Elamur Research⌕
☰
home / papers / 2026 / Constructing a Japanese Rap Lyric Generation Model with GRPO

Constructing a Japanese Rap Lyric Generation Model with GRPO

Paper · 2026 · Hayato Ogawa and Daisuke Kawahara

Summary

Rhyming is difficult for LLMs because it depends on pronunciation, which plain text does not show directly. Japanese makes this harder: kanji often have several readings depending on context. Copyright guidance from Japan's Agency for Cultural Affairs also discourages training on existing artists' lyrics.

The authors fine-tune Sarashina2.2-3B with GRPO to generate two-bar Japanese rap. The reward is computed explicitly. Using readings obtained from MeCab, it checks how well the final vowel and moraic-nasal strings of the first and second bars match within a fixed-length suffix. It also penalizes repetition and checks structural validity. The setup needs no rap lyric corpus, preference pairs or separately trained reward model. At inference it uses no rhyme dictionary, target-vowel specification or post-editing.

In human evaluation, average scores barely changed. However, more outputs reached moderate or high ratings on rhyme-related criteria, and some showed multi-word end rhymes. Output collapse also became more frequent. The work is a copyright-conscious proof of concept, and the trained model could generate scored candidates for later DPO or PPO training.

Key points

  • Method: GRPO with a rule-based end-rhyme reward based on vowel-sequence similarity, with MeCab used for readings. No existing rap lyrics are used for training.
  • Model: Only Sarashina2.2-3B was tested, limited to structured two-bar generation.
  • Human evaluation: GRPO increased the share of outputs scoring 3 or higher on rhyme criteria, although mean gains were limited.
  • Output collapse: Among the 100 evaluated outputs, 17 collapsed after GRPO versus 2 before. Collapse included broken tags, invalid Japanese and stray Romaji or Latin tokens.
  • Future work: The authors plan to add consonant similarity to the reward, following Kawahara (2007). They also plan to build preference data from reward-scored outputs for DPO or PPO.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Waseda UniversityTokyo · Japan20 works · 584 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Hayato Ogawa and Daisuke Kawahara. 2026. Constructing a Japanese Rap Lyric Generation Model with GRPO. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.27

@inproceedings{ogawa2026constructing,
  title = {{Constructing a Japanese Rap Lyric Generation Model with GRPO}},
  author = {Hayato Ogawa and Daisuke Kawahara},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.27},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.27}
}

References

22 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 4302–4310, Red Hook, NY, USA. Curran Associa… find
  2. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  3. Shigeto Kawahara. 2007. Half rhymes in japanese rap lyrics and knowledge of similarity. Journal of East Asian Linguistics, 16(2):113–144. find
  4. Taku Kudo, Kaoru Yamamoto, and Yuji Matsumoto. 2004. Applying conditional random fields to Japanese morphological analysis. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 230–237, Barcelona, Spain. Association for Computational Linguistics.
  5. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. find
  6. Eric Malmi, Pyry Takala, Hannu Toivonen, Tapani Raiko, and Aristides Gionis. 2016. Dopelearning: A computational approach to rap lyrics generation. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 195–204, New York, NY, USA. Associa… find
  7. Enrique Manjavacas, Mike Kestemont, and Folgert Karsdorp. 2019. Generation of hip-hop lyrics with hierarchical modeling and conditional templates. In Proceedings of the 12th International Conference on Natural Language Generation, pages 301–310, Tokyo, Japan. Association for Computational Linguistic… find
  8. Ryota Mibayashi, Daichi Takamura, and Hitomi Yanaka. 2026. Yomi-bench: A benchmark for evaluating japanese reading understanding in large language models. In Proceedings of the 32nd Annual Meeting of the Association for Natural Language Processing (NLP2026), pages 1929–1934. find
  9. Ryota Mibayashi, Takehiro Yamamoto, and Hiroaki Ohshima. 2025. Japanese rhyme generation based on mora similarity and generation probability. In Information Integration and Web Intelligence: 27th International Conference, IiWAS 2025, Matsue, Japan, December 8–10, 2025, Proceedings, page 95–111, Berl… find
  10. Ryota Mibayashi, Takehiro Yamamoto, Kosetsu Tsukuda, Kento Watanabe, Tomoyasu Nakano, Masataka Goto, and Hiroaki Ohshima. 2024. Response verse generation from reverse direction including rhymes in rap battles. Journal of Information Processing Society of Japan, 17(2):28–39. find
  11. Nikola I. Nikolov, Eric Malmi, Curtis Northcutt, and Loreto Parisi. 2020. Rapformer: Conditional rap lyrics generation with denoising autoencoders. In Proceedings of the 13th International Conference on Natural Language Generation, pages 360–373, Dublin, Ireland. Association for Computational Lingui… find
  12. Uraku Oda, Hayato Ogawa, and Daisuke Kawahara. 2025. Unsupervised learning in the style of rap using a large language model. In Proceedings of the 39th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI 2025), pages 2Win567– 2Win567. find
  13. Aitor Ormazabal, Mikel Artetxe, Manex Agirrezabal, Aitor Soroa, and Eneko Agirre. 2022. PoeLM: A meter- and rhyme-controllable language model for unsupervised poetry generation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 3655–3670, Abu Dhabi, United Arab Emirates… find
  14. Peter Potash, Alexey Romanov, and Anna Rumshisky. 2015. GhostWriter: Using an LSTM for automatic rap lyric generation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1919–1924, Lisbon, Portugal. Association for Computational Linguistics. find
  15. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741.
  16. Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. 2020. Leveraging pre-trained checkpoints for sequence generation tasks. Transactions of the Association for Computational Linguistics, 8:264–280.
  17. Richard J. Savery, Lisa Zahray, and Gil Weinberg. 2020. Shimon the rapper: A real-time system for human-robot interactive rap battles. In ICCC. find
  18. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  19. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  20. Copyright Division Subcommittee on Legal Systems. 2024. Approach to ai and copyright. find
  21. Dekai Wu and Karteek Addanki. 2015. Learning to rap battle with bilingual recursive neural networks. In Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI’15, page 2524–2530. AAAI Press. find
  22. Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ogawa-2026-constructing-japanese-rap-lyric-generation/