Domain Challenges
Editor’s note, 13 October 2026: this post was written at launch and corrected on these points: the NER taxonomy has nine labels (see the guidelines), the lexicon is planned for a later release, and the figures in the last section now come from the live corpus. The original version listed four entity types, a lexicon size and annotation statistics that could not be verified. Updated 25 October 2026: the figures now cover nine codes, and a sentence crediting research partners the project does not have was removed.
Ukrainian construction terminology presents unique challenges for natural language processing. Unlike general-purpose Ukrainian text, construction documents mix formal literary Ukrainian with technical jargon, Soviet-era terminology inherited from Russian-language standards (ДБН/СНиП), and modern international terms adapted from English. A single concept — such as "reinforced concrete" — may appear as залізобетон, ж/б, or армований бетон depending on context and register.
Existing Ukrainian dictionaries and terminology databases cover general vocabulary well but have significant gaps in construction-specific terms. Many specialized terms have no standardized Ukrainian form — practitioners often use Russian calques or English borrowings. This inconsistency creates serious problems for NLP systems: tokenizers split compound terms incorrectly, NER models fail to recognize domain entities, and machine translation produces inaccurate or ambiguous output.
BUDOVA plans to address these gaps with a curated, structured lexicon built from primary sources — official Ukrainian building codes (ДБН) and university textbooks — and validated by practicing construction engineers and terminologists. The lexicon is planned for a later release; its terms will appear in the glossary.
Lexicon Structure
The BUDOVA construction lexicon is designed as a machine-readable, linguistically rich resource. Each entry will contain the Ukrainian term, its English equivalent, morphological information (gender, declension pattern, stress), a definition in Ukrainian, and usage examples drawn from the text corpus. Terms will be organized into a hierarchical domain taxonomy with top-level categories: Materials, Structural Elements, Processes, Equipment, Safety, and Regulatory.
The data format is structured JSON optimized for integration with NLP pipelines:
- term_uk: canonical Ukrainian form with stress marking
- term_en: English translation
- pos: part of speech (noun, verb, adjective)
- morphology: gender, number, declension class
- definition_uk: Ukrainian-language definition
- examples: array of attested usage examples with source references
- domain: hierarchical category path (e.g., Materials > Concrete > Admixtures)
- synonyms: variant forms, abbreviations, colloquial equivalents
The lexicon is planned for a later release. The target is 5–10K entries, each to be reviewed by domain specialists.
NER Annotation Schema
The NER annotation schema grew alongside this terminology work. The first sketch at launch had four entity types — MATERIAL, ELEMENT, PROCESS and PROPERTY. As real building codes were annotated, the taxonomy grew to nine labels, defined in the public annotation guidelines:
- material: construction substances, composites and named products (e.g., портландцемент, арматура А500С)
- tool: machinery, equipment and instruments (e.g., кран, риштування, опалубка)
- process: construction work somebody performs (e.g., бетонування, монтаж опалубки, гідроізоляція)
- measurement: numeric quantities with units, class codes and dimensions
- structure: structural and architectural elements, engineering works and buildings (e.g., несуча стіна, фундаментна плита)
- safety: means, measures and named regimes that keep people from harm
- regulation: normative references and legal acts (e.g., ДБН В.2.6-98:2009)
- property: named technical properties and performance classes (e.g., міцність на стиск, клас вогнестійкості)
- load_effect: loads and other actions on a structure, and the mechanical state they cause (e.g., навантаження, прогин)
Spans are stored as character offsets in JSON-lines records (see the documentation); the platform can also export IOB2-tagged CoNLL.
As of October 2026, the corpus holds 2,927 annotated segments from nine building codes (ДБН), with about 16,700 gold entity spans. Each segment receives three independent annotation passes, followed by adjudication against the published guidelines and an audit of unanimous records; agreement figures are on the IAA page.
