farunawebservices/igala-english-nmt

Neural Machine Translation for Igala (low-resource Nigerian language) using fine-tuned mBERT with 3,253 parallel sentences # 🌐 Igala-English Neural Machine Translation Fine-tuned mBERT model for bidirectional translation between Igala (a low-resource Nigerian language) and English. ## 🎯 Overview This project addresses the critical need for NLP tools in underrepresented African languages. Igala, spoken by ~2 million people in Nigeria, has had virtually no machine translation systems until now. ## 🚀 Live Demo Try the translator: huggingface.co ## 📊 Dataset - **Size**: 3,253 parallel Igala-English sentence pairs - **Collection Method**: Field collection from native speakers - **Domain**: General conversation, cultural texts, educational materials - **Quality**: Human-verified translations ## 🛠️ Model Architecture - **Base Model**: `bert-base-multilingual-cased` (mBERT) - **Fine-tuning**: Sequence-to-sequence with encoder-decoder architecture - **Training**: 10 epochs, learning rate 5e-5 - **Vocabulary**: Extended with 500 Igala-specific tokens ## 📈 Performance Metrics | Direction | BLEU Score | Translation Confidence | |-----------|------------|------------------------| | Igala → English | 18.3 | 72% average | | English → Igala | 14.7 | 65% average | *Note: Low BLEU scores typical for low-resource languages* ## 🔍 Example Translations **Igala → English:** Input: "Ọma ẹdu la" Output: "Good morning" (Confidence: 89%) Input: "Ẹ́ nụ́ ọ́wá?" Output: "How are you?" (Confidence: 82%) **English → Igala:** Input: "I am learning Igala language" Output: "Mí ń kọ́ èdè Igala" (Confidence: 76%) ## 📦 Installation ```bash git clone github.com cd igala-english-nmt pip install -r requirements.txt # Download fine-tuned model weights python download_model.py 🚀 Usage from igala_nmt import IgalaTranslator # Initialize translator translator = IgalaTranslator() # Translate Igala to English english = translator.translate("Ọma ẹdu la", direction="ig-en") print(english) # "Good morning" # Translate …