Nepali Dependency Parsing Using Transfer Learning

This paper presents a dependency parser developed for the low resource Nepali language along with the creation of annotated dataset conforming to Universal Dependencies treebank. The created dataset is used to train along with other languages to create a neural dependency parser based on graph-based parsing. Training is done in the state-of-the art graph-based neural architecture that makes use of multilingual BERT embeddings. Various experiments have been conducted varying the source training languages and dataset sizes. In zero-shot case for Nepali language the UAS and LAS scores obtained are 59.52 and 47.47 respectively, whereas for few-shot case the scores are 80.73 and 72.67 respectively. This work presents a good enough baseline as well as high quality data that can now be used for further research in the direction of dependency parsing for Nepali language.