Literature Review β Annotated Bibliography
Organized by theme. Papers marked π are downloaded as PDF under
research/papers/. All links verified 2026-08-22.
A. Arabic NLP & Information Retrieval Foundations
A1. π AraBERT: Transformer-based Model for Arabic Language Understanding
- Antoun, Baly, Hajj β OSACT/LREC 2020 Β· arXiv:2003.00104 Β·
research/papers/arabic-nlp/arabert-2020.pdf - Pre-trained BERT for Arabic; SOTA on most Arabic NLU tasks at publication; released AraBERTv0.2 with pre-segmentation variants.
- Relevance: The default Arabic encoder baseline; basis for Arabic sentence embeddings used in your dense retriever.
A2. π ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic
- Abdul-Mageed, Elmadany, Nagoudi β ACL 2021 Β· arXiv:2101.01785 Β·
research/papers/arabic-nlp/arbert-marbert-2021.pdf - Trained on 24B words of MSA + 20 dialects. MARBERT is the go-to model for dialectal Arabic.
- Relevance: Directly addresses RQ5 (dialect robustness); candidate embedding backbone for dialect-heavy queries.
A3. π GATE: General Arabic Text Embedding (Matryoshka + Hybrid Loss)
- Nacar, Koubaa et al. β 2025 Β· arXiv:2505.24581 Β·
research/papers/arabic-nlp/gate-arabic-embeddings-2025.pdf - State-of-the-art Arabic semantic textual similarity on MTEB; 135M params beat OpenAI ada-002 and e5-mistral-7B on Arabic STS; Matryoshka dims 768β64.
- Relevance: Strongest open Arabic embedding model β primary candidate for the dense/kNN side of your hybrid retrieval. Matryoshka property lets you trade latency vs accuracy.
A4. π Improving Stemming for Arabic IR / Light Stemming (classic line of work)
- Larkey, Ballesteros, Connell β SIGIR 2002 (ACM) (paywalled β cite only)
- Darwish & Oard, TREC-2002 CLIR experiments Β· π
research/papers/arabic-nlp/darwish-trec2002-arabic-light-stemming.pdf - Established that light stemming (strip affixes) + normalization beats root-based stemming for Arabic retrieval.
- Relevance: Justifies Elasticsearch's
arabicanalyzer design (light stemmer + normalization) that forms your lexical baseline. Cite when defending why BM25+analyzer is a strong baseline.
A5. π Normalization of Arabic Dialects into MSA using BERT and GPT-2
- Alnajjar, HΓ€mΓ€lΓ€inen β JDMDH 2024 Β· DOI Β·
research/papers/arabic-nlp/arabic-dialect-normalization-bert-gpt2-2024.pdf - Encoder-decoder models mapping dialectal sentences to MSA (~46% fully correct).
- Relevance: Alternative to LLM expansion: normalize the query instead of expanding it. Good "related approach" to contrast in RQ2.
B. Arabic-Centric LLMs (candidates for the expansion engine)
| Model | Org | Size | License | Paper |
|---|---|---|---|---|
| ALLaM | SDAIA (Saudi) | 7B public | Apache-2.0 | π arXiv:2407.15390 |
| Jais | Core42/G42 (UAE) | 13B/70B | Apache-2.0 | π arXiv:2308.16149 |
| AceGPT | KAUST/CUHKSZ | 7B/13B | Apache-2.0 | π arXiv:2309.12053 |
| Qwen2.5 | Alibaba | 0.5β72B | Apache-2.0 | Qwen repo |
| Commercial | OpenAI/Anthropic/Google | β | API | β |
All four PDFs under research/papers/arabic-nlp/. Also see:
B4. π A Survey of Large Language Models for Arabic Language and its Dialects
- Mashaabi, Al-Khalifa, Al-Khalifa β KSU 2024 Β· arXiv:2410.20238 Β·
research/papers/arabic-nlp/arabic-llms-survey-2024.pdf - Map of all Arabic LLMs by architecture and data.
