Bonaparte003/African-Language-Health-QA-Challenge

Multilingual Health Q&A Retrieval-first semantic routing for low-resource African health question answering Zindi Challenge · Notebook · Experiment log --- ## Overview This repository documents an end-to-end solution for the **Multilingual Health Question Answering in Low-Resource African Languages** challenge hosted on Zindi in partnership with ITU and the HASH consortium. Given a health-related question in one of five African languages, the system returns an expert answer **in the same language**. The final approach is a **retrieval-first, per-subset semantic routing pipeline** that indexes curated question–answer pairs and returns verbatim in-corpus answers ranked by dense multilingual similarity, with sparse TF-IDF blending for low-resource subsets. Development followed a structured workflow: **EDA → baselines → semantic routing → seven targeted ablations** All runs are logged in `experiment_log.json` and scored locally with **ROUGE-1** and **ROUGE-L** F1 on `Val.csv`. --- ## Results | Metric | Value | |---|---| | **Public score** | 0.572643 | | **Rank** | 199 | | **Submissions** | 8 | | **Primary approach** | Semantic routing with expanded train+validation index (E16) | | Evaluation mode | Index | Best ROUGE-1 F1 | Experiment | |---|---|---|---| | Honest validation (no leakage) | `Train.csv` only | **0.474** | E18 | | Leaderboard proxy | `Train + Val` | **0.493** | E16b | --- ## Approach ```mermaid flowchart LR A["Raw CSVs Train / Val / Test"] B["EDA & Insights"] C["Preprocessing clean & language map"] D["Baselines TF-IDF & zero-shot LLM"] E["Semantic routing per-subset retrieval"] F["Answer selection verbatim from corpus"] G["Local evaluation ROUGE on Val"] H["Export predictions CSV submission"] A --> B --> C --> D --> E --> F --> G --> H ``` ### Design principles 1. **Evidence-driven modelling** — architectural choices trace to quantitative EDA findings. 2. **Baseline anchoring** — sparse and generative baselines define the p …