Abstract
Searching for similar entities between diverse and heterogeneous data sources in knowledge graphs (KGs) has always been a major challenge. The advent of large language models (LLMs) has opened up new avenues for research and improvement of existing solutions. For example, fine-tuning an LLM was rapidly adopted due to its ability to specialize in specific tasks. This challenge notably involves capturing subtle similarities between the entities concerned, at the linguistic, syntactic, and semantic levels. In this article, we propose an approach to fine-tune GPT-2 and BERT language models to tackle the generalization of entity alignment problems on many different datasets by one single model. We also propose a protocol for entity alignment based on the Kolmogorov Arnold Network (KAN) to address the limitation of interpretability, redundancy, and computational cost of LLMs. The evaluations conducted in this study highlight the significant contribution of the fine-tuned GPT-2 language model, ensuring superior performance on entity alignment challenges compared to BERT and KAN.