[4G2-OS-24c] 日常生活知識とAI

[4G2-OS-24c-03] Analyzing Issues in Modifying English Entity Linking Models to Japanese Entity Linking Models for Large Knowledge Graphs

〇Yuki Sawamura^1,2, Motoki Yatsu¹, Takeshi Morita^1,2, Shusaku Egami², Takanori Ugai^2,3, Kenichiro Fukuda² (1. Aoyama Gakuin University, 2. National Institute of Advanced Industrial Science and Technology (AIST), 3. Fujitsu Ltd.)

Keywords:Knowledge Graph, Entity Linking, Wikidata

Entity linking (EL) has attracted attention as a fundamental technology for question answering and other applications. The state-of-the-art EL research is focused on English, and there is little research on Japanese EL. EL models are built using language models and knowledge graph embedding, and building Japanese EL models requires Japanese support for language models and knowledge graph embedding. In this study, we analyze the issues of Japanese adaptation of English EL models to Japanese, and construct a Japanese EL model by changing the language-dependent embedding in the English EL model PNEL (Pointer Network based Entity Linker) construction for Wikidata. The Japanese EL model was constructed by changing the language-dependent embedding. We translated the English datasets, WebQSP, SimpleQuestions, and LC-QuAD2, and analyzed issues in model construction from the perspective of embedding methods by comparing and evaluating Japanese and English EL models.