Urdu-to-English Machine Translation using LSTM with Attention

Overview

This project implements an LSTM-based attention encoder-decoder model for machine translation from Urdu to English using PyTorch. The model is trained on a parallel dataset and evaluated using BLEU scores to measure translation quality.

Dataset

• Source: Provided parallel dataset (sentence-aligned).
• Preprocessing:
  • Merge .dev and .devtest files.
  • Shuffle and split into 70% training, 15% validation, and 15% test.

Model Architecture

• Encoder: Bi-directional LSTM processes Urdu input sequences.
• Attention Mechanism: Enhances focus on relevant words during decoding.
• Decoder: LSTM generates translated English text based on encoder outputs and attention.
• Implementation: Custom PyTorch model (no pre-built RNN libraries used).

Evaluation

• Train & Validation Metrics:
  • Loss curve visualization.
  • BLEU score tracking during training.
• Final Performance:
  • 0.46 BLEU score using Moses multi-bleu.perl on the test set.

🚀 This project demonstrates an end-to-end Urdu-to-English translation pipeline using deep learning and attention mechanisms!