|
--- |
|
language: |
|
- en |
|
pretty_name: "Arxiv-tables Dataset" |
|
tags: |
|
- scientific-tables |
|
- document-understanding |
|
- arxiv |
|
license: cc-by-4.0 |
|
task_categories: |
|
- table-to-text |
|
- image-to-text |
|
--- |
|
|
|
# Arxiv-tables Dataset |
|
|
|
## Dataset Summary |
|
|
|
The Arxiv-tables dataset is a collection of tables extracted from scientific papers published on arXiv, primarily focused on ML papers. It includes both the LaTeX source of the tables and their corresponding rendered images from the PDF versions of the papers. |
|
|
|
## Supported Tasks |
|
|
|
This dataset can support several tasks, including but not limited to: |
|
|
|
- Table structure recognition |
|
- LaTeX to image generation for tables |
|
- Image-to-LaTeX conversion for tables |
|
- Table information retrieval |
|
|
|
## Dataset Structure |
|
|
|
The dataset is organized as follows: |
|
```bash |
|
arxiv_tables/ |
|
├── images/ |
|
│ └── [arxiv_id]/ |
|
│ ├── table_0.png |
|
│ ├── table_1.png |
|
│ └── ... |
|
├── pages/ |
|
│ └── [arxiv_id]/ |
|
│ ├── page_0.png |
|
│ ├── page_1.png |
|
│ └── ... |
|
└── gt.json |
|
``` |
|
|
|
Each arXiv paper has its own directory containing: |
|
- Individual table images |
|
- Full page images |
|
- A ground truth JSON file with table information |
|
|
|
### Data Instances |
|
|
|
Each instance in the dataset corresponds to a table and includes: |
|
|
|
- Table image |
|
- Full page image containing the table |
|
- LaTeX source of the table |
|
- Extracted content from the PDF |
|
- Bounding box coordinates |
|
- Page number |
|
- Similarity score between LaTeX and extracted content |
|
|
|
### Data Fields |
|
|
|
The `gt.json` file contains the following fields for each table: |
|
|
|
- `id`: Unique identifier for the table |
|
- `page`: Page number where the table appears |
|
- `bounding_box`: Coordinates of the table in the PDF |
|
- `latex_content`: Original LaTeX code of the table |
|
- `extracted_content`: Content extracted from the PDF version |
|
- `similarity_score`: Similarity between LaTeX and extracted content |
|
- `table_image`: Path to the cropped table image |
|
- `page_image`: Path to the full page image |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
This dataset was created to provide a large-scale resource for researchers working on table understanding in scientific documents. |
|
|
|
### Source Data |
|
|
|
The source data comes from arXiv, a free distribution service and open-access archive for scholarly articles. The dataset focuses on papers in the biology domain. |
|
|
|
### Annotations |
|
|
|
The dataset does not contain manual annotations. The ground truth is automatically generated by matching LaTeX source code to the rendered PDF content. |
|
|
|
### Other Known Limitations |
|
|
|
- The matching between LaTeX source and PDF content is based on similarity scores and may not be perfect. |
|
- Some tables may be missed or incorrectly extracted due to limitations in the extraction process. |
|
- The dataset is limited to papers that have both LaTeX source and PDF versions available on arXiv. |
