【作者】 郭建华；

【导师】 吕青；

【作者基本信息】 太原理工大学 ， 控制科学与工程， 2025， 硕士

【摘要】 本体匹配作为知识化和智能化过程中的重要环节,在完善知识服务、智慧服务,提升系统的智能化水平中起着关键性的作用。本体是对知识中实体和实体间关系的正式规范化描述。不同领域不同专家学者定义本体时,受开发人员主观偏好和看法的影响,同一或者相似概念会被描述为不同形式,从而导致本体异构性问题,而本体匹配技术可以有效处理语义歧义以及本体异质。神经网络技术凭借深度特征学习以及高速非线性建模的能力成为解决本体匹配问题的关键方法,但仍面临一些问题与挑战,为此本文利用神经网络集成学习方法解决一些关键问题,主要工作如下:(1)针对单一的神经网络无法保证其在多种异构环境下的有效性,提出基于有序加权平均-层次分析(OWA-AHP)的自适应集成方法为匹配对构建一种更为精确的相似度特征。采用基于上下文的神经网络以及基于语法的神经网络分别构造相似度矩阵,从不同角度保留匹配对之间的相似度特征。为了进一步提高相似特征的质量,OWA-AHP技术用于集成多种相似度矩阵。最后提出相似度矩阵维持策略以保证匹配对的合理性。(2)针对多数神经网络方法忽略本体描述独特性研究的问题,提出一种自学习的串并联集成神经网络匹配架构(SPINN)查找对齐以提高匹配质量。为了减少网络的过拟合,采用分层抽样的方法,只选取本体中重要程度高的实体构建训练集。串联深度孪生卷积网络和不同的深度孪生循环网络研究不同粒度下的语义特征的独特性。通过基于结构和语义的对齐工程整合最终的匹配结果。(3)针对无标签的匹配任务,提出一种深度无监督学习匹配框架。该框架对文本的上下文信息进行学习,提取到抽象文本特征,以此在无标签的匹配任务中找到对齐。在自编码器的基础上,针对本体的多种描述,构建卷积模块并且和不同的循环神经网络串行连接实现特征降维。一种改进的基于双向长短期神经网络的注意力机制提取较好的抽象特征。最后,采用一种多主导的对齐集成策略将本体不同层次的对齐进行合并从而提高匹配质量。(4)针对多个测试案例上分别进行训练极其耗费时间的问题,提出一种基于深度域适应集成网络的本体匹配方法,利用迁移学习中域适应的思想,构建了深度域适应集成神经网络,大大节省了在整个测试集的训练时间。并且当新的测试集的数据较少时,源网络可以直接进行微调,提高匹配效率,为解决基于深度学习的本体匹配问题提供了新的思路。在国际本体匹配竞赛的案例中测试本文所提出的各种策略和方法。实验结果表明上述各种方法均展现了有效性,不仅提高了有监督学习模式下匹配对的质量,同时提高了无监督学习模式下本体对齐的质量,另外利用域适应的思想,大大节省了整个案例集的测试时间。本文的研究成果为实现知识融合,语义共享,数据互操作提供基础理论和方法。更多还原

【Abstract】 As an important link in the process of knowledge and intelligence,ontology matching plays a key role in improving knowledge service and wisdom service and enhancing the intelligence level of the system.ontology refers to the formal and standardized description of entities and inter-entity relations defined by different experts and scholars in different fields.When different experts and scholars in different fields define ontologies,the same or similar concepts will be described in different forms due to the subjective preferences and opinions of developers,resulting in the problem of ontology heterogeneity,ontology matching technique can effectively deal with semantic ambiguity and ontology heterogeneity.Neural network has become a key method to solve ontology matching problem with its deep feature learning and high-speed nonlinear modeling capabilities,it still faces some problems and challenges,so this paper uses the neural network integrated learning method to solve some key problems,the main work is as follows.(1)Aiming at the problem that a single neural network cannot guarantee its effectiveness in a variety of heterogeneous environments,an adaptive integration method based on ordered weighted average Hierarchical analysis(OWA-AHP)is proposed to construct a more accurate similarity feature for alignment.Context-based neural network and syntax-based neural network were used to construct similarity matrix respectively,and the similarity features between matching pairs were preserved from different perspectives.In order to further improve the quality of similar features,OWA-AHP technique is used to integrate multiple similarity matrices.Finally,the similarity matrix maintenance strategy is proposed to ensure the rationality of matching pairs.(2)Aiming at the problem of most neural network methods ignore the research on the uniqueness of ontology description,a self-learning series-parallel integrated neural network matching architecture(SPINN)is proposed to find alignment.In order to reduce the overfitting of the network,the hierarchical sampling method is adopted to construct the training set by selecting only the entities with high importance in the ontology.The series deep twinning convolutional networks and different deep twinning RNN study the uniqueness of semantic features at different granularity.The final matching results are integrated through structure and semantics based alignment engineering.(3)A deep unsupervised learning matching framework is proposed for unlabeled matching tasks.The framework learns the context information of text and extract abstract text features to find alignment.On the basis of autoencoder,the convolutional module is constructed for various ontology descriptions,and the feature dimension is reduced by serial connection with different recurrent neural networks.An improved bidirectional long short term neural network based attention mechanism extracts better abstract features.Finally,a multi-dominant alignment integration strategy is used to combine the alignment at different levels of ontology to improve the quality of matching.(4)Aiming at the problem that training on multiple test cases is extremely time-consuming,a deep domain adaptive ensemble network based ontology matching method is proposed.Using the idea of domain adaptive in transfer learning,deep domain adaptive ensemble neural network is constructed,which greatly saves the training time in the whole test set.In addition,when the data of the new test set is small,the source network can be directly fine-tuned,which can improve the matching efficiency and reduce the overfitting of the neural network,providing a new idea for solving the deep learning based ontology matching problem.The strategies and methods proposed in this paper are tested in cases in international ontology matching competitions.The experimental results show that all the above methods are effective,which not only improves the quality of matching pairs under supervised learning mode,but also improves the quality of ontology alignment under unsupervised learning mode.In addition,the idea of domain adaptation greatly saves the testing time of the whole case set.The research results of this paper provide basic theories and methods for knowledge fusion,semantic sharing and data interoperability.更多还原