Research On Matching Clues And Methods For Large-Scale Knowledge Fusion

Knowledge fusion is to realize semantic interaction and integration between heterogeneous knowledge bases.It achieves alignment and association between the concepts,attributes,and instances through discovering the semantic mapping between different knowledge bases.Ontology matching is one of the key technologies to solve the fusion of multi-source heterogeneous knowledge.In the past two decades,a variety of ontology matching methods have been proposed.However,there still exists some problems in these methods that require to be urgently explored.On the one hand,there has not yet existed systematical study and guide about the performance,applying scenarios and combination effects of the matching clues.Therefore,it is difficult for users to select suitable matching clues and apply them reasonably while matching ontologies.Effectively extracting matching clues and selecting a matching strategy that accord with the ontological characteristics is the key point of ontology matching.On the other hand,traditional matching methods rely on feature engineering and artificially determining the terminological,structural,and semantic information,which makes it difficult to design a general matching method with promising matching performance.Applying deep learning technology to achieve automatic selection and capture of various features is an effective supplement to the traditional matching methods.In view of above problems,the paper mainly focuses on the followings:(1)How to effectively extract and construct matching clues in the ontology.(2)Analyse the impact of different matching combination strategies for different matching tasks.(3)Discuss the application of deep learning to solve the problem of insufficient mining of potential semantics and improve the generality of the matching method.The main works are as follows:1.Empirically research the effectiveness of ontology matching clues.The atomic matching clues obtained in three dimensions,such as terminology,structure,and external resources,are systematically analyzed.And the corresponding matching strategies are constructed based on the matching clues of different dimensions to evaluate the performance of different matching clues.There are regularly abundant descriptions in the ontology.In different matching scenarios and tasks,it is necessary to rely on the characteristics of the ontology and utilize different strategies to discover the semantic mapping between entities.Empirical results show that different matching clues have different degrees of influence on matching performance.And it is obvious that not all matching clues can effectively improve matching performance for specific matching tasks.The conclusion of the study points out that for different knowledge fusion tasks,the task scenarios and data characteristics should be taken into account while extracting and selecting reasonable matching clues.2.Systematically explore the matching performance of different combination strategies.Based on the single-dimensional matcher built on matching clues from terminology,structure,and external resources,the performance of multiple strategy selections and combinations are systematically analyzed and verified.Specifically,it compares the performance of multi-dimensional matchers based on the comprehensive consideration of multiple ontology features.Then,an appropriate selection and combinations strategy is able to be customized for a given matching scene.Experimental results show that compared with the matching strategies based on single-dimension clues,both the matching strategy that combines the terminological clues and structural clues and the matching strategy that combines the terminological clues and external clues can achieve significant improvement.Simultaneously,it is found that the combination of too many matching clues may not equivalently improve the matching performance.Specifically,Matching strategies that combining all dimensions of matching clues,such as terminology,structure,and external resources,does not significantly outperform matching strategies that combine partial matching clues.The conclusion of the study points out that for the knowledge fusion task,effective selection and combination strategy of matching clues exert a significant impact on the ontology matching performance.3.Propose a hyperbolic graph attention network-based ontology matching method.The method utilizes knowledge embedding to jointly encode the terminological descriptions and structural features of the ontology to better represent the semantic information of elements in the ontology.Firstly,the ontology is embedded in hyperbolic space to learn the entity vector representation,so as to better capture the hierarchical characteristics of the tree-like ontology.Then,a graph attention mechanism is used to aggregate the multi-hop neighboring information of the central node to learn the local and global network structure of the ontology.After that,the hyperbolic representation of entities is learned and updated iteratively by hyperbolic aggregating with a graph-attention mechanism.Finally,the learned hyperbolic representation vectors are taken as the input of the matching module,and the ontological matching problem is converted into a binary classification problem to obtain the final matching results.The experiments demonstrate that,compared with other representation learning methods,our proposed matching method HOM-OM achieves the best performance.In addition,it can be found that reasonable integration of feature engineering can significantly improve the matching performance and achieve a much more competitive performance.