Elamur Research⌕
☰
home / papers / 2026 / Stereotyped by Silence: How LLMs Erase Northeast Indian Languages Through Omission and Orthographic Corruption

Stereotyped by Silence: How LLMs Erase Northeast Indian Languages Through Omission and Orthographic Corruption

Paper · 2026 · Badal Nyalang

Summary

The paper argues that LLMs can harm Northeast Indian language communities by leaving them out entirely, not only through biased associations. It names this harm "stereotype-by-omission" and measures two forms of it. The first is entity invisibility. The author built NortheastNER by fine-tuning xlm-roberta-base on about 25,000 weakly labeled sentences, using gazetteers, tribal and biodiversity data, and six region-specific categories. Untuned encoders score zero on these entities, and IndicNER also failed on them in a qualitative check. NortheastNER reaches F1 = 0.964 on the development set, slightly above a MuRIL model trained the same way. The second is orthographic erasure. Five open LLMs were checked for how they handle the Garo middle dot (U+00B7) and the special Khasi letters ï and ñ. Four of them alter these meaning-bearing characters in 18.8–50% of cases. A custom multilingual tokenizer cuts token counts by 26–50%. Because results differ across models, the author concludes these failures can be fixed. The paper proposes cultural representation audits as a standard part of multilingual evaluation.

Key points

  • NortheastNER uses six categories instead of the PER/LOC/ORG schema: PLACES, TRIBES, FESTIVALS, TOURIST, FLORA and FAUNA. Its development-set F1 is 0.964, compared with 0.951 for MuRIL and 0.000 for untrained xlm-roberta-base. On authentic regional texts, F1 falls to 0.778, with precision 0.980 and recall 0.645.
  • Performance follows the amount of training data. PLACES has about 45,000 entries and reaches F1 = 0.966. FAUNA has only 15 examples and scores F1 = 0.000. TOURIST scores just 0.143.
  • In the character-preservation test, only Gemma-2-2B kept every special character intact. Corruption rates were 18.8% for Falcon3-3B, about 40% for both Nemotron-Mini-4B and Llama-3.2-3B, and 50% for Falcon-H1-3B.
  • The proposed audits would check three things: coverage of culturally specific entities, orthographic integrity, and tokenization efficiency compared with a language-specific reference tokenizer.
  • The study has several limitations:
    • Only Khasi and Garo are covered.
    • Closed-source models are not tested.
    • No validation with the affected communities has been done yet.
    • NortheastNER and its gazetteers are only promised for release upon acceptance.

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

 Fields, topics and concepts

📚 Shelves 6The 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 worksSocial Sciences · 1,772 worksSafety Research · 543 works

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

Cite

Cite as Badal Nyalang. 2026. Stereotyped by Silence: How LLMs Erase Northeast Indian Languages Through Omission and Orthographic Corruption. Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.stereacult-1.6

@inproceedings{nyalang2026stereotyped,
  title = {{Stereotyped by Silence: How LLMs Erase Northeast Indian Languages Through Omission and Orthographic Corruption}},
  author = {Badal Nyalang},
  booktitle = {Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.stereacult-1.6},
  url = {https://doi.org/10.18653/v1/2026.stereacult-1.6}
}

References

10 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of “bias” in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454–5476. Association for Computational Linguistics.
  2. Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vish Subramanian, and Partha Talukdar. 2021. MuRIL: Multilingual representations for Indian langua…
  3. Tyler A. Chang and Benjamin K. Bergen. 2024. Language model behavior: A comprehensive survey. Computational Linguistics, 50(1):293–350.
  4. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  5. Daniil Maksymenko and Oleksii Turuta. 2025. Tokenization efficiency of current foundational large language models for the Ukrainian language. Frontiers in Artificial Intelligence, 8:1538165. find
  6. Badal Nyalang. 2026. NE-BERT: A multilingual language model for nine Northeast Indian languages. In Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026), pages 1–12, Rabat, Morocco. Association for Computational Linguistics. find
  7. Abdellah El Mekki, Houdaifa Atou, Omer Nacar, Shady Shehata, and Muhammad Abdul-Mageed. 2025. NileChat: Towards linguistically diverse and culturally aware LLMs for local communities. In Vladyslav Radchenko and Nazarii Drushchak. 2025. Improving named entity recognition for low-resource languages us… find
  8. Phillip Rust, Jonas Pfeiffer, Ivan Vulić, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? On the monolingual performance of multilingual language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Join…
  9. Yan Tao, Olga Viberg, Ryan S. Baker, and René F. Kizilcec. 2024. Cultural bias and cultural alignment of large language models. PNAS Nexus, 3(9):pgae346. find
  10. Sunita Warjri, Partha Pakray, Saralin A. Lyngdoh, and Arnab Kumar Maji. 2021. Part-of-speech (POS) tagging using deep learning-based approaches on the designed Khasi POS corpus. ACM Transactions on Asian and Low-Resource Language Information Processing, 21(3). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nyalang-2026-stereotyped-silence-how-llms-erase/