AN INSTANCE EMBEDDING METHOD FOR KNOWLEDGE GRAPHS ENTITY ALIGNMENT

Existing entity alignment techniques often require large amounts of labelled data, and are unable to encode multi-modal data simultaneously. Aiming at this problem, we propose an instance embedding method for knowledge graphs entity alignment. It captured the linked information by graph convolution networks and multi-order relation embedding. It constructed instantiation embedding based on relationship and attribute to form cross-language entity alignment matrix. Experimental results show that our scheme can eliminate data label requirements and obtain a general increase in the accuracy of evaluating indicator by about 20% compared with the existing methods, while it can get an effective hyperparameter sensitivity and consistency conflict robustness of aligning results, which provides support for cross language knowledge graph fusion.