SenseWiki-UA

SenseWiki-UA is a sense-aligned Ukrainian image-text training corpus introduced in From Sparse to Sense-Grounded: Wikipedia Training for Ukrainian Visual-WSD.

Dataset construction

The dataset was created to provide Ukrainian multimodal supervision explicitly anchored to lexical senses. Its construction begins with dictionary entries and sense definitions rather than generic web image-caption pairs.

For each dictionary entry, candidate Ukrainian Wikipedia pages are retrieved. The dictionary definition is compared with the lead definition or summary of each candidate page in a shared embedding space. The closest Wikipedia definition is selected, and alignments below a cosine-similarity threshold of 0.5 are discarded. Images associated with accepted Wikipedia pages are then paired with Ukrainian textual supervision.

This process connects three elements:

dictionary-defined sense
→ aligned Ukrainian Wikipedia page
→ real image and Ukrainian text

The resulting pairs teach an image-text model about the visual realization of a specific lexical meaning, rather than only broad topical similarity. SenseWiki-UA complements RA-Wiki-UA, which instead retrieves Wikipedia images from the visual neighborhoods of the Ukrainian Visual-WSD benchmark.

This release contains 27,543 image-text rows, 24,140 unique images, and coverage of 7,557 dictionary headwords. The two construction-time CSV partitions are combined into one training split, matching the way the corpus is reported and used as training supervision in the paper.

Data fields

• sample_id: Stable identifier for the image-text pair.
• image: Packaged Wikipedia image exposed through the Hugging Face Image feature.
• caption_uk: Ukrainian text paired with the image.
• word: Dictionary headword.
• word_definition: Dictionary definition used for sense alignment.
• alignment_similarity: Cosine similarity between the dictionary definition and selected Wikipedia page definition.
• page_title: Title of the aligned Ukrainian Wikipedia page.
• wikipedia_page_url: Source Ukrainian Wikipedia page.
• page_summary: Lead text or summary extracted from that page.

Limitations

• Automatic definition alignment can select the wrong lexical sense.
• A Wikipedia image can be weakly related or unrelated to the aligned sense.
• Ukrainian Wikipedia has uneven coverage across words, senses, and domains.
• Some textual supervision is generic, very short, or only indirectly visual.
• The paper's manual audit found 68% of sampled pairs usable, 12% noisy but usable, and 20% not usable.

Citation

@inproceedings{laba-hryniv-2026-sparse,
  title = {From Sparse to Sense-Grounded: Wikipedia Training for Ukrainian Visual-WSD},
  author = {Laba, Yurii and Hryniv, Rostyslav O.},
  booktitle = {Proceedings of the 30th Conference on Computational Natural Language Learning},
  pages = {501--514},
  year = {2026}
}