Mixed-language expressions in construction safety violation and warning reports: A domain-specific framework for...

Area of Science:

• Construction Safety
• Natural Language Processing
• Data Analytics

Background:

• Construction sites generate vast safety data, often hindered by inconsistent terminology and mixed-language expressions (MLEs).
• Korean safety violation and warning reports (SVWRs) present challenges due to irregular formatting and hybrid vocabulary, limiting data-driven safety management.
• Existing methods struggle with the linguistic variability in localized safety reports.

Purpose of the Study:

• To develop and validate a domain-specific text normalization framework for SVWRs.
• To enhance the linguistic consistency and analytical reliability of construction safety reports.
• To enable more effective data-driven safety management through improved text data.

Main Methods:

• Analysis of 64,999 SVWRs from 39 South Korean construction sites.
• Development of a rule- and dictionary-based normalization pipeline to standardize terms and MLEs.
• Application of topic modeling (with symmetric priors) to identify eight safety-related topics.

Main Results:

• Text normalization improved topic model coherence by 20.6% (from 0.412 to 0.497).
• Clarified risk structures in categories like falls, electrical hazards, and fire prevention.
• Revealed previously obscured co-occurring risk patterns, highlighting the importance of linguistic preprocessing.

Conclusions:

• The framework converts fragmented reports into standardized, analyzable data, enhancing methodological reliability and practical use.
• The dictionary-based approach is extensible to other languages, supporting scalable safety management.
• Linguistic preprocessing is crucial for accurate and reliable text-based safety analytics in construction.