Dataset Card for WIT

Dataset Summary

Wikipedia-based Image Text (WIT) Dataset is a large multimodal multilingual dataset. WIT is composed of a curated set of 37.6 million entity rich image-text examples with 11.5 million unique images across 108 Wikipedia languages. Its size enables WIT to be used as a pretraining dataset for multimodal machine learning models.

A few unique advantages of WIT:

• The largest multimodal dataset (time of this writing) by the number of image-text examples.
• A massively multilingual (first of its kind) with coverage for over 100+ languages.
• A collection of diverse set of concepts and real world entities.
• Brings forth challenging real-world test sets.

Dataset Preprocessing

This dataset doesn't download the images locally by default. Instead, it exposes URLs to the images. To fetch the images, use the following code:

from concurrent.futures import ThreadPoolExecutor
from functools import partial
import io
import urllib

import PIL.Image

from datasets import load_dataset
from datasets.utils.file_utils import get_datasets_user_agent


def fetch_single_image(image_url, timeout=None, retries=0):
    for _ in range(retries + 1):
        try:
            request = urllib.request.Request(
                image_url,
                data=None,
                headers={"user-agent": get_datasets_user_agent()},
            )
            with urllib.request.urlopen(request, timeout=timeout) as req:
                image = PIL.Image.open(io.BytesIO(req.read()))
            break
        except Exception:
            image = None
    return image


def fetch_images(batch, num_threads, timeout=None, retries=0):
    fetch_single_image_with_args = partial(fetch_single_image, timeout=timeout, retries=retries)
    with ThreadPoolExecutor(max_workers=num_threads) as executor:
        batch["image"] = list(executor.map(fetch_single_image_with_args, batch["image_url"]))
    return batch


num_threads = 20
dset = load_dataset("wit")
dset = dset.map(fetch_images, batched=True, batch_size=100, fn_kwargs={"num_threads": num_threads})

Supported Tasks and Leaderboards

• image-captioning: This dataset can be used to train a model for image captioning where the goal is to predict a caption given the image.

• text-retrieval: The goal in this task is to build a model that retrieves the text closest to an image.

In these tasks, any combination of the caption_reference_description, caption_attribution_description and caption_alt_text_description fields can be used as the input text/caption.

Languages

The dataset contains examples from all Wikipedia languages, with the following stats:

Image-Text	# Lang	Uniq. Images	# Lang
total > 1M	9	images > 1M	6
total > 500K	10	images > 500K	12
total > 100K	36	images > 100K	35
total > 50K	15	images > 50K	17
total > 14K	38	images > 13K	38

Dataset Structure

Data Instances

{
  'language': 'en',
  'page_url': 'https://en.wikipedia.org/wiki/Oxydactylus',
  'image_url': 'https://upload.wikimedia.org/wikipedia/commons/5/5f/Oxydactylus_longipes_fm.jpg',
  'page_title': 'Oxydactylus',
  'section_title': None,
  'hierarchical_section_title': 'Oxydactylus',
  'caption_reference_description': None,
  'caption_attribution_description': 'English: Mounted skeleton of Oxydactylus longipes in the Field Museum of Natural History.',
  'caption_alt_text_description': None,
  'mime_type': 'image/jpeg',
  'original_height': 3564,
  'original_width': 2748,
  'is_main_image': True,
  'attribution_passes_lang_id': True,
  'page_changed_recently': True,
  'context_page_description': 'Oxydactylus is an extinct genus of camelid endemic to North America. It lived from the Late Oligocene to the Middle Miocene, existing for approximately 14 million years. The name is from the Ancient Greek οξύς and δάκτυλος.\nThey had very long legs and necks, and were probably adapted to eating high vegetation, much like modern giraffes. Unlike modern camelids, they had hooves, rather than tough sole-pads, and splayed toes.',
  'context_section_description': 'Oxydactylus is an extinct genus of camelid endemic to North America. It lived from the Late Oligocene to the Middle Miocene (28.4–13.7 mya), existing for approximately 14 million years. The name is from the Ancient Greek οξύς (oxys, "sharp")and δάκτυλος (daktylos, "finger").\n \nThey had very long legs and necks, and were probably adapted to eating high vegetation, much like modern giraffes. Unlike modern camelids, they had hooves, rather than tough sole-pads, and splayed toes.'
}

