Osmanoor/arabic-rag-query-enhancement

LLM-based query enhancement for Arabic retrieval on MIRACL Arabic (CSQE + hybrid sparse-dense fusion). University of Khartoum graduation project, 2026. # Arabic RAG Query Enhancement LLM-based query enhancement for Arabic information retrieval, evaluated on MIRACL Arabic (2,896 dev queries, 2.06M passages, Modern Standard Arabic). Undergraduate graduation project — Department of Electrical and Electronic Engineering, University of Khartoum, 2026. --- ## What this does Short Arabic queries are information-poor: they under-specify the information need and suffer vocabulary mismatch against the corpus. This project expands the query before retrieval using open-weight LLMs, and asks *where* in a hybrid sparse–dense pipeline that expansion actually helps. The pipeline was built in stages, each of which is a separate experiment: 1. **Baselines** — mDPR (dense) and BM25S (sparse), evaluated separately. 2. **Query2Doc** — the LLM writes a pseudo-answer document; query + pseudo-doc becomes the new query. Compared across 11 open-weight LLMs (2B–20B). 3. **Query repetition** — the original query is repeated β times inside the expanded query to counter BM25 term dilution. 4. **CSQE (Corpus-Steered Query Expansion)** — a first-pass BM25 retrieval feeds the LLM real corpus sentences, which are mixed with blind expansions (2 corpus + 2 blind, α = 4). 5. **Hybrid fusion** — BM25 and mDPR runs combined via Reciprocal Rank Fusion (RRF) and convex combination (CC). 6. **Asymmetric placement** — the final finding: apply expansion to the *sparse* retriever only. ## Headline results MIRACL Arabic dev set, 2,896 queries. | System | nDCG@10 | R@10 | R@100 | MRR | |---|---|---|---|---| | BM25S, no enhancement | 0.4621 | 0.5964 | 0.8577 | 0.4836 | | mDPR, no enhancement | 0.4993 | 0.6156 | 0.8407 | 0.5328 | | mDPR + Query2Doc (Aya Expanse 8B) | 0.6164 | 0.7256 | 0.9001 | 0.6493 | | BM25S + Query2Doc + repetition (β=2, Aya Expanse 8B) | 0.5855 | — | — | — | | Hybrid RRF (k=20), no enhancement | 0.6267 | 0.7597 | 0.9466 | 0.6517 | | BM25S + CSQE (Aya Expanse 8B) | 0.6157 | 0.7447 | 0.9422 | 0.6380 | | **CSQE → BM25 only, fused with …