SamAbr/PSA-MT-Ekegusii

# Fine-Tuning Neural Machine Translation Models for Kenyan Public Service Announcements Using a Synthetic Parallel Corpus > **United States International University–Africa** · Department of Data Science and Analytics · 2026 --- ## 👥 Team | Name | Role | |------|------| | Weldesenbet Zeray | Team Member | | Samuel Abrha | Team Member | | Hetal Kumbharana | Team Member | | Halima Mohammed | Team Member | | Peter Kidiga | Team Member | | Mitchelle Moraa | Team Member | **Supervisor:** Professor Edward Ombui --- ## 📌 Overview This project addresses the challenge of domain-specific machine translation for **low-resource Kenyan languages**. General-purpose MT systems struggle with the specialised style of Public Service Announcements (PSAs) because they are trained on broad multilingual data rather than domain-specific content. We built an end-to-end pipeline that: 1. **Synthesises** a large-scale English PSA corpus using a constrained LLM (GPT-5 mini via Azure OpenAI). 2. **Translates** the corpus into Kiswahili and Somali using NLLB-200. 3. **Fine-tunes** NLLB-200 and mBART-50 on the generated parallel corpus. 4. **Evaluates** translation quality using BLEU and chrF metrics. --- ## 📊 Datasets All generated datasets are available in the `output/` directory and committed to the repository: | File | Description | |------|-------------| | `output/english_psas.csv` | 50,318 validated English PSAs | | `output/psa_parallel_dataset.csv` | Final parallel corpus (English · Kiswahili · Somali · Luo) | | `output/bible_en_guz_swh.csv` | 30,863 English · Ekegusii · Kiswahili aligned triples (Bible + merged Storybooks) | | `output/lughayangu_sentences.csv` | 316 contemporary English-Ekegusii sentence pairs scraped from lughayangu.com | | `output/corpus_manifest.json` | Source translation IDs, counts, and filter statistics | --- ## 🏆 Results ### Fine-Tuned vs. Zero-Shot Translation Performance | Model | BLEU | chrF | |-------|------|------| | NLLB-200 Zero-shot | 30.8 …