【作者】 刘涛；

【导师】 于瑞国； 邵鹏；

【作者基本信息】 天津大学 ， 电子信息， 2023， 硕士

【摘要】 大数据时代,信息爆炸式增长引发了信息过载。如何从冗杂信息中抽取有效信息亟待研究。知识图谱作为结构化语义知识库,能有效地组织、存储和处理数据,受到业界广泛关注。然而,知识图谱普遍存在的数据稀疏问题限制了知识图谱的应用。因此,学者们提出知识图谱补全方法,并逐渐成为人工智能领域研究热点。现有知识图谱补全方法多基于静态知识图谱。相较于静态知识图谱,时序知识图谱引入时序信息,更符合真实世界发展规律,但由于其数据异质性和时间依赖性,其关系模式异常复杂,难以建模,时序知识图谱补全面临更大挑战。针对现有模型将关系建模为单个几何运算,无法充分建模时序知识图谱中复杂关系模式的问题,论文提出基于齐次变换结合平移旋转的时序知识图谱补全模型(Combination of Translation and Rotation Based on Special Euclidean Group for Temporal Knowledge Graph Completion,ComTR-SE),使用齐次变换结合平移和旋转,将实体表示为齐次坐标向量,将关系和时间戳表示为齐次变换矩阵向量。模型使用齐次变换矩阵构建包含平移、旋转和平移旋转复合变换的统一框架,对时序知识图谱中关键关系模式进行建模,并自适应地为每种关系选择最佳表示。论文在理论上证明ComTR-SE对时序知识图谱中各种关系模式的表达能力。并通过实验证明,模型ComTR-SE能够在多个公开数据集上取得有竞争力的效果。ComTR-SE模型通过平移旋转复合变换,可有效进行复杂关系模式建模。然而该模型将实体表示为齐次坐标向量,相较于关系和时间戳的齐次变换矩阵向量表示,对实体建模能力较弱,造成实体和关系、实体和时间戳表达的不一致性,影响模型表达能力。论文在ComTR-SE的基础上,进一步提出基于对偶四元数的平移旋转复合变换时序知识图谱补全模型(Combination of Translation and Rotation in Dual Quaternion Space for Temporal Knowledge Graph Completion,ComTR-DQ),引入对偶四元数表示法,将实体、关系和时间戳统一表示为对偶四元数向量。模型使用对偶四元数之间的乘法表示平移旋转复合变换,可以对时序知识图谱中关键关系模式进行建模。论文证明ComTR-DQ模型对时序知识图谱中各种关系模式的表达能力。实验结果表明,模型ComTR-DQ能够在多个公开数据集上取得最佳效果。更多还原

【Abstract】 In the era of big data,the explosive growth of information has caused information overload.It is urgent to study how to extract effective information from redundant infor-mation.Knowledge graph,as a structured semantic knowledge base,can effectively or-ganize,store and process data,and has attracted widespread attention from the industry.However,the commonly existing problem of data sparsity in knowledge graphs limits the large-scale application of knowledge graphs.Therefore,scholars propose knowl-edge graph completion methods,which have gradually become a research hotspot in artificial intelligence.Most existing researches are based on static knowledge graphs.Compared with static knowledge graphs,temporal knowledge graphs introduce tem-poral information and are more consistent with the real-world development laws,but due to its data heterogeneity and time dependence,the relation patterns in the temporal knowledge graph are exceptionally complex and difficult to model,and the temporal knowledge graph completion still has great challenges.To address the problem that existing models model relations as individual geomet-ric operations and cannot adequately model the complex relation patterns in the tem-poral knowledge graph,thesis proposes Combination of Translation and Rotation based on Special Euclidean Group for Temporal Knowledge Graph Completion(ComTR-SE).The model uses homogeneous transformations to combine translation and rotation,rep-resenting entities as homogeneous vectors,and representing relations and timestamps as homogeneous transformation matrix vectors.The model uses homogeneous trans-formation matrices to construct a unified framework containing translation,rotation,and translation-rotation composite transformations to model the key relation patterns in the temporal knowledge graph and adaptively select the best representation for each relation.Thesis theoretically proves the expression ability of ComTR-SE for various re-lation patterns in temporal knowledge graphs.Experimental results on multiple public temporal knowledge graph datasets demonstrate the superiority of the model.The ComTR-SE model can effectively model complex relation patterns by translation-rotation composite transformation.However,the model represents entities as homoge-neous vectors,which is weaker for entity modeling compared to the homogeneous trans-formation matrix vectors representation of relations and timestamps.It causes inconsis-tency in entity and relationship,entity and timestamp representations,which affects the model representation capability.To address this issue,thesis proposes Combination of Translation and Rotation in Dual Quaternion Space for Temporal Knowledge Graph Completion(ComTR-DQ)based on the ComTR-SE model.The model introduces a more expressive dual quaternion representation,which unifies entities,relations,and timestamps as dual quaternion vectors.The model uses multiplication between dual quaternions to represent translation-rotation composite transformations and can model the main relation patterns in temporal knowledge graphs.Thesis proves the expression ability of the ComTR-DQ model for various relation patterns in temporal knowledge graphs.Experimental results on multiple public temporal knowledge graph datasets demonstrate the superiority of the model.更多还原