preetammukherjee/sanskrit_morph_prakriya

Vidyut-Prakriya Tinanta Dataset for Morphological Rendering

This dataset contains pairs of Sanskrit morphological metadata and their corresponding surface forms, generated and verified using the vidyut-prakriya library.

Dataset Structure

The dataset is provided in JSON Lines (.jsonl) format. Each line is a JSON object with the following fields:

• llm_input (string): A textual representation of the morphological metadata. This serves as the input for a sequence-to-sequence LLM tasked with morphological rendering. Example: "Dhātu: BU (BvAdi), Lakāra: la~w, Prayoga: kartari, Puruṣa: praTama, Vacana: eka"
• surface_form_vidyut (string): The Sanskrit surface form derived by vidyut-prakriya for the given metadata. Example: "Bavati"

Generation Process

1. Dhātu Lexicon: Dhātus (verb roots) are sourced from the dhatupatha.tsv provided with vidyut-prakriya (version 0.4.0 data).
2. Metadata Combination: For each dhātu, combinations of the following morphological features are generated:
   • Lakāra (tense/mood)
   • Prayoga (voice: kartari, karmani, bhave)
   • Puruṣa (person: prathama, madhyama, uttama)
   • Vacana (number: eka, dvi, bahu)
3. Derivation & Verification: The vidyut.prakriya.Vyakarana.derive() method is used to generate the surface form for each metadata combination. Only combinations that yield a valid surface form are included in the dataset.
4. LLM Input Format: The llm_input string is formatted to be human-readable and suitable for sequence-to-sequence models. Enum values (Lakāra, Prayoga, etc.) are represented by their SLP1 strings (e.g., la~w for Lakāra.Lat).

Intended Use

This dataset is primarily intended for training and evaluating language models on the task of Sanskrit morphological rendering (i.e., generating a surface form from its underlying grammatical specification).

It can also be used for:

• Analyzing the coverage of vidyut-prakriya.
• Studies in Sanskrit computational linguistics.

Project Context

This dataset was generated as part of a project inspired by Rohan Pandey's call for RL projects for Sanskrit. The goal is to use this data to train a model for morphological rendering and subsequently evaluate its impact on English-to-Sanskrit translation quality.

Citation

If you use this dataset, please consider citing the vidyut-prakriya library and/or this repository (once created).