Positioning
How BUDOVA compares
Against other Ukrainian-language corpora, across the criteria that matter for domain NLP research.
| Criterion | BUDOVA | UberText 2.0 | CC-100 UK | UA-GEC |
|---|---|---|---|---|
| Domain coverageConstruction terminology | Dedicated | General web | General web | Grammar errors |
| Speech + text | Planned | |||
| NER annotations9 entity types | ||||
| Inter-annotator agreement | Cohen’s κ per document | Not reported | Not applicable | κ reported |
| Multi-dialect speech | Planned: 27 regions | |||
| License | CC BY 4.0 (planned) | No CC license stated (fair use) | Mixed / fair use | CC BY 4.0 |
| Size (approx.) | Target: 100M raw-text tokens, 100h speech; now 2,927 annotated segments | 3.27B tokens (2023) | 84.6 GiB (≈6–7B tokens) | 30k sentences |
| Annotation platformOpen contribution | ||||
| Hugging Face availability | Planned (v1.0) |
Other corpora: size as of their latest release; BUDOVA: project targets and current status. BUDOVA targets narrow-domain depth; general-purpose corpora target breadth. The two complement each other for downstream Ukrainian NLP. Sources: UberText 2.0 — Chaplynskyi, UNLP 2023 (aclanthology.org/2023.unlp-1.1) and lang.org.ua/ubertext; CC-100 — Conneau et al., ACL 2020 (XLM-R), Table 6 (arxiv.org/abs/1911.02116).
