Nepali Sentiment Analysis
Multilingual transformers benchmarked on Devanagari sentiment classification.
Overview
A comparative study of multilingual transformer architectures — mBERT, XLM-RoBERTa, MuRIL and NepBERTa — on Nepali sentiment classification, documented in research-paper format.
The problem
Multilingual models claim broad coverage, but tokenisation and transfer behaviour on Devanagari script vary widely and are rarely measured side by side for Nepali.
Approach
- Standardised preprocessing and evaluation so the four architectures were compared on equal footing.
- Analysed tokenisation behaviour on Devanagari and its effect on downstream accuracy.
- Examined cross-lingual transfer for a low-resource classification task.
- Carried out error analysis and documented methodology, results and limitations.
Outcome
A reproducible comparison showing where language-specific pretraining pays off against general multilingual models for Nepali.