- Relevance: Your related-work chapter skeleton for the "Arabic LLM" section; use it to justify model shortlist.
C. LLM-Based Query Expansion & Reformulation
C1. π Query Expansion by Prompting Large Language Models
- Jagerman, Zhuang, Qin, Wang, Bendersky (Google) Β· arXiv:2305.03653 Β·
research/papers/llm-query-expansion/query-expansion-prompting-llms-google-2023.pdf - Introduced PRF-style LLM expansion: prompt LLM to imagine hypothetical relevant docs ("query-to-document" expansion); large gains over T5-based expansion on BEIR.
- Relevance: The canonical method your strategy (i)/(iii) in RQ2 builds on; copy its prompting templates and adapt prompts to Arabic.
C2. π Corpus-Steered Query Expansion (CSQE)
- Lei, Cao, Zhou, Shen, Yates β EACL 2024 Β· PDF Β·
research/papers/llm-query-expansion/csqe-eacl2024.pdf - Uses LLM relevance-assessment to pick corpus sentences as grounded expansions; fixes hallucination/outdated-knowledge issues; training-free.
- Relevance: Template for keeping your Arabic expansions grounded in the product catalog (e.g., only suggest terms that exist in the catalog). Strong answer to "LLMs hallucinate synonyms" objection.
C3. π LLM-QE: Aligning LLMs with Ranking Preferences
- Yao et al. β NAACL 2025 Β· arXiv:2502.17057 Β·
research/papers/llm-query-expansion/llm-qe-naacl2025.pdf - RL alignment of the expansion LLM with retriever preferences; +8% over Contriever zero-shot.
- Relevance: Upper-bound/"advanced" variant if you have time for light fine-tuning; also its related-work section maps the whole QE-with-LLM field.
C4. π Exploring Best Practices of Query Expansion with LLMs
- Wang et al. Β· arXiv:2401.06311 Β·
research/papers/llm-query-expansion/best-practices-query-expansion-llms-2024.pdf - Systematic study: what to expand (keywords vs pseudo-docs), where to insert (query vs doc side), effect on sparse vs dense retrievers.
- Relevance: Directly informs your experiment matrix design (RQ2) β reuse their factor structure for Arabic.
C5. π Knowledge-Aware Query Expansion with LLMs
- Xia et al. β NAACL 2025 Β· PDF Β·
research/papers/llm-query-expansion/knowledge-aware-query-expansion-naacl2025.pdf - Augments LLM expansion with knowledge-graph relations for semi-structured (attribute-style) queries β exactly the shape of product queries.
- Relevance: Product catalogs ARE structured knowledge (category trees, attributes); motivates feeding category metadata into the expansion prompt.
C6. π LLM-based Query Expansion Fails for Unfamiliar and Ambiguous Queries
- Abe, Takeoka, Kato, Oyamada β SIGIR 2025 Β· DOI (not downloaded)
- Documents failure modes: unfamiliar jargon, ambiguous queries β bad expansions.
- Relevance: Use to frame limitations chapter; motivates fallback-to-original-query design (never replace, always OR-augment).
C7. π GenQREnsemble: Zero-shot LLM Ensemble Prompting for Generative QR
- Dhole, Agichtein β ECIR 2024 Β· Springer
- Ensemble multiple rewrite generations for robustness. Optional citation.
D. E-Commerce Product Search & Industrial Query Rewriting
D1. π Semantic Product Search (Amazon)
- Nigam et al. β KDD 2019 Β· arXiv:1907.00943 Β·
research/papers/ecommerce-search/semantic-product-search-amazon-kdd2019.pdf - Deep semantic matching between queries and products using behavioral data; deployed at Amazon.
- Relevance: Foundational industrial reference for dense product retrieval; defines the queryβproduct matching problem formally.
