Entity Linking for Wikidata using Large Language Models and Wikipedia Links

Entity Linking (EL), a task that maps named entities in text to corresponding entities in a knowledge base, has gained attention as a fundamental technology in knowledge processing and natural language processing. Conventional EL methods typically tokenize input text and utilize multiple features such as word embeddings and knowledge graph embeddings. Adapting these conventional EL methods to specific languages requires modifying language-dependent modules like tokenizers and word embedding models for the target language. In this study, we propose an EL method targeting Wikidata, based on Large Language Models (LLMs) and links from Wikidata to Wikipedia. Our method prompts LLMs to extract entity names from the input text and generate the corresponding Wikipedia URLs. Furthermore, it queries the Wikidata SPARQL endpoint to obtain Wikidata IDs from the Wikipedia URLs, outputting the entity names and their Wikidata IDs. This method can be applied to various languages by modifying the prompts. To evaluate, we compared the proposed method with conventional EL methods (PNEL and Japanese PNEL) on Japanese and English datasets from LC-QuAD2.0, SimpleQuestions, and WebQSP; using GPT-3.5, GPT-4, and Llama 2 as LLMs. The results showed that our method using GPT-4 outperformed conventional EL methods in recall and F-scores on datasets except for Japanese SimpleQuestions.