Publications

Publications

29 publications, 22 of them peer-reviewed, 10 as first author. Search by title, author or venue, filter by theme and type, and copy a BibTeX entry from the list or any paper’s page.

285 citations · h-index 9 (Google Scholar, September 2026) Atom feed

29 publications

2026 8 publications

  1. arXiv 2026Preprint

    An Empirical Study of Recent Large Language Models for Knowledge Graph to Text Generation on Standard Benchmarks

    Priyanka, R. Delbru, S. Bakkali, M. Coustaty, A. Doucet

    arXiv preprint (under review)

    Benchmarks recent LLMs against fine-tuned encoder-decoder models for knowledge-graph-to-text generation, with reference-based metrics and graph-grounded, reference-free LLM evaluation validated against human annotations. Prompted LLMs close much of the gap on reference-based metrics and produce fewer contradictions and triple omissions.

    arXiv Cite Language
  2. arXiv 2026Preprint

    Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made

    A. Merimi, S. Bakkali, A. Sakat

    arXiv preprint

    A timestamp theory of LLM decision validity and a decision ledger that realises it. Six timestamps define a decision, a freshness ceiling bounds what up-to-date evidence can guarantee at action time, and accuracy factorises into routing new information to affected decisions and judging it for the time of action. Controlled decision streams confirm four predictions across five LLMs.

    arXiv Cite Memory
  3. ICPR Workshops 2026Workshop

    BHCP-DOC: A Real-World Bilingual Dataset for Handwritten Clinical Prescription Detection and Recognition

    F. K. Bukhari, M. Coustaty, M. M. Luqman, S. Bakkali, M. G. Madden, N. I. Kajla

    Pattern Recognition. ICPR 2026 International Workshops (LNCS), pp. 448–463

    A real-world bilingual dataset of handwritten clinical prescriptions for text detection and recognition.

    DOI Cite Text
  4. IJDAR 2026Journal

    HSB-GDL: A Heterogeneous Subgraph Benchmark for Classification and Query Spotting

    M. Farhan, M. D. A. Awan, M. M. Luqman, S. Bakkali, M. Coustaty, N. I. Kajla

    International Journal on Document Analysis and Recognition (IJDAR)

    A benchmark of heterogeneous subgraphs for graph classification and query spotting.

    DOI Cite Graphs
  5. arXiv 2026Preprint

    Amnesia by Design, Memory By Necessity: Persistent State for Document Intelligence

    S. Bakkali, A. Merimi

    arXiv preprint

    A survey that names the statelessness bottleneck of Document AI, formalises persistent evidence-grounded document state, audits ten benchmarks against eight statefulness criteria, and derives a longitudinal harness with five counterfactual metrics: Experience Gain, Cost Efficiency, Memory Harm, Forgetting Fidelity and Coverage Retention.

    Interactive explorerarXiv Cite MemoryUnderstanding
  6. EAMT 2026Conference

    Terminology-Aware Retrieval-Augmented Knowledge Distillation for Biomedical Neural Machine Translation

    M. Zafar, S. Bakkali, R. Haque

    Proceedings of the 26th Annual Conference of the European Association for Machine Translation (EAMT), Volume 1, pp. 399–411

    Combines knowledge distillation with retrieval-augmented generation for few-shot French-to-English biomedical translation: the student learns to retrieve in-domain terminology from an external database, complementing the teacher, and matches or exceeds it.

    PDFLink Cite Language
  7. AICS 2025Conference

    Investigating Model Compression for Neural Machine Translation in the Biomedical Domain

    M. Zafar, S. Bakkali, R. Haque

    Irish Conference on Artificial Intelligence and Cognitive Science (AICS 2025)

    Studies knowledge distillation and quantisation for compressing biomedical neural machine translation models under low-resource, domain-specific conditions.

    arXivDOI Cite Language
  8. WACV Workshops 2026Workshop

    Cross-Lingual Transfer for Complex Scripts: A Benchmark on End-to-End Khmer Scene Text Spotting

    V. Nom, S. Keo, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pp. 729–738

    A benchmark for end-to-end Khmer scene text spotting that studies cross-lingual transfer to complex scripts.

    DOI Cite Text

2025 10 publications

  1. arXiv 2025Preprint

    MTRACE: Multilingual Retrieval-Augmented Generation for Temporally Diverse Text Corpora

    S. Bakkali, A. Mudet

    arXiv preprint

    A modular multilingual RAG pipeline evaluated on the French and English subsets of MIRACL: semantic query expansion with reciprocal rank fusion, a generation prompt that enforces strict grounding and explicit abstention, and ablations on entity extraction and retriever choice. It gives faithful answers for well-supported queries and abstains correctly on unanswerable ones.

