# Enhancing Access to Kenyan Education PSAs through Kiswahili, Luo, and Maasai Translation
**United States International University – Africa**
**Contributors:** Shalyn Muita, Arnold Bophine, Jemimah Bochaberi, Chanteel Kimathi, Melvin Lebo
**Date:** August 6, 2025
---
## Project Overview
Public Service Announcements (PSAs) are a vital communication tool for education, civic awareness, and health. However, language barriers often prevent these messages from reaching all communities, especially in multilingual regions like Kenya.
This project aims to improve access to educational PSAs by translating them into Kiswahili, Luo, and Maasai — ensuring that **every voice is heard, and every message is understood**.
---
## Objectives
- Bridge the language gap in civic communication.
- Fine-tune machine translation models on domain-specific, low-resource data.
- Enable reliable, confidence-aware translation workflows.
- Support cultural preservation through localized language use.
---
## Methodology
### 1. **Data Collection & Preprocessing**
- Scraped and curated **1,100+ education-related PSAs**.
- Cleaned and validated the dataset:
- Removed duplicates and short/incomplete entries.
- Standardized date, format, language, and URL fields.
- Flagged anomalies for manual review.
### 2. **Initial Translation (Swahili)**
- Used **Google Translate API** to generate initial Swahili translations.
- Manually aligned and cleaned outputs for fine-tuning quality.
### 3. **Model Fine-Tuning**
- Fine-tuned **MarianMT (opus-mt-en-sw)** on English–Swahili PSA pairs.
- Training/Testing split: 90% / 10%
- Result: **BLEU score: 59.71** (after 1 epoch)
### 4. **Advanced Pipeline: NLLB-200**
- Developed a two-stage pipeline:
- English → Kiswahili → Luo using `facebook/nllb-200-distilled-600M`
- Integrated:
- Language tags (`eng_Latn`, `swh_Latn`, `luo_Latn`)
- Token-level **confidence scoring** using softmax probabilities
- Maasai (MAA) handled via custom civic domain dictionary. …