【作者】 杨静；

【导师】 杨天若（YANG T.Laurence）；

【作者基本信息】 华中科技大学 ， 计算机系统结构， 2022， 博士

【摘要】 知识的积淀与传承铸就了人类文明的璀璨,也将推动机器智能的创新发展。通过建立数据到知识库中实体、概念等的逐一映射,并利用知识要素之间的关联关系对事物进行深度理解,知识图谱弥合了人类智能与人工智能之间的语义鸿沟,赋予智能体逻辑推理、精准决策等能力,是机器由感知智能过渡到认知智能的必要驱动。知识推理是大规模知识图谱构建和知识驱动的上层应用的核心技术,然而现有知识推理方法在邻域结构信息捕获、复杂关系语义学习、时序相邻模式构建等方面存在挑战,因此论文从知识图谱自身关联复杂、时间依赖、动态演化等特性出发,依托张量在全面知识表达、潜在交互特征捕获、语义可解释性等方面的优势,在静态知识推理、时序知识推理、增量知识推理三个方面开展了一系列理论、方法、技术的研究。主要研究内容和创新成果如下:首先,针对知识图谱具有高度关系异构性,而现有图神经网络存在无法同时捕获关系语义以及局部结构信息的问题,在单知识图谱推理方面,研究了基于张量图神经网络的知识图谱补全方法:提出面向知识表示学习的张量图注意力网络,采用张量运算来建模实体、关系以及三元组之间的交互以突破矩阵运算带来的限制,并进一步利用张量塔克算子及其性质对补全网络进行压缩及约简;在跨知识图谱推理方面,研究了基于多关系图神经网络的实体及关系联合对齐方法:进一步提出了多关系图注意力机制来学习关系表示,在关系聚合的过程中为关联实体分配不同权重,并提出新的实体及关系联合对齐网络及全局评估指标。实验结果表明,与现有先进静态推理方法相比,提出模型最高可将推理准确性提升7.6%。其次,针对时序知识图谱中的实体分布具有稀疏性和可变性,而静态模型缺乏在时间上下文中捕获事实相关性的能力,在时序内推方面,研究了基于张量的时序交互嵌入方法:显式地讨论了时序知识图谱中实体、关系和时间之间的交互影响,提出交叉卷积以捕获不同四元组背景下的潜在交互模式,进一步采用张量神经网络保持交互信息结构,并从不同的角度提取有效特征以提升时序预测;在时序外推方面,基于张量图注意力网络研究了演进时序推理方法:面向并发事实,提出细粒度张量图注意力模型,以充分利用稀疏时序子图的信息,学习实体关系级注意力系数及自注意力系数,面向时序相邻事实,采用门控循环单元递归建模其顺序模式,使学习到的嵌入涵盖历史信息,并在时序和结构因素的双重作用下增强知识表示,提高知识演化学习效果。实验结果表明,与现有先进时序推理方法相比,提出的模型最高可将时序推理准确性提升5.9%。最后,在大数据不断生成以及知识提取技术不断提高的背景下,知识图谱中的事实随时间推移而增加和演化,针对现有方法在进行知识图谱推理时存在效率低下并缺乏可解释性的问题,研究基于布尔张量分解的增量知识推理方法:首先提出了增量布尔张量分解方法,包括知识图谱表示张量的因子更新和布尔特征合并算法,避免了大量重复计算并提高了分解效率;进一步,通过利用布尔张量分解捕获知识图谱潜在语义信息,基于因子重构张量进行增量知识推理,并结合布尔张量分解的可解释性,对推理结果进行合理解释。实验结果表明,与非增量推理方法相比提出模型可将推理效率提高10倍以上。论文提出的基于张量的知识推理方法可以赋能知识图谱突破自身局限,为人工智能向更高层次的认知智能形态迈进积蓄动能。更多还原

