Entity Alignment Through Joint Utilization of Multiple Pretrained Models for Attribution Relationship

• Published in: IEEE International Conference on Cyber Security and Cloud Computing (CSCloud) (Online) pp. 1 - 6
• Main Authors: Yujia, Liu, Tao, Xu, Zhehan, Su, Xinsong, Li, Kaipeng, Xue, Yuxin, Liu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.06.2024
• Subjects: Accuracy; BERT; Cloud computing; Computational modeling; Deep learning; entity alignment; GloVe; joint alignment; knowledge graph; Measurement; Semantics; Training
• ISSN: 2693-8928
• DOI: 10.1109/CSCloud62866.2024.00008

Most existing models for entity alignment tasks are based on a single pre-training mode. Through multiple experiments, it has been verified that a single pre-training model is no longer sufficient to meet the needs of current entity alignment tasks. Based on considerations of multiple pre-training models, this paper proposes an entity alignment method that utilizes multiple pre-training models. The purpose of this method is to address the limited accuracy in entity alignment tasks, with a particular focus on separately pre-training the embedding of entity attribute information and link relationship information for matching. To achieve the goal of improving entity alignment accuracy, this paper adopts joint alignment using pre-training with BERT and GloVe word embeddings. Experimental evaluations are conducted on the DBP15k dataset, and the results demonstrate that leveraging multiple pre-training models for attribute and relationship joint alignment can enhance model performance. The evaluation metrics of the model on three different language datasets, Hits@1 and MRR, are as follows: 0.899, 0.930, 0.964, and 0.932, 0.951, 0.975, respectively. These metrics indicate that the model performs well in aligning entities and outperforms current mainstream entity alignment methods.