Guidelines

BUDOVA covers construction-domain Ukrainian across four registers: normative documents (ДБН / ДСТУ), specifications, estimates, and field speech. Out-of-domain text is rejected before annotation.

Every source fragment receives NER labels (8 entity types), a register label, and — for speech samples — dialect region and speaker ID.

Figure captions, formulas, table headers, and numeric-only fragments without surrounding context are excluded. Annotate sentences that carry domain information, not typography.

Construction substances, composites, and named products — concrete, rebar, insulation, window units, drywall, cement. Multi-word canonical names are kept whole.

Machinery, equipment, and instruments used on-site — excavator, crane, drill, trowel, scaffolding, formwork.

Construction activities — casting, welding, installation, plastering, drilling, waterproofing. Prefer verbal-noun forms.

Numeric quantities with units, class codes, dimensions, and concentration percentages. Always mark the unit as part of the span.

Structural and architectural elements — foundations, columns, beams, slabs, roofs, stairs, ramps.

Safety equipment, procedures, and protective measures — PPE, helmets, harnesses, fire extinguishers, evacuation routes, protective railings. Standard citations go to regulation; numeric protection classes go to property.

Normative references — ДБН, ДСТУ, СОУ, ISO and EN standards, Eurocodes, technical specifications (ТУ). Keep the full code with the year as one span (see boundaries-3).

Named technical properties and performance classes — fire-resistance class, thermal conductivity, sound-insulation index, load-bearing capacity, strength grade. A numeric value with its unit is a measurement; the property name itself is property.

When a domain adjective or quantifier changes the meaning of the head noun, include it in the span.

Do not include bare articles, prepositions, or discourse fillers at the edges of a span.

Terms with a fixed collocation (e.g. ДБН + code, class + number) never split across annotations.

If a longer material name contains a numeric class, mark the full material span; the measurement stays inside and is not separately annotated.

When two distinct entities abut, give each its own span.

Use the exact dialectal word as it appears. Canonical mapping is a separate metadata field, not a substitution.

For speech samples, attach region + dialect group (Northern / Southwestern / Southeastern — the standard three-group Ukrainian taxonomy) to the utterance, not to individual spans.

Use a dedicated mic if possible (USB condenser, lavalier, decent phone). Wear headphones to prevent speaker bleed. Grant microphone permission to the browser when prompted — the platform stops the audio track when recording ends, so the indicator should go away.

Record in a quiet room (no HVAC blast, no traffic). Take 3–5 seconds of silence first and check the noise floor on your meter. Corpus target is SNR ≥ 20 dB; the upload response includes a warning when a WAV measures below.

Loud enough to be clearly above noise, quiet enough not to clip. Aim for peaks between −12 and −3 dBFS. The server rejects nothing on level alone but stores peak_dbfs so clipped takes can be filtered at export.

Prefer mono 48 kHz 16-bit PCM WAV for external-DAW uploads. The in-browser recorder produces 48 kHz mono WebM/Opus at 64 kbit/s (Safari falls back to MP4/AAC). One utterance per file, 30 s – 3 min. Max 50 MB.

Each recording should have a single speaker. If you are interviewing someone, split into one file per speaker. If a true dialogue is captured, note it in source_ref ("2 speakers, foreman + worker").

No full names, addresses, phone numbers, license plates, or personal medical info. Dialectal vocabulary is welcome — tag your region so dialect researchers can find it. Construction-domain only.

When a brand name qualifies a material, annotate the full span as material and store the brand in a sub-field.

A brand name without a construction-material context is not annotated.

A number without unit is not a measurement span. Include mm/cm/m/МПа/°C/%/class-code suffix in the annotation.

For multi-axis dimensions (400×400 mm), keep the entire product in one span. Same for tolerance ranges.

Class labels (C25/30, XC3, B400C) are measurement spans; include the prefix letter and slash.

ДБН / ДСТУ / СОУ text and other standards — formal legal register with explicit numeric thresholds and "must/shall" modals.

Product datasheets, technical specs — numeric-heavy, brand references common, imperative and descriptive sentences.

Cost estimates, work-item lines, unit-rate tables. Recognized by work-type verb phrases and per-unit coefficients.

On-site speech and informal field notes — dialectal vocabulary, approximate numbers, discourse fillers.

If a term could be a material OR a structure, choose whichever is narrower in the current sentence.

Tasks where two annotators disagree after the first pass go to the admin-review queue with a "needs adjudication" flag.

Each task is annotated independently by N annotators (default 3). Spans are projected onto characters; a character keeps an entity label when at least ⌈N/2⌉ annotators assigned it. Consensus spans whose average per-character support falls below 0.6 are dropped from the merged entry — they stay visible in the per-word agreement table.

The headline coefficient is Fleiss’ κ computed only over the entity window — character positions where at least one annotator marked an entity. Unannotated text does not enter the calculation, so the value is comparable between a two-sentence fragment and a multi-page clause: the same labeling pattern yields the same κ regardless of text length.

For every annotator pair, spans match when they share a label and overlap with IoU ≥ 0.5. The mean pairwise F1 measures whether annotators find the same entities — complementing κ, which measures positional consistency.

validation_score = 0.35 · max(0, κ_entity) + 0.65 · F1. Span detection is weighted higher because it is what NER training quality depends on. Entries are colour-coded: ≥ 0.7 high, 0.4–0.7 medium, < 0.4 low (queued for adjudication).

κ ≤ 0 — no better than chance; 0.2–0.4 fair; 0.4–0.6 moderate; 0.6–0.8 substantial; > 0.8 almost perfect. κ = 1 with an empty entity window means all annotators agreed the fragment contains no entities.

ДБН В.1.1-12:2014 (seismic design), ДБН В.2.6-98:2009 (concrete structures), ДБН А.3.1-5:2016 (organization of construction).

The eight-label set adapts Ontonotes-style named-entity taxonomy to the construction domain. Full rationale in the project paper (v1.0).

When citing BUDOVA guidelines: "BUDOVA Annotation Guidelines v0.2 (2026)". Deep-link individual rules with the anchor, e.g. /guidelines#entities-material.