Skip to content

sarveshchaudhari/praTran

Repository files navigation

Maharashtri Prakrit to English Neural Machine Translation

This project implements a Neural Machine Translation (NMT) system to translate Maharashtri Prakrit, an ancient Indo-Aryan language, into English. The project features a Flask web interface with a model fine-tuned for this unique, low-resource language.


Project Overview

Model

The core of this project is the fine-tuned M2M100 multilingual model by Facebook AI, adapted specifically for Maharashtri Prakrit.

Key Features

  • Transformer Architecture: Leverages self-attention for robust translation.
  • Long-range Dependency Handling: Optimized for complex ancient sentence structures.
  • Transfer Learning: Improved accuracy for low-resource language scenarios.

Dataset

The custom dataset contains 1,474 parallel sentences from Gāhā Sattasaī, using data augmentation and transfer learning to address the data scarcity challenges.

Evaluation

The model was evaluated with:

  • BLEU score: 15.34
  • METEOR score: 0.47