Parsers for Sanskrit / संस्कृतम्

NOTE: This project is still under development. Both over-generation (invalid forms/splits) and under-generation (missing valid forms/splits) are quite likely. Please see the Sanskrit Parser Stack section below for detailed status. Report any issues here.

Please feel free to ping us if you would like to collaborate on this project.

A web interface is available here - https://kmadathil.github.io/sanskrit_parser/ui/index.html

This project has been tested and developed using Python 2.7. A port to Python 3 has been completed, and everything should now work in both versions of Python.

pip install sanskrit_parser

• See generated sphynx docs.
• PS: Command line usage is also documented there.

Run:

sudo mkdir /var/www/.sanskrit_parser
sudo chmod a+rwx /var/www/.sanskrit_parser

• Generate docs: cd docs; make html

Stack of parsing tools

Sandhi splitting subroutine Input: Phoneme sequence and Phoneme number to split at Action: Perform a sandhi split at given input phoneme number Ouptut: left and right sequences (multiple options will be output). No semantic validation will be performed (up to higher levels)

Module that performs sandhi split/join and convenient rule definition is at lexical_analyzer/sandhi.py.

Rule definitions (human readable!) are at lexical_analyzer/sandhi_rules/*.txt

• From dhatu + lakAra + puruSha + vachana to pada and vice versa
• From prAtipadika + vibhakti + vachana to pada and vice versa
• Upasarga + dhAtu forms - forward and backwards
• nAmadhAtu forms
• Krt forms - forwards and backwards
• Taddhita forms - forwards and backwards

To be done.

However, we have a usable solution with inriaxmlwrapper + Prof. Gerard Huet's forms database to act as queriable form database. That gives us the bare minimum we need from Level 1, so Level 2 can work.

Sanskrit Sentence

• Traverse the sentence, splitting it (or not) at each location to determine all possible valid splits

• Traverse from left to right

• Using dynamic programming, assemble the results of all choices

  To split or not to split at each phoneme

  If split, all possible left/right combination of phonemes that can result

  Once split, check if the left section is a valid pada (use level 1 tools to pick pada type and tag morphologically)

  If left section is valid, proceed to split the right section

• At the end of this step, we will have all possible syntactically valid splits with morphological tags

All semantically valid sandhi split sequences

Module that performs sentence split is at lexical_analyzer/SanksritLexicalAnalyzer.py

Semantically valid sequence of tagged padas (output of Level 1)

• Assemble graphs of morphological constraints

  viseShaNa - viseShya

  karaka/vibhakti

  vachana/puruSha constraints on tiGantas and subantas

• Check validity of graphs

1. Is the input sequence a morphologically valid sentence?
2. Enhanced sequence of tagged padas, with karakas tagged, and a dependency graph associated

Early experimental version (simple sentences only) is at morphological_analyzer/SanskritMorphologicalAnalyzer.py

See: Grammar as a Foreign Language : Vinyals & Kaiser et. al. Google http://arxiv.org/abs/1412.7449

• Method: Seq2Seq Neural Network (n? layers)
• Input Embedding with word2vec (optional)

Sanskrit sentence

Sentence split into padas with tags

DCS corpus, converted by Vishvas Vasuki

Not begun