【Abstract】 The accumulation and inheritance of knowledge has created the brilliance of human civilization,and will also promote a new highland for the innovation and development of machine intelligence.By establishing the mapping of data to entities,concepts,etc.,and using the associations to deeply understand things,the knowledge graph bridges the semantic gap between human intelligence and artificial intelligence,and empowers the agents with logical reasoning,accurate decision-making,and other capabilities.Knowledge graph is an essential driver for machines to transition from perceptual intelligence to cognitive intelligence.Knowledge reasoning is the core technology for building large-scale knowledge graph and knowledge-enabled upper-layer applications.Existing knowledge reasoning methods face challenges in capturing neighborhood structural information,learning complex relational semantics,and modeling temporal sequential patterns.Starting from the characteristics of complex association,multi-source conflict,dynamic evolution,etc.,and relying on the advantages of tensors in fully knowledge expression,latent interaction features capture,and semantic interpretability,etc.,research on theories,methods and technologies are carried out in static,temporal,and incremental knowledge reasoning.The main innovations and research contents are as follows:Knowledge graph has high relational heterogeneity,while the traditional graph neural networks have the defects of capturing relational semantics and local structural information.In terms of single knowledge graph reasoning task,this thesis studies tensor graph neural network for knowledge graph completion.The tensor graph neural network for knowledge graph is proposed,which adopts tensor operations to model the interaction between entities,relations,and triples to break through the limitations brought by matrix operations.Besides,the Tucker operator and its properties are adopted to compress parameters and reduce computations for the network.In terms of cross-knowledge graph reasoning task,this thesis studies multi-relational graph attention networks for joint entity and relation alignment.Based on the proposed tensor graph attention network,a multi-relational graph attention mechanism is further proposed to assign different weights to associated entities in the process of relation learning.In addition,an effective joint entity and relation alignment network and global evaluation criteria are proposed.The experimental results show that compared with the existing advanced static inference methods,the proposed model can improve the reasoning accuracy by up to 7.6%.The distribution of entities in the temporal knowledge graph is sparse and variable,and the static models lack the ability to capture factual relevance in the temporal context.In terms of temporal interpolation,this thesis studies temporal interaction embedding based on tensors.The interaction among entities,relations and time in temporal knowledge graph is explicitly discussed.Cross-convolution is proposed to capture the potential interaction mode in different quadruple contexts.Tensor neural network is adopted to preserve the interaction information structure,and extract effective features from different perspectives to improve temporal prediction.In terms of temporal extrapolation,this thesis studies finegrained tensor graph attention network for evolutional temporal reasoning.For concurrent facts,fine-grained tensor graph attention model is proposed to learn entity,relation-level attention coefficients and self-attention coefficients,which can fully use the sparse temporal subgraph information.For temporally adjacent facts,Gated Recurrent Unit(GRU)is employed to recursively model the sequential patterns and make the embeddings cover historical information.Enhanced representations are learned under the dual action of historical and structural factors to improve the knowledge evolution learning.The experimental results show that compared with the existing advanced temporal inference methods,the proposed model can improve the accuracy of temporal reasoning by up to 5.9%.In the context of the continuous generation of big data and the improvement of knowledge extraction technology,the facts in the knowledge graph increase and evolve over time.Existing reasoning methods suffer from inefficiency and lack of interpretability.This thesis studies incremental knowledge reasoning based on Boolean tensor factorization.The incremental Boolean tensor factorization method is developed,and the factor update and the Boolean features merging algorithm of the knowledge graph representation tensor are proposed,which avoids repeated calculations and improves the decomposition efficiency.By using Boolean tensor factorization to capture the latent semantic information of knowledge graph,an incremental knowledge inference method based on factor reconstruction tensor is proposed.According to the interpretability of the Boolean tensor decomposition,the inference results are reasonably explained.Experimental results show that compared with the non-incremental reasoning method,the proposed model can improve the reasoning efficiency by more than 10 times.The tensor-based knowledge reasoning approaches proposed in this thesis can empower the knowledge graph to break through its own limitations,and accumulate momentum for artificial intelligence to move towards a higher level of cognitive intelligence.更多还原