表題番号：2025C-092
日付：2026/02/02
研究課題計算機の文章読解・生成能力の向上、評価に関する研究
| 研究者所属（当時） | 資格 | 氏名 | |
|---|---|---|---|
| （代表者） | 理工学術院 基幹理工学部 | 教授 | 河原 大輔 |
- 研究成果概要
- To improve computers’ text understanding and generation capabilities, we conducted research and development on the training and evaluation of large language models using large-scale text corpora, as well as on application systems built upon them. Regarding the training of large language models, we examined large language model training using synthetic data, constructed a large-scale Japanese vision–language model, and developed a highly efficient process reward model. Regarding evaluation, we built a safety evaluation benchmark for Japanese multi-turn dialogue, a fact-checking framework targeting procedural responses, and a meta-evaluation framework for evaluators of Japanese in-vehicle dialogue. Regarding application systems, we developed an automatic Japanese rap lyrics generation system that controls rhyme and content, as well as a smart home assistant using local language models. We believe that these research and development efforts have advanced research on text understanding and generation by one step.