    Version 1 (December 2025) was titled “Hybrid Retrieval-Augmented Generation for Robust Multilingual Document Question Answering”.

    arXiv Cite RetrievalLanguage
  2. ICDAR 2025Conference

    WildKhmerST: A Comprehensive Dataset and Benchmark for Khmer Scene Text Detection and Recognition in the Wild

    V. Nom, S. Keo, S. Bakkali, M. M. Luqman, M. Coustaty, M. Rusiñol, J.-M. Ogier

    Document Analysis and Recognition – ICDAR 2025 (LNCS), pp. 351–368

    A large-scale dataset of Khmer scene text captured in real-world environments, with 29,601 annotated text lines from 10,000 images, and a benchmark for detection and recognition.

    PDFDOIHAL Cite Text
  3. ICDAR Workshops 2025Workshop

    Cross-Lingual Learning for Low-Resource Khmer Scene Text Detection and Recognition

    V. Nom, S. Keo, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    Document Analysis and Recognition – ICDAR 2025 Workshops (LNCS), pp. 347–365

    Cross-lingual learning strategies for detecting and recognising Khmer scene text with little labelled data.

    PDFDOIHAL Cite Text
  4. ICDAR Workshops 2025Workshop

    Visual Text Generation in Khmer Language: Challenges and Trends with Diffusion Models

    S. Keo, V. Nom, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    Document Analysis and Recognition – ICDAR 2025 Workshops (LNCS), pp. 134–152

    Challenges and trends in generating Khmer visual text with diffusion models.

    PDFDOIHAL Cite Text
  5. Applied Sciences 2025Journal

    Confidence-Based Knowledge Distillation to Reduce Training Costs and Carbon Footprint for Low-Resource Neural Machine Translation

    M. Zafar, P. J. Wall, S. Bakkali, R. Haque

    Applied Sciences 15(14), 8091

    Confidence-based knowledge distillation strategies that cut the training cost and carbon footprint of low-resource neural machine translation.

    PDFDOI Cite Language
  6. arXiv 2025Preprint

    Evaluating the Impact of Khmer Font Types on Text Recognition

    V. Nom, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    arXiv preprint

    Measures how 19 Khmer font types affect OCR accuracy with Tesseract: Khmer, Odor MeanChey, Siemreap, Sithi Manuss and Battambang read well, while iSeth First, Bayon and Dangrek do poorly.

    arXiv Cite Text
  7. GbRPR 2025Workshop

    Fusion of GNN and GBDT Models for Graph and Node Classification

    M. Farhan, N. I. Kajla, M. D. A. Awan, M. Coustaty, M. M. Luqman, S. Bakkali

    Graph-Based Representations in Pattern Recognition (GbRPR 2025, LNCS), pp. 167–178

    Combines graph neural networks with gradient-boosted decision trees for graph and node classification.

    PDFDOIHAL Cite Graphs
  8. WACV Workshops 2025Workshop

    IDTrust: Deep Identity Document Quality Detection with Bandpass Filtering

    M. Al-Ghadi, J. Voerman, S. Bakkali, M. Coustaty, O. Lessard, N. Sidère

    IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), AI4MFDD, pp. 668–675

    Quality detection for identity document images using deep learning with bandpass filtering.

    PDFarXivDOI Cite Trust
  9. WACV Workshops 2025Workshop

    DocSum: Domain-Adaptive Pre-training for Document Abstractive Summarization

    P. P. M. Chau, S. Bakkali, A. Doucet

    IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), VisionDocs, pp. 1213–1222

    A domain-adaptive abstractive summarisation framework for administrative documents: pre-training on OCR-transcribed text and fine-tuning with question-answer pairs, to cope with domain terminology, OCR errors and scarce annotations.

    PDFarXivDOI Cite Understanding
  10. WACV 2025Conference

    GlobalDoc: A Cross-Modal Vision-Language Framework for Real-World Document Image Retrieval and Classification

    S. Bakkali, S. Biswas, Z. Ming, M. Coustaty, M. Rusiñol, O. Ramos Terrades, J. Lladós

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1436–1446

    A cross-modal transformer pre-trained with three self-supervised pretext tasks that unify language and vision, reducing reliance on OCR positional information. It introduces two industrial-style tasks: few-shot document image classification and content-based document image retrieval.

    PDFarXivDOI Cite UnderstandingRetrieval

2024 4 publications

  1. ACCV 2024Conference

    KhmerST: A Low-Resource Khmer Scene Text Detection and Recognition Benchmark

    V. Nom, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    Asian Conference on Computer Vision (ACCV 2024, LNCS), pp. 383–399

    A benchmark for detecting and recognising scene text in Khmer, a low-resource script with complex character structure.

