Elamur Research⌕
☰
home / papers / 2026 / Would LLMs be Good Historical Linguists and Chinese Dialect Learners?

Would LLMs be Good Historical Linguists and Chinese Dialect Learners?

Paper · 2026 · Yicheng Liu, Shumin Shi, Youchao Zhou et al.

Summary

LLMs handle Standard Chinese well but often invent pronunciations for low-resource Chinese dialects, largely because they treat Mandarin as a stand-in for all varieties. The authors test whether Middle Chinese, the shared historical ancestor of most modern dialects, can serve as an explicit intermediate layer for dialect phonology. They define two tasks: inducing conditional sound change rules from Middle Chinese to a dialect, written as executable Python functions (a variant of Sound Law Induction), and producing dialect-specific IPA for whole sentences (a variant of G2P). They built a dataset for Middle Chinese and 12 modern dialects by cleaning and aligning MCPDict and Polyphone, with special handling for polyphonic characters. Qwen3-8B was trained jointly with LoRA-based SFT on both tasks, followed by GRPO for rule induction only. Overall, the system outperformed DeepSeek-V3.2 and ChatGPT-5.2, though gains differed by dialect. The work shows that historical linguistic knowledge can make dialect modelling both more accurate and easier to interpret.

Key points

  • Data: The multi-source dataset covers Middle Chinese plus 12 dialects. It includes character-level correspondences, hand-built rule exemplars and sentence-level IPA, derived from MCPDict and Polyphone through an automated pipeline for polyphone normalization and alignment.
  • Method: Qwen3-8B uses LoRA SFT, then GRPO with a reward targeting the generated Python rules. A single-character tokenization strategy is also applied. Ablations show that removing GRPO, this tokenization or joint training each lowers performance.
  • Task 1 (rule induction): The full method achieves the best Score on most dialects, and its Accuracy ranks second only to ChatGPT-5.2. ChatGPT-5.2 tends to overfit by writing complex rules with special cases for irregular characters.
  • Task 2 (sentence IPA): Evaluation uses a modified Levenshtein distance over candidate pronunciations, plus a Panphon feature edit distance. Removing Middle Chinese features hurts results, especially on unseen characters. ChatGPT-5.2's edit distance exceeds 1 on most dialects; the exceptions are Guangzhou (Cantonese) and Harbin.
  • Limitations: The approach assumes regular, tree-like descent from Middle Chinese. It does not model tone sandhi, the split between literary and colloquial readings, loanwords, or contact between neighbouring dialects.

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

🎓Beijing Institute of TechnologyBeijing · China34 works · 3,652 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Yicheng Liu, Shumin Shi, Youchao Zhou and Xingchen Zhang. 2026. Would LLMs be Good Historical Linguists and Chinese Dialect Learners?. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.831

@inproceedings{liu2026would,
  title = {{Would LLMs be Good Historical Linguists and Chinese Dialect Learners?}},
  author = {Yicheng Liu, Shumin Shi, Youchao Zhou and Xingchen Zhang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.831},
  url = {https://doi.org/10.18653/v1/2026.acl-long.831}
}