Data Fields

• language: Language code depicting wikipedia language of the page
• page_url: URL to wikipedia page
• image_url: URL to wikipedia image
• page_title: Wikipedia page's title
• section_title: Section's title
• hierarchical_section_title: Hierarchical section's title
• caption_reference_description: This is the caption that is visible on the wiki page directly below the image.
• caption_attribution_description: This is the text found on the Wikimedia page of the image. This text is common to all occurrences of that image across all Wikipedias and thus can be in a language different to the original page article.
• caption_alt_text_description: This is the “alt” text associated with the image. While not visible in general, it is commonly used for accessibility / screen readers
• mime_type: Mime type associated to the image.
• original_height: Image height
• original_width: Image width
• is_main_image: Flag determining if the image is the first image of the page. Usually displayed on the top-right part of the page when using web browsers.
• attribution_passes_lang_id: Compared language field with the attribution language (written in the prefix of the attribution description).
• page_changed_recently: [More Information Needed]
• context_page_description: Page description corresponds to the short description of the page. It provides a concise explanation of the scope of the page.
• context_section_description: Text within the image's section.

Figure: WIT annotation example.

Details on the field content can be found directly in the paper, figure 5 and table 12.

Data Splits

All data is held in train split, with a total of 37046386 rows.

Dataset Creation

Curation Rationale

From the repository:

Multimodal visio-linguistic models rely on a rich dataset to help them learn to model the relationship between images and texts. Having large image-text datasets can significantly improve performance, as shown by recent works. Furthermore the lack of language coverage in existing datasets (which are mostly only in English) also impedes research in the multilingual multimodal space – we consider this a lost opportunity given the potential shown in leveraging images (as a language-agnostic medium) to help improve our multilingual textual understanding.

To address these challenges and advance research on multilingual, multimodal learning we created the Wikipedia-based Image Text (WIT) Dataset. WIT is created by extracting multiple different texts associated with an image (e.g., as shown in the above image) from Wikipedia articles and Wikimedia image links. This was accompanied by rigorous filtering to only retain high quality image-text sets.

The resulting dataset contains over 37.6 million image-text sets – making WIT the largest multimodal dataset (publicly available at the time of this writing) with unparalleled multilingual coverage – with 12K+ examples in each of 108 languages (53 languages have 100K+ image-text pairs).

Source Data

Initial Data Collection and Normalization

From the paper, section 3.1:

We started with all Wikipedia content pages (i.e., ignoring other pages that have discussions, comments and such). These number about ∼124M pages across 279 languages.

Who are the source language producers?

Text was extracted from Wikipedia.

Annotations

Annotation process

WIT was constructed using an automatic process. However it was human-validated.

From the paper, section 3.7:

To further verify the quality of the WIT dataset we performed a study using (crowd-sourced) human annotators. As seen in Fig. 3, we asked raters to answer 3 questions. Given an image and the page title, raters first evaluate the quality of the attribution description and reference description in the first two questions (order randomized). The third question understands the contextual quality of these text descriptions given the page description and caption. Each response is on a 3-point scale: "Yes" if the text perfectly describes the image, "Maybe" if it is sufficiently explanatory and "No" if it is irrelevant or the image is inappropriate.

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

From the paper, section 3.4:

Lastly we found that certain image-text pairs occurred very frequently. These were often generic images that did not have much to do with the main article page. Common examples included flags, logos, maps, insignia and such. To prevent biasing the data, we heavily under-sampled all such images

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

[More Information Needed]

Citation Information

@article{srinivasan2021wit,
  title={WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning},
  author={Srinivasan, Krishna and Raman, Karthik and Chen, Jiecao and Bendersky, Michael and Najork, Marc},
  journal={arXiv preprint arXiv:2103.01913},
  year={2021}
}

Contributions

Thanks to @thomasw21, @nateraw and hassiahk for adding this dataset.