MultimodalQA_Doc Documentation

This repository contains Wikipedia data crawled for the MultimodalQA dataset, a dataset designed for multimodal question answering. The dataset combines text, tables, and images from Wikipedia articles to enable research on answering questions that require understanding and reasoning across multiple modalities.

Included in this repository is a data loading script. It reads text, table, and image data stored in Parquet format and reconstructs them into individual JSON documents along with their corresponding image files. Each JSON document is organized by doc_title and contains structured components for text, tables, and images, preserving headings, captions, and hyperlinks for each element.

Folder Structure

Input Parquet Folder

• text.parquet : Text components per document
• table.parquet : Table components per document
• image.parquet : Image component metadata per document
• image_dump.parquet : Actual byte data of images (mapped by image_name)

Example:

parquet_data/
├── text.parquet
├── table.parquet
├── image.parquet
└── image_dump.parquet

Output Folder Structure

• <save_path>/ : Restored JSON documents
• <save_path>/images/ : Image files

Example:

restored_data/
├── images/
│   ├── image_1.png
│   ├── image_2.png
│   └── ...
├── Document_1.json
├── Document_2.json
└── ...

JSON Structure

Each document is stored in the following structure:

{
  "title": "Document Title",
  "text": {
    "component_id_1": {
      "text": "Main text content",
      "heading_path": ["Section 1", "Subsection A"],
      "hyperlinks": ["https://example.com"],
      "label_id": 0
    }
  },
  "table": {
    "component_id_2": {
      "table": [[...], [...]],
      "heading_path": ["Section 2"],
      "hyperlinks": [],
      "label_id": 1
    }
  },
  "image": {
    "component_id_3": {
      "image_name": "image_1.png",
      "heading_path": ["Section 3"],
      "caption": "Image caption",
      "hyperlinks": [],
      "label_id": 2
    }
  }
}

• text, table, image use the component ID as the key, containing detailed information for each component.
• Image files are saved in the images/ folder with .png format.

Usage

python load.py --parquet_path /path/to/parquet_data --save_path /path/to/restored_data

• --parquet_path : Path to the folder containing Parquet files
• --save_path : Path where restored JSON documents and images will be saved (must not already exist)

After successful restoration, the terminal will display output like:

Loading Parquet Files...
Restoring JSON Documents: 100%|████████████| 3236/3236 [00:33<00:00, 96.34it/s]
Successfully restored 3236 documents. Folder Path: /workspace/mmqa_load

• The example above shows the actual number of documents restored.