    PDFarXivDOI Cite Text
  2. ICDAR 2024Conference

    Multimodal Adaptive Inference for Document Image Classification with Anytime Early Exiting

    O. Hamed, S. Bakkali, M. B. Blaschko, M.-F. Moens, J. Van Landeghem

    Document Analysis and Recognition – ICDAR 2024 (LNCS), pp. 270–286

    The first multimodal early-exit design for visually rich document understanding: training strategies, exit types and placements that reach a Pareto-optimal trade-off, cutting latency by over 20% while fully retaining baseline accuracy.

    arXivDOI Cite Understanding
  3. COMPSAC 2024Conference

    LLMChain: Blockchain-Based Reputation System for Sharing and Evaluating Large Language Models

    M. A. Bouchiha, Q. Telnoff, S. Bakkali, R. Champagnat, M. Rabah, M. Coustaty, Y. Ghamri-Doudane

    IEEE 48th Annual Computers, Software, and Applications Conference (COMPSAC), pp. 439–448

    A decentralised, blockchain-based reputation system that combines automatic evaluation with human feedback to assign contextual reputation scores to large language models.

    arXivDOI Cite Trust
  4. ACET 2024Conference

    State-of-the-Art Khmer Text Recognition Using Deep Learning Models

    S. Keo, M. Coustaty, S. Bakkali, et al.

    ASEAN Conference on Emerging Technologies (ACET)

    Deep learning models for Khmer text recognition.

    PDFHAL Cite Text

2023 2 publications

  1. Pattern Recognition 2023Journal

    VLCDoC: Vision-Language Contrastive Pre-Training Model for Cross-Modal Document Classification

    S. Bakkali, Z. Ming, M. Coustaty, M. Rusiñol, O. Ramos Terrades

    Pattern Recognition, vol. 139, 109419

    Learns cross-modal document representations from language and vision by aligning intra- and inter-modality relationships with a contrastive objective, instead of merging features into one joint space; effective on both small- and large-scale document classification datasets.

    arXivDOI Cite Understanding
  2. SSRN 2023Preprint

    TransferDoc: A Self-Supervised Transferable Document Representation Learning Model Unifying Vision and Language

    S. Bakkali, S. Biswas, Z. Ming, M. Coustaty, M. Rusiñol, O. Ramos Terrades, J. Lladós

    SSRN preprint

    A self-supervised, transferable document representation model that unifies vision and language.

    DOILink Cite Understanding

2022 1 publication

  1. PhD thesis 2022Thesis

    Multimodal Document Understanding with Unified Vision and Language Cross-Modal Learning

    S. Bakkali

    PhD thesis, La Rochelle Université

    Unified vision-and-language cross-modal learning for multimodal document understanding, prepared at L3i, La Rochelle Université, with a European Doctorate label (Computer Vision Center, Universitat Autònoma de Barcelona).

    PDFHAL Cite Understanding

2021 1 publication

  1. IJDAR 2021Journal

    EAML: Ensemble Self-Attention-Based Mutual Learning Network for Document Image Classification

    S. Bakkali, Z. Ming, M. Coustaty, M. Rusiñol

    International Journal on Document Analysis and Recognition (IJDAR) 24(3), pp. 251–268

    A self-attention fusion module inside an ensemble network learns image and text features together, and mutual learning transfers positive knowledge between the two modalities through a truncated Kullback-Leibler regulariser.

    arXivDOI Cite Understanding

2020 2 publications

  1. ICIP 2020Conference

    Cross-Modal Deep Networks for Document Image Classification

    S. Bakkali, Z. Ming, M. Coustaty, M. Rusiñol

    IEEE International Conference on Image Processing (ICIP), pp. 2556–2560

    A cross-modal network that captures both the textual content and the visual information of document images, with NASNet-Large and BERT features learned jointly; it outperforms single-modal methods.

    PDFDOI Cite Understanding
  2. CVPR Workshops 2020Workshop

    Visual and Textual Deep Feature Fusion for Document Image Classification

    S. Bakkali, Z. Ming, M. Coustaty, M. Rusiñol

    IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 562–563

    A two-stream architecture that jointly learns visual features of document images and embeddings of their text, with an investigation of widely used backbones and word-embedding procedures.

    PDFDOI Cite Understanding

2019 1 publication

  1. ICDAR Workshops 2019Workshop

    Face Detection in Camera Captured Images of Identity Documents Under Challenging Conditions

    S. Bakkali, M. M. Luqman, Z. Ming, J.-C. Burie

    International Conference on Document Analysis and Recognition Workshops (ICDARW), CBDAR 2019, pp. 55–60

    Evaluates three face detectors (Cascade-CNN, MTCNN and PCN) on camera-captured identity documents from MIDV-500, showing how far they degrade under complex backgrounds, lighting and defocus.

    arXivDOI Cite Trust