Questions answered
Common questions about the corpus, access, licensing, and how to contribute.
Yes. The subsets will be released under open licenses — CC BY 4.0 for most of them, which requires only attribution (citation of the dataset card) and permits both commercial and non-commercial use. Each subset’s card will state its license.
The annotated text on the platform currently comes from Ukrainian building codes (ДБН). Each segment receives three independent annotation passes, followed by adjudication against the published annotation guidelines and an audit of unanimous records (those where all three passes agree). Cohen’s κ is reported per document; expert validation is pending.
Speech: contributors read short construction texts aloud in the browser (WebM/Opus, or MP4/AAC on Safari) or upload audio files (WAV preferred); in the data, speakers appear only under pseudonyms (
SPK-NNN).Nine entity classes:
material,tool,process,measurement,structure,safety,regulation,property,load_effect. Label definitions and span rules are formalized in the annotation guidelines (see /guidelines).Yes. Apply to join the annotation platform — after approval you can annotate existing tasks, read short construction texts aloud, submit new source texts, and add lexicon terms. The most active contributors appear in the contributors table on the home page; crediting contributors in the dataset card is planned.
Each segment receives three independent annotation passes. The passes are adjudicated against the published guidelines, and unanimous records (all three passes agree) go through a separate audit. Cohen’s κ is reported per document, and the IAA page shows agreement metrics for the merged entries. Expert validation is pending.
Not yet — the first sample (text corpus with datasheet and guidelines) is planned for Q4 2026 on Hugging Face and Zenodo. Until then, the annotated segments can be browsed in the corpus browser.
Use the BibTeX entry in the Citation section of the Dataset page. It covers the project as a whole; per-subset citations (speech, text, lexicon) will be provided in each Hugging Face repository as the subsets are released.
Audio files are stored under server-assigned names and without the metadata their file format can carry (tags, recording software, device or browser, creation time); the sound itself is not changed. Speakers appear in the data only under pseudonyms (
SPK-NNN) — this is pseudonymisation, not anonymisation: a voice stays recognisable. Contributors can export their data, withdraw a consent or delete their account at any time through the annotation workspace (GDPR panel). Details are in the Privacy Policy.
