sanskrit_parser

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.

Try it out!

• A simple web interface is available at https://sanskrit-parser.appspot.com/
• Launch the example notebook on Binder -
• Launch the example notebook on Google colab -

Installation

This project has been tested and developed using Python 3.7 - 3.9. To install the package:

pip install sanskrit_parser

To enable statistical scoring based on DCS, please also install gensim and sentencepiece:

pip install gensim sentencepiece

See next section for some options if gensim installation fails, and you need the scoring feature.

Gensim installation: Alternate options if pip install fails

The scoring implementation in sanskrit_parser depends on gensim for scoring, which requires the capability to build C extensions for Python. If you have an appropriate C compiler for your system, gensim should be installed automatically during pip install. We have seen some cases where pip install is unable to install gensim on Windows, and the following instructions are for those situations.

On Windows, gensim typically requires the installation of Microsoft build tools for Visual studio 2019 as documented here. If you cannot, or do not want to install MS build tools to compile extensions, some alternate options are:

1. Install the pre-built Windows library from https://www.lfd.uci.edu/~gohlke/pythonlibs/. (Please follow the instructions on the website to install the dependencies first.)
2. Run your code in the cloud (either on Binder or Colab) - See links in the Try it out section
3. Use the REST API of sanskrit-parser.appspot.com documented here - https://sanskrit-parser.appspot.com/sanskrit_parser/docs. You can use the try it out option under the default version -> splits -> Try it out. It will show you the sample commands for CURL or the URL itself, as well as the response.

Usage

• For a tour of the basic features, check out the example notebook.
• For more detailed documentation, see generated sphynx docs.
• PS: Command line usage is also documented there.

Deploying REST API server

Run:

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

Contribution

• Generate docs: cd docs; make html

Sanskrit Parser Stack

Stack of parsing tools

Level 0

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

Current Status

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

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

This is not accessed standalone from the command line.

Level 1

• 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

Current Status

Bootstrapped using a lexical lookup module built from

1. inriaxmlwrapper + Prof. Gerard Huet's forms database
2. the sanskrit_data project, suitably wrapped

(Either or both of these can be enabled at runtime)

That gives us the minimum we need from Level 1, so Level 2 can work. As the generator sub-project matures, that will take over the role of this Level

Use sanskrit_parser tags on the command line to access this

Level 2

Input

Sanskrit Sentence

Action

• 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

Output

All semantically valid sandhi split sequences

Current Status

Module at parser/sandhi_analyer.py

Use sanskrit_parser sandhi on the command line

Level 3

Input

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

Action:

• Assemble graphs of morphological constraints

  viseShaNa - viseShya

  karaka/vibhakti

  vachana/puruSha constraints on tiGantas and subantas

• Check validity of graphs

Output

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

Current Status

Module at parser/vakya_analyer.py

Use sanskrit_parser vakya on the command line

Sanskrit Generator

Generate any valid sanskrit pada using Ashtadhyayi rules, plus vartikas where necessary.

Rules are input in a high level meta-language (currently yaml with imposed semantics - this may change), and the internal rule engine executes rules till a valid pada form is output. Input may be

1. prakriti + pratyaya
2. prakriti + sentence semantics

subantas of ajanta prAtipadikas are currently implemented. Other features are being rolled in.

Use sanskrit_generator on the command line

Seq2Seq based Sanskrit Parser

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)

Input

Sanskrit sentence

Output

Sentence split into padas with tags

Train/Test data

DCS corpus, converted by Vishvas Vasuki

Current Status

Not begun