An Ontology-Based Classification Framework for Situation Extraction from Concrete Pouring Radio Logs

Technical Summary: An Ontology-Based Classification Framework for Situation Extraction from Concrete Pouring Radio Logs

Problem Statement
Concrete pouring involves complex, simultaneous operations (e.g., pump delivery, formwork checks, worker movement) where safety relies heavily on real-time radio communication. However, radio utterances in construction sites are distinct from standard documents; they are short, elliptical, and frequently omit speakers, locations, or objects. They often contain site-specific slang, Japanese-derived expressions, shouting, profanity, and urgent commands. Consequently, simple keyword detection is insufficient for identifying equipment conditions, temporary structure states, hazard signals, or control actions. The core challenge is to normalize these unstructured, field-specific utterances into standard concepts and derive safety-critical situations (hazards, controls, states) based on the semantic relations between elements rather than isolated words.

Methodology
The study proposes a two-layer ontology-based framework designed to process radio logs through normalization, semantic tagging, and relation-based judgment.

1. Field Expression Normalization: Before ontology mapping, raw utterances are normalized. Site slang (e.g., "wakku" for formwork, "noro" for cement paste) and non-standard expressions are converted into standard construction terminology. This stage also segments utterances into object, phenomenon, act, and discourse/emotion units, preserving profanity and urgency as attitude markers rather than deleting them.
2. Two-Layer Ontology Structure:
   • Layer 1 (Detailed Semantic Tags): Normalized expressions are mapped to seven semantic axes: Work, Act, Object, Target, Phenomenon, Interjection/Discourse, and Profanity/Emotion. Each axis has specific codes (e.g., Object includes equipment, temporary structures, materials; Phenomenon includes pressure rise, kinks, leakage).
   • Layer 2 (Situation Derivation): Hazard signals, control actions, and situation states are not assigned to single words but are derived from relation patterns between Layer 1 tags. Key relations include object-phenomenon binding (e.g., Formwork + Imminent Failure = Formwork Failure Risk), act-target binding (e.g., Command + Worker + Evacuation Verb = Worker Clearance), and state transition rules (e.g., Hazard Signal + Confirmation = Resolved State).
3. Evaluation Strategy: Due to the difficulty of collecting and disclosing real-world radio data (which may contain sensitive or inappropriate content), the study utilizes a researcher-constructed one-day virtual radio log containing 25 utterances. A reference ontology mapping table was created for this log. The framework's performance was evaluated by comparing rule-based extraction results against this reference table. Additionally, a local Large Language Model (LLM, qwen2.5:7b) was tested as a baseline for detailed tag mapping to assess its ability to handle field expressions compared to the rule-based approach.

Key Contributions

• Structured Semantic Framework: The study defines a detailed tagging scheme that separates site slang, interjections, and profanity into specific semantic axes rather than treating them as noise or simple hazard keywords.
• Relation-Based Judgment: It shifts the paradigm from keyword matching to relation-based inference. Hazard signals and control actions are derived only when specific semantic relations (e.g., object-phenomenon binding) are confirmed, preventing false positives from isolated words.
• Evaluation and Review-Hold Mechanism: The paper introduces a rigorous evaluation framework using a reference mapping table. Crucially, it defines "review-hold" criteria for utterances that cannot be automatically confirmed (e.g., missing targets or ambiguous slang), distinguishing between automatically determinable items and those requiring human review.
• LLM Role Definition: The study positions LLMs not as final decision-makers for safety monitoring but as tools for preliminary tag mapping, highlighting their current limitations in stable axis separation and code selection compared to rule-based systems.

Results
Using the 25-utterance virtual log, the rule-based extraction system achieved the following agreement rates against the reference table:

• Detailed Tags: 95.9% agreement (3.4% review-hold).
• Relations: 86.4% agreement (9.1% review-hold).
• Hazard Signals: 90.0% agreement (10.0% review-hold).
• Control Actions: 97.0% agreement (3.0% review-hold).
• Situation States: 100.0% agreement (0.0% review-hold).
• Non-Hazard Judgment: 100.0% accuracy in identifying normal work utterances as non-hazardous.

In contrast, the LLM-based mapping achieved only 34.5% agreement on detailed tags and generated 49 excess items not present in the reference table. The LLM struggled with separating axes (e.g., confusing inspection targets with actual phenomena) and selecting correct codes for field expressions.

Significance and Claims
The paper claims that its primary contribution is providing a detailed tagging scheme, relation-based judgment rules, and an error analysis framework suitable for future validation with real or simulated data. It explicitly states that it does not verify the operational performance of an automatic monitoring system using real field data. Instead, the study demonstrates that:

1. Radio utterances can be treated as structured units for monitoring judgment rather than simple text for keyword detection.
2. A two-layer ontology can successfully derive complex safety situations (hazards, controls, states) from the relations between semantic elements.
3. A "review-hold" mechanism is essential to prevent errors such as over-classifying normal utterances as hazards or misinterpreting confirmation requests as actual incidents.
4. While LLMs can capture approximate meanings of field expressions, they currently lack the stability required for the precise axis separation and code selection necessary for safety-critical decision-making without human oversight or rule-based verification.

The study concludes that the proposed framework effectively structures the intended processing procedure for the virtual log, but the existence of review-hold items highlights the need for further refinement of terminology dictionaries, anaphora rules, and multi-act decomposition rules before deployment in actual field environments.