D2. π Shopping Queries Dataset (Amazon ESCI benchmark)
- Reddy et al. β KDD Cup 2022 Β· arXiv:2206.06588 Β·
research/papers/ecommerce-search/esci-shopping-queries-amazon-2022.pdf - ~130K hard queries, 2.6M human judgments, ESCI labels (Exact/Substitute/Complement/Irrelevant). Languages: EN/JA/ES only β no Arabic.
- Relevance: (a) Methodological template for building YOUR Arabic benchmark; (b) explicit evidence of the gap you fill; (c) multilingual translation of ESCI is a recognized shortcut for low-resource benchmarks.
D3. π WANDS: Dataset for Product Search Relevance Assessment (Wayfair)
- Chen et al. β ECIR 2022 Β· Paper page (PDF paywalled β cite only) Β· Dataset: MIT license
- 480 queries Γ 43K furniture products, 233K judgments, 3-grade labels (Exact/Partial/Irrelevant), published annotation guidelines.
- Relevance: Smaller and simpler than ESCI β better template for an annotation protocol you can actually execute during a master's timeline.
D4. π Large Language Model based Long-tail Query Rewriting in Taobao Search (BEQUE)
- Peng et al. β WWW 2024 (Taobao) Β· arXiv:2311.03758 Β·
research/papers/ecommerce-search/taobao-longtail-query-rewriting-llm-2024.pdf - First industrial fine-tuned-LLM query rewriter: multi-instruction SFT + offline feedback + objective alignment. Deployed Oct 2023; boosted long-tail GMV.
- Relevance: THE closest industrial analog to your thesis (minus Arabic). Structure your pipeline discussion around it; your novelty = language (Arabic) + hybrid open-source stack + systematic comparison.
D5. π CLE-QR: Query Rewriting in Taobao Search
- Li et al. β CIKM 2022 Β· ACM (not downloaded)
- Pre-LLM generation-era rewriter via contrastive retrieval from historical queries. Baseline lineage for D4.
D6. π MAAQR: Multi-Agent Adaptive Query Rewriting in Alipay Search
- SIGIR 2025 Β· ACM (not downloaded)
- Knowledge-enhanced fine-tuning + multi-agent rewriting, deployed Dec 2024. Shows current frontier = agentic rewriting.
D7. π Enhancement of E-commerce Sponsored Search Relevancy with LLM (Walmart)
- Rokon et al. β SIGIR eCom 2024 Β· PDF Β·
research/papers/ecommerce-search/walmart-sponsored-search-llm-2024.pdf - LoRA-fine-tuned LLaMA for ad relevance at Walmart; cost/efficiency argument for smaller adapted models.
- Relevance: Supports H4/RQ4 argument that small local LLMs can be production-viable.
D8. π Advancing Query Rewriting via Shopping Intent Learning (Amazon)
- Zhang et al. β SIGIR eCom 2022 Β· PDF Β·
research/papers/ecommerce-search/amazon-shopping-intent-query-rewriting-sigir-ecom-2022.pdf - Rewrites tail queries into head-like queries preserving shopping intent.
- Relevance: Defines "intent preservation" β an evaluation criterion for judging whether an Arabic expansion is faithful.
D9. π Improving E-commerce Search with Category-Aligned Retrieval
- Aliev β arXiv 2025 Β· arXiv:2510.21711 (not downloaded)
- Predict category from query embedding then boost within category; Top-3 category accuracy 43.8%β83.2%.
- Relevance: Cheap post-hoc boost you can add on top of fusion; good ablation idea.
E. Hybrid Retrieval & Fusion
E1. π Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods
- Cormack, Clarke, BΓΌttcher β SIGIR 2009 Β· PDF Β·
research/papers/hybrid-retrieval/rrf-cormack-sigir2009.pdf - RRF: score(d)=Ξ£ 1/(k+rank), k=60; no score calibration needed.
- Relevance: The fusion function of your hybrid layer; Elasticsearch implements it natively (
rrfretriever).