References

28 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Thiago Costa Chacon and Johann-Mattis List. 2016. Improved computational models of sound change shed light on the history of the tukanoan languages. Journal of Language Relationship, 13(3-4):177–204. find
  2. Kalvin Chang, Chenxuan Cui, Youngmin Kim, and David R. Mortensen. 2022. WikiHan: A new comparative dataset for Chinese languages. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3563–3569, Gyeongju, Republic of Korea. International Committee on Computational L… find
  3. Kalvin Chang, Nathaniel Robinson, Anna Cai, Ting Chen, Annie Zhang, and David Mortensen. 2023. Automating sound change prediction for phylogenetic inference: A tukanoan case study. In Proceedings of the 4th Workshop on Computational Approaches to Historical Language Change, pages 129–142, Singapore.… find
  4. DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenhao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, and 245 others. 2025. Deepseek-v3.2: Pushing the frontier of open…
  5. Antonios Dimakis, John Pavlopoulos, and Antonios Anastasopoulos. 2025. Dialect normalization using large language models and morphological rules. In Findings of the Association for Computational Linguistics: ACL 2025, pages 23696–23714, Vienna, Austria. Association for Computational Linguistics. find
  6. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685.
  7. Zhige Huang, Haoan Jin, Mengyue Wu, and Kenny Q. Zhu. 2024. Automatic reconstruction of Ancient Chinese pronunciations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5689–5698, Miami, Florida, USA. Association for Computational Linguistics. find
  8. Jiyue Jiang, Pengan Chen, Liheng Chen, Sheng Wang, Qinghang Bao, Lingpeng Kong, Yu Li, and Chuan Wu. 2025. How well do LLMs handle Cantonese? benchmarking Cantonese capabilities of large language models. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4464–4505, Albuq… find
  9. Young Min Kim, Kalvin Chang, Chenxuan Cui, and David R. Mortensen. 2023. Transformed protoform reconstruction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 24–38, Toronto, Canada. Association for Computational Linguistics. find
  10. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. Preprint, arXiv:2309.06180.
  11. Jinggui Liang, Dung Vo, Yap Hong Xian, Hai Leong Chieu, Kian Ming A. Chai, Jing Jiang, and Lizi Liao. 2025. Colloquial singaporean English style transfer with fine-grained explainable control. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long…
  12. Fangru Lin, Shaoguang Mao, Emanuele La Malfa, Valentin Hofmann, Adrian de Wynter, Xun Wang, Si-Qing Chen, Michael J. Wooldridge, Janet B. Pierrehumbert, and Furu Wei. 2025. Assessing dialect fairness and robustness of large language models in reasoning tasks. In Proceedings of the 63rd Annual Meetin… find
  13. Yanchen Liu, William Held, and Diyi Yang. 2023. DADA: Dialect adaptation via dynamic aggregation of linguistic rules. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13776–13793, Singapore. Association for Computational Linguistics.
  14. Jiaming Luo. 2021. change discovery. find
  15. Automatic methods for sound Yun Mai. 2009. Introduction to Chinese Phonology. Jiangsu Education Press. In Chinese. find
  16. Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2021. Ab antiquo: Neural proto-language reconstruction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4460–4473, Online. Association for Comput… find
  17. David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin. 2016. PanPhon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3… find
  18. The COLING 2016 Organizing Committee. find
  19. Atharva Naik, Darsh Agrawal, Hong Sng, Clayton Marr, Kexun Zhang, Nathaniel Romney Robinson, Kalvin Chang, Rebecca Byrnes, Aravind Mysore, Carolyn Rose, and David R. Mortensen. 2025. Programming by example meets historical linguistics: A large language model based approach to sound law induction. In… find
  20. OpenAI. 2025. Chatgpt-5.2. https://chat.openai.com/. Accessed: December 2025. ↗
  21. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  22. Patrick Sims-Williams. 2018. Mechanising historical phonology. Transactions of the Philological Society, 116(3):555–573. find
  23. Li Wang. 1985. The History of Chinese Phonology. China Social Sciences Press. In Chinese. find
  24. Zhenghui Xiong, Zhenxing Zhang, Xing Huang, Bu Dao, and Jiayan Zou. 2012. Language Atlas of China: 2nd Edition. Volume on Chinese Dialects. Commercial Press. In Chinese. find
  25. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, Yuanjun Lv, Yongqi Wang, Dake Guo, He Wang, Linhan Ma, Pei Zhang, Xinyu Zhang, Hongkun Hao, Zishan Guo, and 19 others. 2025. Qwen3-omni technical report. Preprint, arXiv:2509.17765.
  26. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  27. Siheng Zhang, Xingjun Tan, Yanqiang Lei, Xianxi-ang Wang, Zhizhong Zhang, and Yuan Xie. 2023. Cvte-poly: A new benchmark for chinese polyphone disambiguation. In Proc. find
  28. Interspeech 2023, pages 5526–5530. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-would-llms-be-good-historical/