Show document

A+ | A- | | SLO | ENG

Title: Predobdelava podatkov za zagotavljanje varnosti in zasebnosti pri uporabi velikih jezikovnih modelov v gradbeništvu Authors: ID Brelih, Anja (Author) ID Srdič, Aleksander (Author) ID Dujc, Jaka (Author) ID Klinc, Robert (Author) Files: PDF - Presentation file, download (1,08 MB) MD5: F589CAA19550C104819031603397FC61   URL - Source URL, visit https://www.zveza-dgits.si/gradbeni-vestnik-dec-2025/   Language: Slovenian Typology: 1.01 - Original Scientific Article Organization: ZDGITS - Union of associations of Slovenian civil engineers and technicians Abstract: Prispevek predstavlja izzive zagotavljanja varstva podatkov pri uporabi velikih jezikovnih modelov (VJM) v delovnih tokovih operativnega gradbeništva. Analizira, kako uspešno obstoječa orodja za prepoznavanje imenskih entitet (angl. Named Entity Recognition, NER) zaznajo in anonimizirajo občutljive informacije v tehničnih gradbenih dokumentih, zlasti v slovenskem jeziku. Opravljena je bila kvalitativna evalvacija štirih ogrodij za obdelavo naravnega jezika (SpaCy, SpaCy SLO, Flair, NLTK), ki so bile preizkušene na vzorcu petih dejanskih gradbenih dokumentov in primerjane z ročno anotiranimi referenčnimi podatki. V evalvacijo je bila vključena tudi anonimizacija z VJM, ki je občutljive podatke zakrival z uporabo regularnih izrazov. Rezultati kažejo, da je osnovna anonimizacija sicer mogoča, vendar vsa klasična ogrodja NER slabše prepoznavajo entitete specifične za področje, kot so projektne šifre, inženirski nazivi ter strukturirani šte vilčni podatki. Ugotovitve kažejo na potrebe po prilagojenih orodjih za predobdelavo, saj netočna anonimizacija predstavlja pravna in etična tveganja pri vključevanju VJM v regulirane panoge, kot je gradbeništvo. Prihodnje raziskave se morajo osredotočiti na gradnjo hibridnih anonimizacijskih tokov in učenje modelov na anotiranih podatkih, da bi izboljšali natančnost in skladnost v tehničnih panogah. Keywords: veliki jezikovni modeli, zasebnost podatkov, prepoznavanje imenskih entitet, operativno gradbeništvo, predobdelava dokumentov Publication status: Published Publication version: Version of Record Year of publishing: 2025 Number of pages: str. 210-219 Numbering: Letn. 74 PID: 20.500.12556/DiRROS-27357  UDC: 004.434:004.8:624 ISSN on article: 0017-2774 COBISS.SI-ID: 262447363  Publication date in DiRROS: 03.02.2026 Views: 96 Downloads: 66 Metadata:    Cite this work Plain text BibTeX EndNote XML EndNote/Refer RIS ABNT ACM Ref AMA APA Chicago 17th Author-Date Harvard IEEE ISO 690 MLA Vancouver : Copy citation

Share: Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Record is a part of a journal

Title:	Gradbeni vestnik : glasilo Zveze društev gradbenih inženirjev in tehnikov Slovenije
Shortened title:	Gradb. vestn.
Publisher:	Zveza društev gradbenih inženirjev in tehnikov Slovenije
ISSN:	0017-2774
COBISS.SI-ID:	859140

Document is financed by a project

Funder:	ARIS - Slovenian Research and Innovation Agency
Project number:	P2-0210-2019
Name:	E-Gradbeništvo

Licences

License:	CC BY-SA 4.0, Creative Commons Attribution-ShareAlike 4.0 International
Link:	http://creativecommons.org/licenses/by-sa/4.0/
Description:	This Creative Commons license is very similar to the regular Attribution license, but requires the release of all derivative works under this same license.

Secondary language

Language:	English
Title:	Data preprocessing to ensure security and privacy when using large language models in construction
Abstract:	This paper addresses the challenge of ensuring data privacy when using Large Language Models (LLMs) in Construction Management Workflows. It analyses how well existing Named Entity Recognition (NER) tools can identify and redact sensitive information in technical construction documents, particularly in the Slovenian language. A qualitative evaluation was performed with four NLP frameworks (SpaCy, SpaCy SLO, Flair, NLTK) applied to a sample of five real-world construction documents and compared with manually annotated baseline data. The evaluation also included anonymization with VJM, which masked sensitive data using regular expressions. The results show that while basic anonymisation is possible, all classical NER frameworks underperform in identifying domain-specific entities such as project codes, engineering titles and structured numerical data. These findings emphasise the urgent need for domain-adapted preprocessing tools, as inaccurate redaction po ses legal and ethical risks when integrating LLMs in regulated domains such as construction. Future work should focus on building hybrid redaction pipelines and training custom models on annotated corpora to improve accuracy and compliance in technical domains.
Keywords:	large language models, data privacy, name entity recognition, construction management, document preprocessing

Collection

This document is a part of these collections:

1. Gradbeni vestnik

Back