Construction sites generate large volumes of textual safety data, yet inconsistent terminology and mixed-language expressions (MLEs) reduce the reliability of analysis. Korean safety violation and warning reports (SVWRs), a localized form of safety observation reports, are often written with irregular spacing, abbreviations, and hybrid vocabulary, hindering systematic utilization for data-driven safety management.
Objective
This study aims to develop and validate a domain-specific text normalization framework to improve the linguistic consistency and analytical reliability of SVWRs.
Methods
A dataset of 64,999 SVWRs collected from 39 construction sites in South Korea was analyzed. A rule- and dictionary-based normalization pipeline was designed to unify fragmented terms and standardize MLEs. Topic modeling was conducted using topic modeling with symmetric priors and eight topics aligned with national safety categories.
Results
Normalization increased topic-model coherence from 0.412 to 0.497 (20.6% improvement), clarifying risk structures across categories such as falls, electrical hazards, and fire prevention. It revealed co-occurring risk patterns previously obscured by inconsistent language use, demonstrating that linguistic preprocessing is crucial for reliable text-based safety analytics.
Conclusions
The proposed framework enhances both methodological reliability and practical applicability by converting fragmented field reports into standardized, analyzable data. Its dictionary-based architecture can be extended to other agglutinative or multilingual languages, supporting scalable and data-driven safety management in the construction industry.
CameronIHareBDuffR. An analysis of safety advisor roles and site safety performance. Eng Constr Archit Manag2013; 20: 505–521.
2.
AdnanHYussofFNMJaafarFW, et al.Safety manager competencies in managing construction projects in Malaysia. IOP Conf Ser Earth Environ Sci2019; 385: 012057.
3.
Occupational Safety and Health Research Institute. Statistical survey and analysis of industrial occupational accidents. Ulsan (KR): Korea Occupational Safety and Health Agency, 2023.
4.
KimHYYiJSJangYE. Analyzing patterns of multi-cause accidents from KOSHA’s construction injury case reports utilizing text mining methodology. J Archit Inst Korea2022; 38: 237–244.
5.
OswaldDSherrattFSmithS. Problems with safety observation reporting: a construction industry case study. Saf Sci2018; 107: 35–45.
6.
GadekarHBugaliaN. Automatic classification of construction safety reports using semi-supervised YAKE-Guided LDA approach. Adv Eng Inf2023; 56: 101929.
7.
TixierAHallowellMRajagopalanB, et al.Construction safety clash detection: identifying safety incompatibilities among fundamental attributes using data mining. Autom Constr2017; 74: 39–54.
8.
HughesPShippDFigueres-EstebanM, et al.From free-text to structured safety management: introduction of a semi-automated classification method of railway hazard reports to elements on a bow-tie diagram. Saf Sci2018; 110: 11–19.
9.
AhadhABinishGVSrinivasanR. Text mining of accident reports using semi-supervised keyword extraction and topic modeling. Process Saf Environ Prot2021; 155: 455–465.
10.
MaZChenZ. Mining construction accident reports via unsupervised NLP and Accimap for systemic risk analysis. Autom Constr2024; 161: 105343.
11.
GohYMUbeynarayanaCU. Construction accident narrative classification: an evaluation of text mining techniques. Accid Anal Prev2017; 108: 122–130.
12.
BengfortBBilbroROjedaT. Applied text analysis with Python: Enabling language-aware data products with machine learning. 1st ed. Sebastopol, CA: O’Reilly Media, Inc., 2018, https://doi.org/10.5555/3285754.
ChoiHJ. An analysis on foreign workers’ dangerous work behavior factors in the construction sites. J Stand Certif Saf2022; 12: 15–28.
16.
KwonMSKimLN. Analysis of affective domain trends in mathematics education using dynamic topic modeling. School Maths2024; 26: 445–468.
17.
BakerHHallowellMTixierA. Automatically learning construction injury precursors from text. Autom Constr2020; 118: 103145.
18.
ZouPSunindiioR. Strategic safety management in construction and engineering. Hoboken, NJ: John Wiley & Sons, 2015, https://doi.org/10.1002/9781118839362.
19.
VermaAMaitiJ. Text-document clustering-based cause and effect analysis methodology for steel plant incident data. Int J Inj Contr Saf Promot2018; 25: 416–426.
20.
JingSLiuXGongX, et al.Correlation analysis and text classification of chemical accident cases based on word embedding. Process Saf Environ Prot2020; 158: 698–710.
21.
LinJRHuZZLiJL, et al.Understanding on-site inspection of construction projects based on keyword extraction and topic modeling. IEEE Access2020; 8: 198503–17.
22.
CaoXLuRGuoL, et al.Construction health and safety: a topic landscape study. Organ Technol Manag Constr Int J2021; 13: 2472–2483.
23.
RobinsonSDIrwinWJKellyTK, et al.Application of machine learning to mapping primary causal factors in self reported safety narratives. Saf Sci2015; 75: 118–129.
24.
DasMAlphonsePJA. A comparative study on tf-idf feature weighting method and its analysis using unstructured dataset. arXiv:2308.04037 [Preprint]. 2023 [cited 2025 Feb 10], https://arxiv.org/abs/2308.04037.
