For a Sanskrit parser project that I am collaborating on, we have been discussing and investigating different language models, and their applicability to parsing Sanskrit. I have been particularly interested in deep learning approaches for language modeling, such as Seq2Seq (+ attention), etc.

This git repository contains scripts and ipython notebooks with my experiments.

A building block in many of these deep learning approaches is the embedding of words in a vector space using word2vec or GloVe. This notebook contains some of my experiments with word2vec using the Digital Corpus of Sanskrit to investigate the feasibility of using word2vec on just root words (prAtipadikas/dhAtus) in Sanskrit.

The DCS database is quite small from a deep learning perspective (about 30 MB if we count just the root words), so it was unclear how good the results would be or what to expect. (Spoiler - I was pleasantly surprised by the quality of the results obtained for a first pass).