Danish Wikipedia — Image, Caption, Context
First large-scale Danish image–text dataset that preserves neighboring article context, not just a single caption.
vision-languagedatasetsdanishwikipedia
The first large-scale Danish image–text dataset that keeps neighboring article context, not just a single caption.
The gap this fills
Most Wikipedia-derived image–text datasets reduce each image to caption + alt-text, dropping the surrounding paragraph that usually disambiguates what the picture is actually showing. For English, the loss is recoverable from the sheer volume of data. For Danish — where total available image–text pairs are orders of magnitude smaller — every bit of context you throw away hurts downstream models.
What’s preserved
For each image extracted from Danish Wikipedia, the dataset stores:
- The image URL and basic metadata.
- The literal caption from the figure.
- The section heading the image appears under.
- The neighboring paragraph(s) surrounding the image, so models can learn from longer-form image–text relationships rather than caption-only pairs.
This makes it usable for tasks where the caption alone is ambiguous: a portrait captioned only with a name needs the surrounding article to know whether that person was a 17th-century theologian or a 20th-century footballer.