Fine-grained part-of-speech tagging in Nepali text

Ingroj Shrestha, Shreeya Singh Dhakal

Abstract: Part-of-speech (POS) tagging is essential in Natural Language Processing (NLP) and text analysis applications. POS-tagging is a well-researched problem, but there have been limited efforts in fine-grained Nepali POS-tagging. Highly inflectional languages like Nepali encodes information like gender, person, number, mood, and aspect within their words. Accurate disambiguation of these inflected word forms is essential in Nepali text analysis. This work shows that neural network models are great at disambiguating fine-grained morphological information encoded by words in Nepali texts. We experiment with three neural network-based architectures: BiLSTM, BiGRU, and BiLSTM-CRF. Our results show that deep-learning based models can capture fine-grained morphological information encoded by Nepali words given a large enough corpus. We trained all the models in this work using two embeddings: pre-trained multi-lingual BERT and randomly initialized Bare embeddings. We found that training a randomly initialized Bare embedding is better than the ones trained using large pre-trained multi-lingual BERT embedding for downstream tasks in Nepali like POS tagging. Among the tested models, the BiLSTM-CRF with the Bare embedding performed the best and achieved a new state-of-the-art F1 score of 98.51% for fine-grained Nepali POS tagging.