表題番号：2024C-422
日付：2025/03/28
研究課題計算機の文章読解・生成能力向上に関する研究
| 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 enhance the capabilities of computational text understanding and generation, we conducted research involving the training of large language models (LLMs) on large-scale corpora and the development of application systems based on these models. Regarding LLM training, we focused on integrating knowledge graphs into LLMs, improving knowledge-fusion models using a Mixture of Experts (MoE) approach, automatic quiz generation, and validating the vision-and-language model LongCLIP. The developed application systems included a recognizer for "Honka-dori" (classical poetry allusions) based on waka embeddings, an integrated framework leveraging synthetic data for adapting LLMs to academic domains, and a music generation system employing retrieval-augmented generation with ABC notation. Evaluation studies involved the creation of Japanese datasets related to prompting and the validation of empathy annotations in dialogue contexts. These advancements significantly contribute to the field of computational text understanding and generation.