由此证明了本文模型的可行性和有效性。更多还原

【Abstract】 With the increase of Internet users,the data generated by the interaction between users gradually accumulates,and electronic text information presents an explosive development in the Internet.When emergencies occur,a lot of relevant event information will be published on the Internet.How to extract structured emergency information from massive texts has also become an important aspect of network public opinion analysis.There are two main problems in the current emergency event extraction methods: First,in the trigger word extraction method,the importance of the contextual semantic information of the text cannot be fully utilized for extraction.Second,in the argument extraction method,the dependent syntactic information in the sentence cannot be fully utilized for extraction.In view of the limitations of the current method,this thesis proposes a Chinese emergency event extraction algorithm based on deep learning.In order to facilitate the optimization of each module,the event extraction is carried out in a pipeline manner,and the event extraction is divided into triggers based on the multi-head attention mechanism.A word extraction model and an argument extraction model that incorporates syntactic information of dependencies.The main research contents are as follows:(1)A trigger word extraction model based on multi-head attention mechanism is proposed.The pre-trained model BERT-wwm is used as the text representation.The model is trained in the way of full word coverage for Chinese-specific grammatical expressions,and can learn the semantic information of Chinese words.In order to obtain the global features of the text,BiLSTM is used to extract the forward and backward semantic information of the text.The multi-head attention mechanism is introduced for semantic information extraction.The multi-head attention mechanism can enhance the more important features of trigger word extraction,weaken the irrelevant or unimportant features,make the feature distribution more reasonable,and can capture from different semantic spaces.Text features to enhance text expression.Considering the correlation between adjacent nodes,the results are marked by means of CRF joint decoding.(2)An argument extraction model that fuses dependent syntactic information is proposed: Since there is a certain relationship between trigger words and arguments,the location features of trigger words obtained by the trigger word extraction model and the pre-training model BERT-wwm are fused as text representations,and then combined with BiLSTM concatenation for contextual semantic information extraction.In order to obtain the dependencies between the components of each language unit in a sentence,a dependency syntax tree about emergent events is constructed,the constructed dependency tree is converted into an adjacency matrix,and this matrix is fused with a graph attention network.The graph attention network models the syntactic adjacency matrix,then learns the dependency syntax in the text,and pays attention to the dependency syntax information that is more important for argument extraction through the attention mechanism.Finally,the result is marked by CRF joint decoding.This thesis conducts experiments on the Chinese Emergency Corpus(CEC)constructed by the Institute of Semantic Intelligence of Shanghai University,compares the research results with the current advanced models,and conducts ablation analysis on each module of the model.The experimental results show that the extraction results of the event trigger word extraction model and event argument extraction model proposed in this thesis are better than the comparison model in terms of precision rate,recall rate and F1 value,which proves the feasibility and effectiveness of the model in this thesis.更多还原