MMDocIR: Benchmarking Multi-Modal Retrieval for Long Documents

1. Abstract

Multi-modal document retrieval is designed to identify and retrieve various forms of multi-modal content, such as figures, tables, charts, and layout information from extensive documents. Despite its significance, there is a notable lack of a robust benchmark to effectively evaluate the performance of systems in multi-modal document retrieval. To address this gap, this work introduces a new benchmark, named as MMDocIR, encompassing two distinct tasks: page-level and layout-level retrieval. The former focuses on localizing the most relevant pages within a long document, while the latter targets the detection of specific layouts, offering a more fine-grained granularity than whole-page analysis. A layout can refer to a variety of elements such as textual paragraphs, equations, figures, tables, or charts. The MMDocIR benchmark comprises a rich dataset featuring expertly annotated labels for 1,685 questions and bootstrapped labels for 173,843 questions, making it a pivotal resource for advancing multi-modal document retrieval for both training and evaluation. Through rigorous experiments, we reveal that (i) visual retrievers significantly outperform their text counterparts; (ii) MMDocIR train set can effectively benefit the training process of multi-modal document retrieval; (iii) text retrievers leveraging on VLM-text perform much better than those using OCR-text. These findings underscores the potential advantages of integrating visual elements for multi-modal document retrieval.

2. Task Setting

Page-level Retrieval

The page-level retrieval task is designed to identify the most relevant pages within a document in response to a user query.

Layout-level Retrieval

The layout-level retrieval aims to retrieve most relevant layouts. The layouts are defined as the fine-grained elements such as paragraphs, equations, figures, tables, and charts. This task allows for a more nuanced content retrieval, honing in on specific information that directly answers user queries.

3. Evaluation Set

3.1 Document Analysis

MMDocIR evaluation set includes 313 long documents averaging 65.1 pages, categorized into ten main domains: research reports, administration&industry, tutorials&workshops, academic papers, brochures, financial reports, guidebooks, government documents, laws, and news articles. Different domains feature distinct distributions of multi-modal information. For instance, research reports, tutorials, workshops, and brochures predominantly contain images, whereas financial and industry documents are table-rich. In contrast, government and legal documents primarily comprise text. Overall, the modality distribution is: Text (60.4%), Image (18.8%), Table (16.7%), and other modalities (4.1%).

3.2 Question and Annotation Analysis

MMDocIR evluation set encompasses 1,658 questions, 2,107 page labels, and 2,638 layout labels. The modalities required to answer these questions distribute across four categories: Text (44.7%), Image (21.7%), Table (37.4%), and Layout/Meta (11.5%). The ``Layout/Meta'' category encompasses questions related to layout information and meta-data statistics. Notably, the dataset poses several challenges: 254 questions necessitate cross-modal understanding, 313 questions demand evidence across multiple pages, and 637 questions require reasoning based on multiple layouts. These complexities highlight the need for advanced multi-modal reasoning and contextual understanding.

4. Train Set

Citation

If you use any datasets from this organization in your research, please cite the original dataset as follows:

@misc{dong2025mmdocirbenchmarkingmultimodalretrieval,
      title={MMDocIR: Benchmarking Multi-Modal Retrieval for Long Documents}, 
      author={Kuicai Dong and Yujing Chang and Xin Deik Goh and Dexun Li and Ruiming Tang and Yong Liu},
      year={2025},
      eprint={2501.08828},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2501.08828}, 
}