【作者】 李祥；

【导师】 徐博；

【作者基本信息】 大连理工大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 知识图谱是人工智能研究的重要组成部分,对于解决人工智能的可解释性问题具有重要意义。然而当前知识图谱的研究还主要集中于通用领域,垂直领域的知识图谱构建研究较少,所以本文对垂直领域的政务知识图谱构建展开研究。根据政务领域文本语料特点,本文提出了基于政务领域知识图谱的实体识别和实体关系抽取方法,旨在提高模型的关系抽取性能,为政务领域知识图谱的自动构建提供技术支持。针对政务文本中并列结构实体数量众多,实体分布集中等的特点,本文提出了基于BERT机制的政务实体预训练模型,并且结合双向长短期记忆网络(Bi LSTM)提出了政务实体识别模型。首先使用BERT预训练模型对政务文本进行预训练,得到初始化向量表示,然后利用Bi LSTM对政务文本实体进行识别,并引入注意力机制来确定输出向量的权重。实验结果表明:对于中文政务命名实体识别问题,本文算法比传统神经网络有更好的性能;对不同类型政务文本样本均取得较好的结果。针对现有政务文本关系抽取方法在卷积及池化操作时会损失局部特征信息的问题,本文提出了特征序列分割卷积神经网络的自适应池化方法。首先利用自适应池化和最大池化方法来提取政务文本的局部特征,然后结合Transformer模型来提取政务文本的整体特征,最后将每个子模块编码结果叠加处理,再进行交叉熵损失计算,得到文本局部、全局特征结合的输出向量。实验结果表明:对于中文政务文本关系抽取问题,本文算法比传统单一模型具有更好的抽取结果;对于并列、平行结构的政务实体关系有较好的抽取效果。本文改进了在政务文本数据集上命名实体识别和关系抽取的网络模型,构建了基于政务文本的知识图谱可视化系统,清楚的展示了各实体之间的关联信息。这项研究表明本文模型构建方法的有效性,体现了将自然语言处理和知识图谱技术应用于政务文本的潜力。更多还原

【Abstract】 Knowledge graph is an important part of artificial intelligence research,and it is of great significance to solve the interpretability problem of artificial intelligence.However,the current research on knowledge graphs mainly focuses on general fields,and there are few studies on the construction of knowledge graphs in vertical fields.So this paper studies the construction of government affairs knowledge graphs in vertical fields.According to the characteristics of the text corpus in the government affairs domain,this paper proposes an entity recognition and entity relationship extraction method based on the government affairs domain knowledge graph,aiming to improve the relationship extraction performance of the model and provide technical support for the automatic construction of the government affairs domain knowledge graph.Aiming at the characteristics of large number of parallel structure entity pairs and centralized distribution of entity pairs in government affairs texts,a pre-training model for government affairs entity recognition based on BERT mechanism is proposed,and a government affairs entity recognition model is proposed combined with bidirectional long short-term memory network.Firstly,the BERT pre-training model is used to pre-train the government affairs text to obtain the initialization vector,and then the Bi LSTM is used to extract the government affairs text entities,and the Attention mechanism is introduced to determine the weight of the output vector.The experimental results show that: for the Chinese government affairs named entity recognition problem,the algorithm in this paper has better performance than the traditional neural network;it achieves better results for different types of government affairs text samples.Aiming at the problem that the existing government text relation extraction methods will lose local feature information during convolution operation and pooling,this paper proposes an adaptive pooling method of feature sequence segmentation convolutional neural network.First,the adaptive pooling and maximum pooling methods are used to extract the local features of the government affairs text,and then the Transformer model is used to extract the overall features of the government affairs text.The output vector of the combination of local and global features.The experimental results show that: for the relationship extraction problem of Chinese government affairs text,the algorithm in this paper has better extraction results than the traditional single model;it has a better extraction effect for the entity relationship of parallel structure.This paper improves the network model of named entity recognition and relation extraction on the government affairs text dataset,and builds a knowledge graph visualization system based on government affairs text,which clearly shows the relationship information between entities.This study demonstrates the effectiveness of the above model building methods,embodying the potential of applying natural language processing and knowledge graph techniques to government texts.更多还原