Ilham-sy/kenyan-psa-machine-translation

NLLB-based machine translation system for translating public service announcements from English and Kiswahili into Ekegusii. # Multilingual PSA Machine Translation for Low-Resource Kenyan Languages ## Project Overview This project develops a **multilingual Neural Machine Translation (NMT) system for Public Service Announcements (PSAs)** with a focus on low-resource Kenyan languages. The system investigates translation involving: * **English** * **Kiswahili** * **Ekegusii** * **Kalenjin** The project uses **transfer learning** with pretrained multilingual models rather than training translation models from scratch. The main models explored are: * **NLLB-200 distilled 600M** * **mT5** The overall workflow covers data collection, cleaning, exploratory data analysis, model fine-tuning, evaluation, and deployment through simple **Gradio translation interfaces**. --- # Project Title **Development of a Multilingual Public Service Announcement (PSA) Dataset for Low-Resource Neural Machine Translation in Kenyan Languages** --- # Group Members | Name | Student ID | | ------------- | ---------- | | Harina Chohan | 666983 | | Susan Otieno | 670501 | | Ilham Mohamed | 670152 | | Kevin Korir | 670656 | | Yahya Mohamed | 669622 | --- # Project Motivation Public Service Announcements communicate important information in areas such as: * Health * Education * Agriculture * Security * Governance Much of this information is distributed mainly in English and Kiswahili. Kenyan indigenous languages such as **Ekegusii and Kalenjin** have significantly fewer digital resources and parallel datasets available for machine translation. This project therefore investigates whether pretrained multilingual language models can be adapted using **cross-lingual transfer learning** to support translation of public-service information into low-resource Kenyan languages. --- # Project Objectives The project aims to: 1. Build and curate a multilingual PSA dataset. 2. Prepare clean and aligned parallel translation datasets. 3. Develop Neural Machine Translation models usi …