E2. π Sparse Lexical and Expansion Model for First Stage Ranking (SPLADE)
- Formal et al. β ICTIR 2021 Β· arXiv:2107.05720 Β·
research/papers/hybrid-retrieval/splade-formal-2021.pdf - Learned sparse representations with term expansion inside an inverted index.
- Relevance: Conceptual bridge: LLM expansion at query time β learned-sparse ideas, but model-free and Arabic-tunable without training. Mention in related work.
E3. π Blind Deconvolution / Weighted-Sum Hybrid Score Fusion
- Bruch et al. Β· arXiv:2210.11934 Β·
research/papers/hybrid-retrieval/weighted-sum-hybrid-bruch-2022.pdf - Convex combination of calibrated scores can beat RRF when well tuned.
- Relevance: Second fusion method for your ablation (RRF vs weighted sum) β cheap extra experiment row.
F. Arabic E-Commerce Specific Work (direct competitors/closest prior art)
F1. π Enhancing Arabic Search Functionality for E-Commerce Websites: Challenges and Solutions
- Yadav β IJLRP 2022 Β· PDF Β·
research/papers/arabic-nlp/arabic-ecommerce-search-challenges-ijlrp-2022.pdf - Survey-style: enumerates morphological complexity, diacritics, dialect variation, transliteration issues in Arab e-commerce; proposes NLP roadmap but no system, no evaluation.
- Relevance: Your problem-statement source #1 β cite for motivation; position your thesis as the measured follow-up this paper lacks.
F2. π SHEINfer: Implicit Product Category Inference from Arabic E-commerce Queries
- OSACT @ LREC 2026 Β· PDF Β·
research/papers/arabic-nlp/sheinfer-arabic-product-category-lrec-osact-2026.pdf - Recent (2026!) work on understanding Arabic shopping queries (category inference).
- Relevance: Proof the community is actively working on Arabic e-commerce NLP NOW β strengthens timeliness claim; potential comparison point.
F3. π prestoai/arabic-ecom-search-bench
- Hugging Face dataset Β· Link Β· 10Kβ100K rows, text-retrieval task, MSA + Libyan dialect tags.
- Relevance: Possible starting seed for your benchmark (RQ1 evaluation data); verify license/quality before relying on it.
G. Background Reading (cite lightly)
- BM25 foundations: Robertson & Zaragoza, The Probabilistic Relevance Framework (PDF)
- Sentence-BERT: Reimers & Gurevych 2019, arXiv:1908.10084
- ColBERT late interaction: Khattab & Zaharia 2020, arXiv:2004.12832
- DPSR (JD.com): Qiu et al. SIGIR 2020 β personalized semantic product retrieval
- Que2Search (Facebook): Yiqun Liu et al. KDD 2021 β query/product understanding <1.5ms P99
- Arabic stemmers survey: Alshalabi 2022, Arabic light-based stemmer using new rules, ScienceDirect
H. Where Master's Theses Fit
No openly accessible master's thesis was found combining Arabic + e-commerce + LLM expansion (searched academic repositories Aug 2026) β good news for novelty. Useful repositories to monitor:
- Mandumah Dissertations β largest Arabic thesis database (170K+ theses; check via university library access): https://mandumah.com
- AUC Scholarship (Egypt), KAUST repository, Dar Almandumah
- International: DiVA (Scandinavia), PoliTO/PoliMi theses, DART-Europe (dart-europe.org)
Summary Table β What Exists vs. What You Add
| Capability | English/Chinese | Arabic |
|---|---|---|
| E-commerce search benchmark | β ESCI, WANDS | β (only emerging HF bench) |
| Industrial LLM query rewriting | β Taobao, Alipay, Walmart | β |
| Systematic LLM-QE study | β Google, academia | β |
| Hybrid BM25+dense product search | β | β measured |
| Your thesis | β | β all three combined + benchmark artifact |