Research On Entity Alignment Of Multi-modal Knowledge Graphs For Monolingual And Cross-lingual Scenarios

In recent years,with the continuous development and application of multimodal knowledge graphs,more and more application scenarios require large and accurate multimodal knowledge graphs to support the development of applications.Therefore,entity alignment technology has become one of the current research hotspots to improve the accuracy and efficiency of multimodal knowledge graphs.Through entity alignment,different types of entities can be aligned in the corresponding knowledge graphs to improve the various entity information of multimodal knowledge graphs.However,there are still some problems with current entity alignment technology.The main work of this article is as follows:This paper proposes an enhanced mechanism-based monolingual multi-modal entity alignment method to address issues such as the lack of auxiliary modalities,ignoring the interaction of entity modalities,and the absence of feature fusion schemes in monolingual entity alignment.The proposed method adds a numerical modality to assist in multi-modal entity alignment,enhances entity relationship representation using encoded visual features,and assists in entity attribute feature representation through contrastive learning.An attention layer is added to dynamically allocate modality attention weights,forming a new feature joint embedding for entity alignment tasks to solve the above issues.Due to the diversity of knowledge in the real world and the multiple languages in which it is expressed,it is necessary to consider the various differences in grammar and expression between different languages.Considering the generally low efficiency of existing cross-lingual entity alignment methods,this paper proposes the addition of a cross-lingual entity description module based on the single-language entity alignment method.The description module can extract the core information from entity description sentences to more accurately identify and match entities.Additionally,by leveraging the multimodal enhancement approach proposed in single-language entity alignment work,the cross-lingual entity alignment efficiency can be improved through modal feature enhancement and noise reduction.The enhanced mechanism-based monolingual multimodal entity alignment method and the common mechanism-based cross-lingual entity alignment method based on descriptive information were tested on different standard datasets through relevant experiments.Both showed significant improvements compared to baseline models.The experimental results also demonstrated the feasibility and effectiveness of this paper’s two proposed multimodal entity alignment methods.