README.md

metadata

license: apache-2.0
language:
  - fr
  - en
tags:
  - arXiv
  - multimodal
  - document-type objects

We present the Arxiv Figures & Tables Database (AFTdb), which consists of an aggregation of figures and tables from scientific articles sourced from the arXiv platform.

The purpose of this dataset is to train multimodal models specialized in images of document-type objects (graphs, functional diagrams, tables, etc.), rather than photographic-type images. The idea is that a model trained on this type of data will be more coherent within the context of document corpora than a model trained on pictorial compositions. To establish a connection between the two modalities (image and text), captions for each object are also provided. As captions can sometimes be very brief, the article's summary is also included to add context to the document object if necessary.

For this reason, a corpus of scientific articles was prioritized. Due to the scientific rigor demanded, each document-type object is systematically accompanied by a caption (similar to captions for pictorial images on platforms like Flickr, for example).

The database is divided into two types of document objects: figures and tables. For the table part, it is possible to approach two different types of learning. The first, similar to figures, associates the image with the caption. However, in the data field, the LaTeX source code of the table is also provided. An objective can be to take an image of a table and convert it into text using this source code.

Loading the database

The figure part is relatively substantial, and it is advisable to use the dataset in streaming mode:

aftdb_figure = load_dataset("cmarkea/aftdb", "figure", streaming=True)

The table part is less substantial and can be downloaded locally directly:

aftdb_table = load_dataset("cmarkea/aftdb", "table")

Both categories are compatible, and it is possible to load both types simultaneously:

aftdb = load_dataset("cmarkea/aftdb", "figure+table", streaming=True)

This is the default configuration.

Statistical Description

The descended articles correspond to a portion of the articles that had their last modifications in the year 2023 on the arXiv platform.

Number of
articles	22,397
authors	50,000
figures	157,944
tables	16,415
total words in English titles	229,042
total words in French titles	301,607
total words in English abstracts	3,765,978
total words in French abstracts	4,437,968
total words in English captions	7,506,626
total words in French captions	8,311,727

Field Descriptions

• id: Unique identifier for each observation.
• paper_id: Unique arXiv identifier for each article.
• type: 'figure' for graphic objects such as graphs, functional diagrams, etc., and 'table' for tables.
• authors: Names of the article's authors.
• categories: arXiv categories of the article.
• title: Title of the article.
• summary: Article summary.
• caption: Caption of the document-type object.
• image: Pillow image of the document-type object.
• data: For figures, it represents the filename of the figure; for tables, it is the LaTeX transcription of the table.
• newcommands: List containing the LaTeX newcommands used in the article.

Citation

@online{DeAFTdb,
  AUTHOR = {Cyrile Delestre},
  URL = {https://huggingface.co/datasets/cmarkea/aftdb},
  YEAR = {2024},
  KEYWORDS = {NLP ; Multimodal}
}