Neural Machine Translation with Attention

Overview

This project implements a sequence-to-sequence (seq2seq) neural machine translation (NMT) model with attention. The model consists of an encoder-decoder architecture using bidirectional LSTMs, enhanced with an attention mechanism to improve translation accuracy.

Features

• Encoder-Decoder Architecture: Utilizes bidirectional LSTMs for encoding and a unidirectional LSTM for decoding.
• Attention Mechanism: Implements cross-attention to dynamically focus on relevant parts of the input sequence during translation.
• Minimum Bayes Risk (MBR) Decoding: Enhances translation quality by optimizing based on expected loss.

Installation

To run this project, install the required dependencies:

pip install numpy
pip install tensorflow

Usage

1. Prepare the dataset and preprocess input sequences.
2. Train the model using 'compile_and_train' function
3. Translate sentences using the 'mbr_decode' function

Folder Structure

├── por.txt/       # Dataset and preprocessing scripts
├── utils.py/      # helper functions
├── nmt.py/        # Implementation of encoder, decoder, and attention
├── README.md      # Project documentation