Multi-modal Entity Alignment Based on Multi-level Feature Fusion and Reinforcement Learning
To address the defects of failing to leverage multimodal information and the potential interaction effects between modalities,this study proposes a multimodal entity alignment technique.This approach aims to capitalize on the distinctive modal features of entities to identify comparable entities in disparate multimodal knowledge graphs.Firstly,different feature encoders are used to extract attribute,relation,image and graph structure repre-sentations,and numerical modalities are used to enrich entity semantic information.Secondly,in the feature fusion stage,cross-modal complementarity and relevance modelling are executed simultaneously on the grounds of compar-ative learning.Reinforcement learning is also implemented to enhance the model output and decrease the heterogene-ous disparities between the acquired joint embeddings and the actual modal embeddings.Finally,the cosine similarity between two entities'cosine similarity is analyzed to filter out candidate aligned entity pairs,which are then iteratively added to the alignment seed to direct the new entity alignment.Experimental results demonstrate the effectiveness of the proposed approach in the multimodal entity alignment task.
万方数据
维普
