Bennie-Kibet/TrendLens

# A Low-Resource Machine Translation: English To Dholuo This project builds and evaluates a machine translation pipeline between English and Dholuo, with a focus on adapting the system for medical/telemedicine use. ## Problem Statement Dholuo-speaking patients — particularly rural and elderly patients managing chronic conditions like diabetes or arthritis — often struggle to communicate clearly with doctors on telemedicine platforms. Kenya's doctor-to-population ratio is below WHO guidance, making language-inclusive remote care valuable. No existing Kenyan telemedicine platform (e.g. BYON8) currently supports Dholuo, despite the language having millions of speakers. ## Objective: To build a multilingual English–Dholuo translation model, adapt it toward clinical phrasing, and produce a pipeline that could plug into a telemedicine chat workflow. ## Target audience: Dholuo-speaking patients, Kenyan telemedicine providers, and healthcare workers needing a communication bridge during remote consultations. ## Success criteria: BLEU and chrF score improvement over an untuned pretrained baseline, evaluated on a held-out general-domain test split and a separate medical-phrase test set. ## Dataset Source: * `luo_eng.csv` (English–Dholuo parallel corpus, originally from a Kaggle dataset) * 136,625 sentence pairs, 2 columns (`eng`, `luo`), no missing values, 11 duplicate rows (dropped) * English sentences average ~18.3 words (~90 characters); Dholuo sentences average ~19.1 words (~89 characters) — the two languages are closely length-aligned ## Project Structure / Workflow 1. **Business Understanding** — problem framing, target audience, success 2. **Data Understanding** — loading and inspecting the parallel corpus 3. **Exploratory Data Analysis** * Missing values and duplicates check * Sentence-length outlier detection (IQR, boxplots) * Class/length balance checks * Keyword frequency bar charts and word clouds (English vs. Dholuo) * Zipf's law plot (validates the c …