Cross language entity alignment based on knowledge augmentation of large language model

Cross-lingual entity alignment is crucial for the construction and maintenance of knowledge graphs. However, most of existing methods rely on ample labeled data, and the process of collecting this data is labor-intensive. To address this issue, we propose Cross Language Entity Alignment based on knowledge augmentation of Large Language Model (CLEA-LLM), which combines large language model with semi-supervised learning. Specifically, CLEA-LLM first utilizes the ChatGLM4-9B model to generate knowledge descriptions for entities and relations, which are processed by Sentence-BERT to produce coherent and deep semantic embedding vectors. Then, a joint alignment model is designed to integrate structural information and semantic information, enabling entity alignment through a bidirectional soft voting mechanism. Comparative experiments on three sub-datasets of DBP-15k showed that CLEA-LLM achieved Hits@1, Hits@10, and MRR scores of 79.8%, 95.4%, and 85.3%, respectively. The proposed method significantly improves the effectiveness of cross linguistic knowledge graph alignment tasks, especially in small-scale datasets.