Construction sites run on radio. During a concrete pour, a single crackling utterance — “the wakku is about to burst,” “noro is leaking,” “reduce it” — can carry the difference between routine work and a suspended pour, a reinforcement crew, or a restart that must wait for inspection. A new study published in the journal Architectural Research presents a proof of concept for turning those fragmented, jargon-laden radio messages into structured, machine-readable safety records, using a two-layer ontology and a rule-based extraction pipeline tested on a simulated day of pouring work.
The study, authored by Young-Jun Park and published on 16 October 2026, tackles a problem that has long frustrated efforts to mine construction communication for safety insight. Radio utterances are short and elliptical. Speakers omit subjects, receivers, locations and work objects. Site slang — Japanese-derived expressions such as “wakku” for formwork, “noro” for cement paste, and “narashi” for leveling — does not map cleanly onto standard safety terminology. And the same word can change meaning entirely depending on what it is combined with. The meaning of a message, the author argues, lives in the relations among its semantic elements, not in individual words.
The framework addresses this with two layers of processing. The first layer classifies every expression in an utterance along seven semantic axes: work, act, object, target, phenomenon, interjection, and profanity. Field expressions are first normalized into standard terms, drawing on Korean construction terminology dictionaries from the Ministry of Land, Infrastructure and Transport and the Korea Construction Standards Center. So “the wakku is about to burst” is split into the object “formwork” and the phenomenon “imminent failure or deformation.” Profanity and shouting are not deleted but preserved as utterance-attitude cues, deliberately kept out of hazard judgment.
The second layer is where interpretation happens. Relation rules combine the tags to derive hazard signals, control actions, and situation states. Formwork failure risk, for example, is not a label attached to any single word; it emerges only when the formwork object and the imminent-deformation phenomenon are bound together by an object-phenomenon relation under specific clause and context conditions. The rules also guard against two failure modes: concept omission, where a hazard goes untagged, and concept excess, where a routine report is inflated into an alarm. “Pressure is rising” alone does not trigger a pump stop; “reduce the amount” is classified as a delivery-amount reduction only after its target is confirmed, never over-read as a pump shutdown.
To test the procedure, the author constructed a one-day virtual radio log of 25 utterances covering a concrete pour for a third-floor slab and beam, dated 3 July 2026. The scenario runs chronologically through normal work confirmations, a pump pressure rise and hose kink response, and a formwork deformation sequence ending in a restart prohibition. Three construction practitioners with at least fifteen years of field experience were consulted on the ontology’s core content, and the author prepared a complete reference mapping — tags, ordered relation triples, hazard signals, control actions, and states — for all 25 utterances. The experts reviewed the ontology-derived results and judged them reasonable for the virtual scenario.
The rule-based extractor, implemented with lexical dictionaries, regular expressions and chronological state updates, was scored item by item against that reference. Twenty-one of the 25 utterances showed complete agreement across all categories. Disagreements or review holds appeared in four utterances, at timestamps 08:31:10, 13:42:21, 13:43:36 and 13:44:20. Situation state classification agreed for all 25 utterances, and on the 18 utterances whose reference hazard list was empty, the extractor generated no spurious hazard signals in any of them. Where targets were omitted or ambiguous, the system left items on review hold for human inspection rather than guessing — a design choice the author presents as a feature, not a concealment of errors.
The study also ran an exploratory comparison with five large language models: Qwen2.5-7B, Gemma3-4B, Llama3.1-8B, GPT-6 Astra and Claude Fable 5.1. Each model received the same fixed prompt, one utterance at a time with its metadata and an allowed tag list, and each of the 25 utterances was processed ten times per model, yielding 1,250 outputs in total. Reference tags were withheld, and no conversation history was allowed. The results were starkly uneven. Mean F1 scores for seven-axis tag mapping reached 89.57 percent for GPT-6 Astra and 85.11 percent for Claude Fable 5.1, but fell to 39.65 percent for Qwen2.5-7B, 37.26 percent for Gemma3-4B and 35.86 percent for Llama3.1-8B. Repeatability varied too: the normalized tag sets changed between runs for 12 of 25 utterances under Qwen2.5-7B and 21 under Gemma3-4B, but for only one under GPT-6 Astra and none under Llama3.1-8B. Notably, Gemma3 produced identical aggregate scores despite changing tags on two utterances — a reminder that stable scores do not guarantee stable tag selection.
The author is explicit about the boundaries of these findings. The 25 utterances and their reference mapping were created within the same study that evaluated them, with no held-out scenarios or independently annotated field data. The metrics therefore support internal consistency checks for the development case, not estimates of performance on real sites. No audio was collected or transcribed; the virtual log was treated as error-free text, so the effects of imperfect speech recognition — noise, transmission conditions, dialects, shouting — remain unquantified. Terminal assignment supplies an assumed role, not an authenticated speaker or measured location. Sampling settings and instructions were not fully equivalent across the tested models, so the LLM differences cannot be attributed solely to model capability, and prompt sensitivity was not evaluated.
Within those limits, the demonstration shows what the structured records can do. Deterministic filters over the saved utterance-level records traced how a reported condition was followed by an instruction or status update — for instance, retrieving the site manager’s 13:44:20 formwork-reinforcement instruction and the restart restriction pending inspection. The author notes that such histories could eventually support systems for reconstructing site situations and presenting evidence for managerial review, with links to response procedures such as reinforcement, rescue coordination or material supply. Those connections remain prospective; response routing and execution were not implemented or evaluated.
The contribution, as the paper frames it, is the explicit representation and the implemented processing sequence itself — a procedure for converting linked radio text into records that can support later review of communicated conditions and instructions. Audio acquisition, speech recognition, speaker authentication and field performance all remain subjects for subsequent validation. What the study establishes is feasibility within its constructed case: a path from the clipped, slang-heavy shorthand of a pour site’s radio channel to a structured account of what was reported, what was decided, and what still awaits a human check before any real-world deployment.
Subject of Research: An ontology-based proof of concept for classifying concrete pouring radio utterances into semantic tags, hazard signals, control actions and situation states.
Article Title: An Ontology-Based Classification Framework for Situation Extraction from Concrete Pouring Radio Logs
Article References: Park, Y.-J. (2026). An Ontology-Based Classification Framework for Situation Extraction from Concrete Pouring Radio Logs. Architectural Research, 28(1), Article 24. https://doi.org/10.1007/s44515-026-00029-7
Image Credits: AI Generated
DOI: 10.1007/s44515-026-00029-7
Keywords: ontology, construction safety, concrete pouring, radio communication, natural language processing, knowledge graphs, large language models, rule-based extraction, situation awareness, proof of concept, Ontology-Based, Classification
News Source: Denise Maddox. (October 28, 2026). Ontology Framework Turns Concrete Pouring Radio Chatter into Safety Records. Scienmag.