25.
PenningtonJSocherRManningC. GloVe: Global vectors for word representation. Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014, pp.1532–1543, https://doi.org/10.3115/v1/D14-1162.
26.
MikolovTChenKCorradoG, et al.Efficient estimation of word representations in vector space. arXiv:1301.3781 [Preprint]. 2013 [cited 2025 Sep 12], https://doi.org/10.48550/arXiv.1301.3781.
27.
MoonSHChungSHChiSH. Bridge damage recognition from inspection reports using NER based on recurrent neural network with active learning. J Perform Constr Facil2020; 34: 04020119.
28.
ZhangF. A hybrid structured deep neural network with Word2Vec for construction accident causes classification. Int J Constr Manag2022; 22: 1120–1140.
29.
ChoiSHKimTUSeolJS, et al.A syllable-based technique for word embeddings of Korean words. arXiv:1708.01766 [Preprint]. 2017 [cited 2025 Oct 20], https://arxiv.org/abs/1708.01766v1.
30.
ChanHS. Constructional borrowing from English in Hong Kong Cantonese. Front Commun2022; 7: 796372.
31.
GongSHShimJWHyeonMH, et al.Auto-correction with Levenshtein (edit) distance algorithm to spell error in legal terms. J Korea Contents Assoc2024; 24: 82–93.
32.
KimHJChoSZKangPS. KR-WordRank: an unsupervised Korean word extraction method based on WordRank. J Korean Inst Indus Eng2014; 40: 18–33.
33.
KilHH. The study of Korean stopwords list for text mining. URIMALGEUL: Korean Lang Literat2018; 78: 1–25.
34.
ParkMHChoiHRLeeHC. Study on preprocessing method suitable for Korean aspect extraction based on unsupervised learning: for childcare products reviews. J Korean Inst Indus Eng2021; 47: 56–67.
35.
ZengL. Leveraging large language models for code-mixed data augmentation in sentiment analysis. In Proceedings of the second workshop on social influence in conversations (SICon 2024); 2024 Nov; Miami, Florida, USA. Association for Computational Linguistics, 2024, pp.85–101. https://doi.org/10.18653/v1/2024.sicon-1.6.
36.
PanM. Patterns of code-switching in Mandarin Chinese and English and why does it happen. Commun Humanit Res2025; 70: 110–116.
37.
Treffers-DallerJ. The simple view of borrowing and code-switching. Int J Biling2023; 29: 347–370.
38.
ParkCEKimJA. Robust multi-task learning-based Korean POS tagging to overcome word spacing errors. ACM Trans Asian Low-Resource Lang Inform Processing2023; 22: 1–13.
39.
XuXZouP. Discovery of new safety knowledge from mining large injury dataset in construction. Saf Sci2021; 144: 105481.
40.
ChangTYChiSHImSB. Understanding user experience and satisfaction with urban infrastructure through text mining of civil complaint data. J Constr Eng Manag2022; 148: 04022061.
41.
LimJSHanSGKangYC, et al.Affinity analysis between factors of fatal occupational accidents in construction using data mining techniques. Korean J Constr Eng Manag2021; 22: 29–38.
42.
XuNMaLWangL, et al.Extracting domain knowledge elements of construction safety management: rule-based approach using Chinese natural language processing. J Manag Eng2021; 37: 04021001.
43.
DoQLeTLeC. Uncovering critical causes of highway work zone accidents using unsupervised machine learning and social network analysis. J Constr Eng Manag2023; 150: 04023168.
44.
KangHASongHJ. Constructing sentiment lexicon for subject-specific sentiment analysis. Korean Linguist2021; 93: 83–110.
ParkTMahamaduAMAgyekumK, et al.An inquiry into the health and safety management practices of construction firms in South Korea. J Eng Design Technol2025; 23: 345–367.
48.
LimSJOhARWonJH, et al.Improvement of inspection system for reduction of small-scale construction site accident in Korea. Ind Health2018; 56: 466–474.
49.
DaleAMColvinRBarreraM, et al.The association between subcontractor safety management programs and worker perceived safety climate in commercial construction projects. J Saf Res2020; 74: 279–288.
RöderMBothAHinneburgA. Exploring the space of topic coherence measures. In Proceedings of the eighth ACM international conference on web search and data mining (WSDM ‘15). 2015 Feb, pp. 399–408. https://doi.org/10.1145/2684822.2685324.
54.
MaatenLHintonG. Visualizing data using t-SNE. J Mach Learn Res2008; 9: 2579–2605.
55.
ChengLImGHDoeC, et al.Identifying English language use and communication challenges facing “entry-level” workplace immigrants in Canada. J Int Migr Integr2021; 22: 865–886.
56.
AkihikoY. Word embedding-based semantic analysis of English loanwords in Japanese and Korean. Dissertation, Seoul National University, Seoul (KR), 2021.
57.
De Jesus-RivasMConlonHABurnsC. The impact of language and culture diversity in occupational safety. Workplace Health Saf2016; 64: 24–27.
