BayinganaEdwin/multilingual-health-qa-african-languages

# Multilingual Health Question Answering in Low-Resource African Languages **ALU ML Techniques I — Final Project** **Author:** Edwin Bayingana **Email:** e.bayingana@alustudent.com --- ## Overview This project addresses the Zindi competition *Multilingual Health Question Answering in Low-Resource African Languages*. The task is to build a retrieval system that answers health questions in five African languages — Amharic, Akan, Luganda, Swahili, and English — by retrieving the most semantically similar answer from a provided training corpus. No text generation or fine-tuning is involved; the system returns an existing answer verbatim. Ten experiments are conducted, progressing from a TF-IDF sparse retrieval baseline through dense semantic retrieval using three multilingual sentence transformers, hybrid scoring with per-language weight tuning, two-stage reranking, and MBR ensemble decoding. **Best result (Experiment 8):** Zindi ROUGE-1 = 0.5763, LLM Judge = 0.7401 — a 20.8% improvement over the TF-IDF baseline. --- ## Repository Structure ``` . ├── notebook_multilingual_health_qa.ipynb # Main Kaggle notebook — all 10 experiments ├── Submissions/ # Zindi submission CSV files and screenshots per experiment └── README.md ``` --- ## Experiments Summary | # | Approach | Val ROUGE-1 | Zindi ROUGE-1 | LLM Judge | |---|----------|-------------|---------------|-----------| | 1 | TF-IDF Baseline | 0.3927 | 0.4771 | 0.6469 | | 2 | mT5-base Zero-Shot | 0.0093 | 0.4771 | 0.6469 | | 3 | MPNet Semantic | 0.4354 | 0.5080 | 0.6884 | | 4 | MPNet Hybrid (tuned) | 0.4774 | 0.5404 | 0.7136 | | 5 | LaBSE Semantic | 0.4467 | 0.5340 | — | | 6 | LaBSE Hybrid (tuned) | 0.4840 | 0.5701 | 0.7063 | | 7 | E5-Large Semantic (buggy prefix) | 0.4250 | 0.4785 | 0.6721 | | 8 | **E5-Large Hybrid (tuned)** | **0.5068** | **0.5763** | **0.7401** | | 9 | LaBSE + TF-IDF Reranking | 0.4364 | 0.5044 | 0.6628 | | 10 | MBR Ensemble (LaBSE + E5) | 0.4414 | 0.4790 | 0.6724 | - …