Nagamese is the lingua franca of Nagaland, yet it has almost no NLP resources: no large corpora, no standard benchmarks and no language models. NagaNLP is my attempt to change that. It is a paper on arXiv (2512.12537), and its datasets and models are released on Hugging Face.
Paper: NagaNLP: Bootstrapping NLP for Low-Resource Nagamese Creole with Human-in-the-Loop Synthetic Data. Agniva Maiti, Manya Pandey, Murari Mandal.
Method
- An LLM (Gemini), guided by expert prompts, generates a candidate Nagamese corpus.
- Native speakers correct and annotate it. Inter-annotator agreement (Cohen's κ) is 0.92 for POS and 0.88 for NER.
- The result is a 10,018-pair conversational dataset plus POS-tagged and NER-annotated corpora.
Results (as reported in the paper)
| Task | Model | Result |
|---|---|---|
| Part-of-speech tagging | XLM-RoBERTa-base, fine-tuned | 93.81% accuracy (0.90 macro-F1) |
| Named entity recognition | XLM-RoBERTa-base, fine-tuned | 0.75 macro-F1 |
| Conversation | NagaLLaMA (Llama-3.2-3B Instruct, LoRA) | perplexity 3.85 vs 96.76 few-shot |
Stack
Python, PyTorch, Hugging Face Transformers, LoRA fine-tuning, human annotation guidelines and agreement analysis.
Why it matters
The pipeline is meant to be reusable for any language that has speakers but no corpora: a model drafts the data and people check it. The same work powers NagaTranslate.