README.md

metadata

title: My Translator by Ko Ko
emoji: 🌍
colorFrom: blue
colorTo: indigo
sdk: docker
pinned: false

Multilingual Neural Machine Translation (Project A3)

Developed by: Htut Ko Ko (st126010)

• 👉 Live App : huggingface.co/spaces/shadowsilence/burmese-english-translator

This project implements high-quality machine translation systems for multiple languages (Burmese, Thai, Chinese, Vietnamese, Hindi, Nepali, Urdu, Tagalog, Kazakh, Bengali, German) to English using two approaches:

1. Fine-Tuned NLLB-200: State-of-the-art multilingual model tailored for high-quality translation across all supported languages.
2. Transformer from Scratch: Educational implementation to demonstrate understanding of NMT architecture.

Experiments

Attention Mechanisms (Burmese-English)

I compared General (Dot Product) and Additive (Bahdanau) attention mechanisms using a Seq2Seq GRU model.

Attention Mechanism	Training Loss	Training PPL	Validation Loss	Validation PPL
General (Dot)	4.819	123.868	6.662	782.166
Additive (Bahdanau)	4.447	85.368	6.440	626.673

Observation: Additive Attention achieved lower validation perplexity, indicating better performance.

Demo

Folder Structure

• Burmese_English_NLLB.ipynb: (Recommended) Fine-Tuning NLLB for high-quality translation.
• Burmese_English_Transformer.ipynb: Transformer from Scratch implementation for Burmese-English.
• *_English_Transformer.ipynb: Transformer implementation for Foreign_language_for_AIT_students-English.
• Attention_Experiments.ipynb: Comparison of General vs. Additive Attention (Burmese-English).
• app/: Web Application folder.
  • app.py: Flask application supporting multiple languages.
  • nllb_model/: Fine-tuned NLLB model.

How to Run Locally

1. Requirements

Install dependencies:

cd app
pip install -r requirements.txt

2. Run the App

python app.py

Open http://localhost:5001.

Credits & Acknowledgements

This project respects the academic integrity and usage policies of the following resources:

• Dataset: Asian Language Treebank (ALT), Opus-100
• Base Model: NLLB-200 by Meta AI.
• Tokenization: SentencePiece by Google.