[1G5-OS-26b] OS-26

[1G5-OS-26b-03] GPT-based Entity Linking for Wikidata

〇Fumiya Mitsuji¹, Yuki Sawamura^1,2, Takeshi Morita^1,2 (1. Aoyama Gakuin University, 2. National Institute of Advanced Industrial Science and Technology)

[1G4-OS-26a] 日常生活知識とAI 16:00 〜 16:20 にて発表

Keywords:Entity Linking, Large Language Models, Wikidata, Generative Pretrained Transformer, Knowledge Graph

Entity linking (EL), a task associating named entities in text with entities in a knowledge base, has attracted attention as a fundamental technology for question answering and other applications. Most existing EL methods focus on English and may not support other languages or have poor performance. In this study, we propose an EL method for Japanese and English based on GPT, which has advanced language understanding and generalization capabilities. Our approach extracts entity names and generate corresponding Wikipedia URLs from EL target sentences by providing prompts to GPT-3.5 Turbo and GPT-4. Subsequently, we query Wikidata's SPARQL endpoint to obtain Wikidata IDs from Wikipedia URLs and outputs the sets of entity names and their Wikidata IDs. We compared our proposed method with a prior research method (PNEL) on LC-QuAD2.0, SimpleQuestions, and WebQSP datasets in Japanese and English. Results showed that our method outperformed PNEL on all datasets except Japanese SimpleQuestions.