Systematic Investigation of Strategies Tailored for Low-Resource Settings for Sanskrit Dependency Parsing

Existing state of the art approaches for Sanskrit Dependency Parsing (SDP), are hybrid in nature, and rely on a lexicon-driven shallow parser for linguistically motivated feature engineering. However, these methods fail to handle out of vocabulary (OOV) words, which limits their applicability in realistic scenarios. On the other hand, purely data-driven approaches do not match the performance of hybrid approaches due to the labelled data sparsity. Thus, in this work, we investigate the following question: How far can we push a purely data-driven approach using recently proposed strategies for low-resource settings? We experiment with five strategies, namely, data augmentation, sequential transfer learning, cross-lingual/mono-lingual pretraining, multi-task learning and self-training. Our proposed ensembled system outperforms the purely data-driven state of the art system by 2.8/3.9 points (Unlabelled Attachment Score (UAS)/Labelled Attachment Score (LAS)) absolute gain. Interestingly, it also supersedes the performance of the state of the art hybrid system by 1.2 points (UAS) absolute gain and shows comparable performance in terms of LAS. Code and data will be publicly available at: \url{https://github.com/Jivnesh/SanDP}.