【作者】 徐涛；

【导师】 赵峰；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2020， 硕士

【摘要】 知识图谱的概念由知识库衍生而来,数据实体间通过相互关系链接在一起。知识图谱技术旨在存储开放世界中实体及实体之间的复杂关联信息,能够改善现有知识库的数据查询准确度及搜索效率,在知识自动问答、推荐系统等领域有着广泛的应用价值。现有的知识图谱通常是不完善的,且数据关联稀疏,导致其在自动问答、智能推荐等应用系统上的表现非常糟糕。基于文本增强的知识图谱表示学习技术,充分利用文本数据丰富的语义信息,并将与知识库关联的文本信息进行融合,能够提高实体关系向量的语义解释性,并对知识图谱稀疏结构数据进行补全,提高知识图谱技术在智能系统中计算推理的准确度。为了能够利用知识图谱外部丰富的文本数据信息,对表示学习得到的知识图谱中实体关系结构向量进行语义增强,建立知识联合表示学习模型。利用翻译训练算法的思想学习得到知识图谱内部三元组结构的表示向量,针对知识图谱相关概念的文本描述信息,设计卷积神经网络来抽取句子中的可靠特征信息,采用合理的卷积核参数处理输出向量,将文本表示向量映射到与结构向量一致的嵌入空间。基于注意力机制对不同关系的特征可信度进行区分,根据每个文本与关系的相关程度分配权重参数,通过向量内积的计算方式进行语义组合,从而可以有效地获取知识图谱中关系关联文本嵌入向量。联合模型利用相关文本的表示向量对现有知识库中的实体关系结构向量进行增强表示学习,使知识表示模型的翻译向量更具语义解释性,并可以很好地运用知识图谱外部模态信息对现有知识库的稀疏领域知识进行计算补全。同时,模型借助二维卷积运算对实体和关系的联合表示向量进行处理,提取向量本身具有的非线性特征,增强隐式向量间交互能力的同时,拥有高效的参数利用效率,在一定程度上缓解了复杂关系数据建模的高复杂度问题。为验证计算模型的有效性,分别在FB15k、WN18和YAGO3-10数据集上与通用模型Trans E进行对比实验,在实体预测任务上预测准确度总体提升6%-20%,在三元组分类任务上准确度总体提升4%-12%,充分阐明了联合表示模型的有效性和可扩展性。更多还原

【Abstract】 The concept of knowledge graph is derived from the knowledge base,and the data entities are linked together through mutual relationships.Knowledge graph technology aims to store complex and related information between entities in the open world.It can improve the accuracy and efficiency of data query in the existing knowledge base.It has wide application value in the fields of automatic knowledge question answering and recommendation systems.The existing knowledge graph is usually imperfect,and the data association is sparse,resulting in its poor performance in application systems such as automatic question answering and intelligent recommendation.The textenhanced representation learning technology makes full use of the rich semantic information of text data,and merges the text information associated with the knowledge base,which can enhance the semantic representation of the entity relationship vector and complete the sparse structure data of the knowledge graph.It can improve the accuracy of knowledge inference calculation and inference in intelligent systems.In order to utilize the rich text data information outside the knowledge graph to semantically enhance the entity relationship structure vector in the knowledge graph obtained by representation learning,a knowledge joint representation learning model is established.Using the idea of translation training algorithm to learn to get the representation vector of the triple structure inside the knowledge graph.Aiming at the text description information of related concepts of knowledge graph,a convolutional neural network is designed to extract the reliable feature information in the sentence,the output vector is processed with reasonable convolution kernel parameters,and the text representation vector is mapped to the embedding space consistent with the structure vector.Based on the attention mechanism,distinguish the credibility of the features of different relationships,and assign weight parameters according to the degree of relevance of each text to the relationship.The semantic combination is carried out through the calculation of the inner product of the vector,so that the relation related text embedding vector in the knowledge graph can be effectively obtained.The joint model uses the representation vectors of related texts to perform enhanced representation learning on the entity relationship structure vectors in the existing knowledge base,so that the translation vectors of the knowledge representation model are more semantically interpretable.It can also use the external modal information of the knowledge graph to calculate and complete the sparse domain knowledge of the existing knowledge graph.At the same time,the model uses the 2-dimensional convolution operation to process the joint representation vector of entities and relationships,extract the non-linear characteristics of the vector itself,enhance the interaction between implicit vectors,and have efficient parameter utilization efficiency.It alleviates the high complexity of complex relational data modeling to a certain extent.In order to verify the validity of the calculation model,a comparison experiment with the general model Trans E is conducted on the FB15 k,WN18 and YAGO3-10 datasets respectively.The overall prediction accuracy on entity prediction task is improved by 6%-20%,and the accuracy on triple classification task is improved by 4%-12%,which fully clarifies the effectiveness and scalability of the joint representation model.更多还原