melakteshome/Bidirectional-Awngi-Amharic-Machine-Translation

Bidirectional Awngi (አዊኛ)–Amharic Machine Translation research # Bidirectional-Awngi-Amharic-Machine-Translation Bidirectional Awngi (አዊኛ)–Amharic Machine Translation research # Bidirectional Awngi–Amharic Machine Translation: A Deep Learning Approach This repository accompanies the paper: **"Bidirectional Awngi (አዊኛ)–Amharic Machine Translation: A Deep Learning Approach"** by Melak Teshome, Zewdu Ayinie, and Habtamu Abate (Wollo University, Ethiopia). --- ## 📖 Overview We present the first bidirectional neural machine translation (NMT) system for the low-resource Ethiopian languages **Awngi** and **Amharic**. We compiled a new parallel corpus (~15,000 sentence pairs) and trained Transformer-based models in both directions. | Direction | Model | BLEU | |------------|--------|------| | Awngi → Amharic | Transformer | **33.12** | | Amharic → Awngi | Transformer | **34.24** | --- ## 📂 Repository Structure --- ## ⚙️ Requirements - Python ≥ 3.8 - PyTorch ≥ 1.13 - OpenNMT-py or Fairseq - sentencepiece (for BPE tokenization) - sacrebleu (for evaluation) Install dependencies: ```bash pip install torch sentencepiece sacrebleu 🧰 Dataset The Awngi–Amharic parallel corpus consists of approximately 15,000 sentence pairs collected from: Amhara Media Corporation’s official Facebook page Awngi and Amharic school textbooks Awngi newspaper (ቺርቤዋ ጋዜጣ) Community-verified translations Data split: Training: 90% Testing: 10% Note: The dataset is available upon reasonable request from the authors. pgsql @article{teshome2025awngi_amharic_mt, title={Bidirectional Awngi--Amharic Machine Translation: A Deep Learning Approach}, author={Teshome, Melak and Ayinie, Zewdu and Abate, Habtamu}, year={2025}, institution={Wollo University}, note={Manuscript submitted to Natural Language Engineering} } 📫 Contact For questions or collaboration, please contact: Melak Teshome Department of Information Technology, Wollo University 📧 melak.teshome@wu.edu.et or melemelak@gmail.com --- Would you like me to **fill this README template with your …