sunilregmi/nepali-lemmatizerV1-mt5-base

Nepali Neural Lemmatizer

This model is a part of the study "Evaluating Multilingual Transformer Models for Lemmatization in Nepali: A Low-Resource Case Study" presented at LREC-COLING 2026. It addresses the challenges of lemmatization in the morphologically rich and low-resource Nepali language by leveraging pre-trained multilingual Transformer models.

🚀 Model Description

This project frames Nepali lemmatization as a text-to-text generation problem. We fine-tuned several state-of-the-art architectures to map inflected Nepali words (e.g., खान्छु) to their canonical base forms (e.g., खानु).

Key Features:

• Architecture: Based on mT5-base, mT5-small, and mBART-large-50.
• Script: Optimized for the Devanagari script.
• Robustness: Handles Out-of-Vocabulary (OOV) words and spelling inconsistencies better than traditional rule-based or TRIE-based systems.

📊 Performance

Our experiments show that the mT5-base model achieved the highest precision, while mBART-large-50 demonstrated superior morphological coverage.

Model	CER ↓	Accuracy ↑	BLEU (Char) ↑	Morph. Coverage ↑
mT5-base	1.1%	96.1%	0.980	0.964
mBART-large-50	1.6%	96.0%	0.986	0.970
mT5-small	1.7%	95.2%	0.983	0.955

Downstream Impact:

The use of this lemmatizer significantly improves performance in other NLP tasks:

• Cross-lingual Alignment: Improved accuracy from 12.86% to 41.61% for Hindi-Nepali word alignment.
• Information Retrieval: Increased Mean Average Precision (MAP) from 0.71 to 0.90.

🛠 Usage

You can use this model through the Hugging Face transformers library:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "sunilregmi/nepali-lemmatizerV1-mt5-base" # Example HF path
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

def lemmatize(word):
    # Prompt format used during training
    text = f"lemmatize: {word}"
    inputs = tokenizer(text, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=32)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

print(lemmatize("खैरियतसहितका")) # Output: खैरियत

📂 Training Data

The model was trained on a curated corpus of 8,000 unique word-lemma pairs:

• Gold Standard: 5,000 manually verified pairs.
• Augmented Data: 3,000 samples generated based on common Nepali inflectional patterns and morphological rules.

📝 Citation

If you use this model or the research findings, please cite:

@inproceedings{regmi2026evaluating,
  title={Evaluating Multilingual Transformer Models for Lemmatization in Nepali: A Low-Resource Case Study},
  author={Regmi, Sunil and Dawadi, Sundeep and Bal, Bal Krishna},
  booktitle={Proceedings of the 31st International Conference on Computational Linguistics (LREC-COLING 2026)},
  year={2026}
}

🔗 Resources

• GitHub Repository: sunilRegmi-ai/Nepali-Lemmatizer
• Paper: LREC 2026 - Evaluating Multilingual Transformer Models for Lemmatization in Nepali