danielikenga/low-resource-factcheck

Fact-verification pipeline for low-resource Nigerian languages (Igbo, Yoruba, Hausa) using transformers and LLMs. # Low-Resource Retrieval-Augmented Fact Verification **Daniel Chibike Ikenga (ec25121)** MSc Data Science dissertation project investigating evidence-grounded misinformation detection for low-resource Nigerian languages. The project evaluates fact verification across Hausa, Igbo, and Yoruba using multilingual encoder models, instruction-tuned large language models, sparse retrieval, translation, few-shot prompting, adversarial evidence, and cross-experiment error analysis. ## Research Aim The project investigates how evidence availability, retrieval quality, model configuration, prompting strategy, and translation are associated with fact-verification performance under low-resource multilingual conditions. ## Languages - Hausa - Igbo - Yoruba ## Main Experimental Components - Claim-only classification - Gold-evidence verification - BM25 retrieval-augmented verification - Adversarial evidence evaluation - Model-scale comparison - Translation-based verification - Few-shot prompting experiments - Cross-system behavioural and error analysis ## Models and Methods - XLM-R-based multilingual classification - Qwen instruction-tuned language models - BM25 sparse retrieval - English translation of low-resource-language inputs - Few-shot prompting - Adversarial evidence conditions - Per-language and per-label evaluation - Individual-example transition analysis - Manual taxonomy of persistent hard errors ## Repository Structure - `data/` — dataset files and processed data - `notebooks/` — exploratory notebooks - `results/` — experimental outputs and analysis artefacts - `src/` — core source code - `requirements.txt` — Python dependencies ## Environment Setup Create and activate a virtual environment: `python -m venv .venv` `source .venv/bin/activate` Install dependencies: `pip install -r requirements.txt` ## Research Scope The repository supports a controlled empirical study of low-resource fact verification. Particular attention is given to the relationshi …