SanskritShala: A Neural Sanskrit NLP Toolkit

Official code for the paper "SanskritShala: A Neural Sanskrit NLP Toolkit with Web-Based Interface for Pedagogical and Annotation Purposes". If you use this code please cite our paper.

Web-based tool

You can interact with our SanskritShala's web-based platform: Link. We encourage you to check our demo video to get familiar with our platform.

Neural Modules of SanskritShala for 4 NLP tasks

You may find more details of codebases in Neural Modules folder for word segementaion, morphological tagging, depedency parsing and compound type identification task.

Deploy SanskritShala on your local machine

First you need to install the individual modules on your machine as instructed in the above section. You need not have a GPU in oder to make these pretrained systems work on your local machine. You may find more details on how to deploy toolkit on your local machine in SanShala-Web folder.

EvalSan: Evaluation Toolkit for Sanskrit Embeddings

SanEval is a toolkit for evaluating the quality of Sanskrit embeddings. We assess their generalization power by using them as features on a broad and diverse set of tasks. We include a suite of 4 intrinsic tasks which evaluate on what linguistic properties are encoded in word embeddings. Our goal is to ease the study and the development of general-purpose fixed-size word representations for Sanskrit. You may find more details of codebases in EvalSan folder.

Evaluation tasks

Intrinsic tasks

• SanEval includes a series of Intrinsic tasks to evaluate what linguistic properties are encoded in your word embeddings.
• We use SLP1 transliteration scheme for our data. You can change it to another scheme using this code.

Task	Metric	#dev	#test
Relatedness	F-score	4.5k	9k
Similarity	Accuracy	na	3k
Categorization Syntactic	Purity	na	1.1k
Categorization Semantic	Purity	na	150
Analogy Syntactic	Accuracy	na	10k
Analogy Semantic	Accuracy	na	6.4k

Pretrained word embeddings

• You can download the pretrained models from this link. README.md is given for each model.
• Place the models folder in the parent directory path.
• Pretrained vectors can be downloaded from this link. Place this folder in EvalSan/evaluations/Intrinsic/ path. This vectors are being used in evaluation script.
• Our proposed LCM pretraining is available at EvalSan/LCM folder. For more details please visit this link.

Citation

If you use our tool, we'd appreciate if you cite our paper:

@misc{Sandhan_SanskritShala,
  doi = {10.48550/ARXIV.2302.09527},
  url = {https://arxiv.org/abs/2302.09527},
  author = {Sandhan, Jivnesh and Agarwal, Anshul and Behera, Laxmidhar and Sandhan, Tushar and Goyal, Pawan},
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {SanskritShala: A Neural Sanskrit NLP Toolkit with Web-Based Interface for Pedagogical and Annotation Purposes},
  publisher = {arXiv},
  year = {2023},
  copyright = {Creative Commons Attribution 4.0 International}
}

License

This project is licensed under the terms of the Apache license 2.0.

Acknowledgements

We'd like to say thanks to everyone who helped us make the different neural models for SanskritShala.