Wikipedia Multilingual Image-Text Pairs With 37.6 Million Examples

The Wikipedia-based Image Text (WIT) Dataset contains 37.6 million entity-rich image-text examples paired with 11.5 million unique images across 108 Wikipedia languages. It was created by keshan for pretraining multimodal machine learning models and was last updated in August 2021.

Use Cases

• Pretrain a vision-language model on 37.6 million entity-rich image-text examples to learn cross-modal representations.
• Train a multilingual image captioning model using text examples across 108 Wikipedia languages.
• Fine-tune a model for cross-lingual image retrieval using the 11.5 million unique images and their associated multilingual text.
• Analyze the distribution of visual entities across different language editions of Wikipedia using the curated image-text pairs.

Strengths

• 37.6 million image-text examples provide substantial scale for model training.
• Multilingual coverage spans 108 Wikipedia languages.
• Contains 11.5 million unique images, offering visual diversity.

Limitations

• Dataset specifics like column structure, file formats, and license are unknown.
• The data is from 2021 or earlier, which may not reflect recent Wikipedia content or current events.
• Potential bias exists towards entities and topics well-covered on Wikipedia.

Provenance

Source

Wikipedia

Collection Method

Curated set of image-text examples extracted from Wikipedia.

Freshness

Last updated on 2021-08-07.