RA-Wiki-UA

RA-Wiki-UA is a benchmark-targeted Ukrainian image-text augmentation corpus for Visual Word Sense Disambiguation. It was introduced in From Sparse to Sense-Grounded: Wikipedia Training for Ukrainian Visual-WSD.

Dataset construction

The dataset was created to reduce the visual and linguistic domain gap between generic image-text training data and Ukrainian Visual-WSD. Rather than sampling Wikipedia uniformly, its construction begins from the 381 gold images in the Ukrainian Visual-WSD benchmark. Each gold image is used only as a visual retrieval query and is not included in RA-Wiki-UA as a training image.

For every benchmark instance, the ten nearest images are retrieved from a pool of approximately 150,000 Ukrainian Wikipedia images. Retrieval uses cosine similarity in the embedding space of OpenCLIP ViT-H/14 + XLM-R large (LAION-5B). This produces a Wikipedia neighborhood that is visually close to the intended benchmark sense and concentrates supervision on relevant visual concepts.

Wikipedia captions are sparse and often too short or templatic for contrastive training. Therefore, GPT-5.1 is given each retrieved image together with its Ukrainian Wikipedia page title, available caption, and article summary. The model uses this visual and encyclopedic context to disambiguate concepts in the image and generate ten diverse Ukrainian descriptions.

The resulting structure is:

381 benchmark retrieval queries
× 10 retrieved Wikipedia images
× 10 generated Ukrainian descriptions
= 38,100 image-text training pairs

The 3,810 retrieval positions correspond to 3,288 unique packaged Wikipedia images because the same image can be retrieved for more than one benchmark instance.

RA-Wiki-UA is a benchmark-targeted supervision source rather than a benchmark-independent pretraining corpus. It is designed to test whether Wikipedia images from the visual neighborhood of target senses, paired with context-conditioned Ukrainian descriptions, can improve sense-level image-text alignment.

Data fields

• sample_id: Stable identifier for the image-text pair.
• file_name: Portable relative path to the packaged image.
• generated_caption_uk: Ukrainian image description generated by GPT-5.1.
• word, word_definition, trigger: Sense-disambiguation context.
• similarity, rank: Retrieval score and rank from dataset construction.
• page_title, wikipedia_page_url, page_summary: Ukrainian Wikipedia context.
• original_caption: Caption collected during Wikipedia mining.
• benchmark_uuid: Identifies the benchmark instance whose gold image was used as the retrieval query.
• caption_index: Zero-based generated-caption index for the candidate.

Limitations

• Generated descriptions may contain hallucinations or other model errors.
• Retrieval candidates and captions inherit biases and coverage limitations from Ukrainian Wikipedia and GPT-5.1.
• The data is connected to a 381-instance Visual-WSD benchmark; users should carefully avoid evaluation leakage when designing experiments.
• Image provenance and licensing metadata must be completed before a public release.

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},
  year = {2026}
}