SamAbr/public-service-anouncement-MT

Adding Ekegusii to NLLB-200 with transfer learning, then fine-tuning it for Kenyan public service announcements. 40.97 chrF2++ English into Ekegusii on held-out real announcements, against a 14.56 floor. USIU-Africa, School of Science and Technology, 2026. # Fine-Tuning Neural Machine Translation Models for Kenyan Public Service Announcements > **United States International University–Africa** · School of Science and Technology · Natural Language Processing · 2026 Adding **Ekegusii** to NLLB-200, a language the model was never trained on, so that Kenyan public service announcements can reach 2.7 million more speakers. --- ## 👥 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 --- ## 📌 What this is NLLB-200 supports 200 languages. Ekegusii, a Bantu language of about 2.7 million speakers in Kisii and Nyamira counties, is not one of them, and no amount of prompting will make a model produce a language it has no token for. So we added one. Using **transfer learning** we registered `guz_Latn` in NLLB-200, seeded its embedding from a related language the model already knows, and fine-tuned on **62,669 parallel sentence pairs** collected and built for this project. The result translates English and Kiswahili into Ekegusii. Two things are documented here, because they are the two things the project actually decided: 1. **Where the training data came from** and what had to be thrown away. 2. **The curriculum experiment**: whether teaching the model general Ekegusii first and public service register second beats teaching it both at once. ### Headline result chrF2++ on **real** Kenyan public service announcements the model never saw: | Direction | Stock NLLB-200 | **Our model** | Gain | |---|---:|---:|---:| | English into Ekegusii | 14.56 | **40.97** | +26.41, a 181% relative gain | | Kiswahili into Ekegusii | 14.13 | **39.61** | +25.48, a 180% relative gain | Stock NLLB-200 cannot produce Ekegusii at all. It was asked for the nearest language it supports, so its column is a **floor**, not …