Dense retrieval for Arabic regulation questions
Fine-tuned an Arabic dense retriever on 1.2k regulation questions; Recall@5 rose from 0.61 (BM25) to 0.74. Short paper at a student workshop.
We built the question set with two law students, who wrote questions against 40 public regulations and marked the paragraph that answers each one. I fine-tuned a multilingual encoder with hard negatives mined from BM25 and compared it with BM25 and an untuned encoder on a split by regulation, so no document appears in both train and test. Most of the gain came on paraphrased questions; questions that reuse the exact wording still favour BM25, so the final system combines both. Code, splits and the evaluation script are public.
Role: Lead author · retrieval and evaluation
- Python
- PyTorch
- Natural language processing
- Arabic NLP
- Retrieval-augmented generation
- Hugging Face

