Comparative Evaluation of Transformer-Based Nepali Language Models

Large pre-trained transformer models using self-supervised learning have achieved state-of-the-art performances in various NLP tasks. However, for low-resource language like Nepali, pre-training of monolingual models remains a problem due to lack of training data and well-designed and balanced benchmark datasets. Furthermore, several multilingual pre-trained models such as mBERT and XLM-RoBERTa have been released, but their performance remains unknown for Nepali language. We compared Nepali monolingual pre-trained transformer models with multilingual models to determine their performance using a Nepali text classification dataset as a downstream task based on different number of classes and data sizes, taking machine learning (ML) and deep learning (DL) algorithms as baselines. Under-representation of Nepali language in mBERT resulted in overall poor performance, but, XLM-RoBERTa, which has a larger vocabulary size, produced state-of-the-art performance which is relatively similar to that of Nepali DistilBERT and DeBERTa, which outperformed all of the baseline algorithms. Bi-LSTM and SVM from the baselines also performed very well in variety of settings. Moreover, to assess the cross-language knowledge transfer for the cases when mono-lingual models are not available, we also evaluated HindiRoBERTa, a monolingual Indian language model on Nepali text dataset. This research mainly contributes to the Nepali NLP community by creation of news classification dataset with 20 classes, with over 200,000 articles and performance evaluation of various pre-trained monolingual Nepali transformers with multilingual transformers, DL and ML algorithms.