All Issue

2026 Vol.44, Issue 1 Preview Page

Article

A Comparative Study on AI-Based Pseudonymization Methods for Safe Utilization of Transportation Complaint Texts

교통 민원 텍스트의 안전한 활용을 위한 AI 기반의 가명처리 비교 연구

Keunwook KIM1
,
Younshik CHUNG2*
김 건욱1
,
정 연식2*
1Director, Big Data Utilization Center, Daegu Digital Innovation Promotion Agency, Daegu 42250, Korea
2Professor, Department of Urban Planning and Engineering, Yeungnam University, Gyeongsan 38541, Korea
1대구디지털혁신진흥원 빅데이터활용센터 센터장
2영남대학교 도시공학과 교수
*Corresponding Author
28 February 2026. pp. 99-113
PDF XML Citation
Abstract

Local governments increasingly rely on civil complaint data to inform urban policymaking; however, the presence of personal information and offensive language significantly restricts data sharing and public disclosure. Although pseudonymization techniques based on regular expressions or named entity recognition (NER) are widely adopted, they often inadequately preserve contextual meaning and fail to balance data utility with privacy protection. This study proposes a secure and practically applicable pseudonymization framework and conducts a quantitative comparison of alternative approaches. The analysis is based on 4,889 traffic-related civil complaints collected in Daegu Metropolitan City between December 2023 and March 2025. Four methods are evaluated: NER-based substitution, synthetic text generation using a multilingual language model (Qwen 2.5), synthetic text generation using a Korean-specific language model (EXAONE 3.5), and an “intentional recombination” approach that reconstructs complaint texts by reorganizing key nouns and verbs according to the complaint title as an explicit representation of intent. The results indicate that while the NER-based method most closely preserves the original textual structure and word distribution, it exhibits a comparatively high residual rate of personal information and profanity. By contrast, the Korean language model–based intentional recombination approach achieves the lowest residual rate (5.27%) while maintaining approximately 94% semantic similarity to the original texts. Multilingual language models demonstrate limitations for practical deployment due to contextual distortion and unintended foreign language mixing. These findings suggest that Korean-specific language models combined with intent-based reconstruction provide an effective foundation for the secure utilization of traffic complaint data in local governments and highlight the need for further research to generalize this framework across other urban policy domains.

Keywords
complaint text
large language models
personal information protection
pseudonymization
synthetic data

도시 내 사회문제를 해결하기 위해 대부분의 지자체는 민원 데이터를 정책 근거로 활용하고 있으나, 민원 텍스트에는 주소·차량번호·이름 등 개인정보와 공격적 표현, 비속어가 함께 포함되어 있어 제3자 제공이나 개방에 한계가 있다. 이에 따라 정규표현식이나 개체명 인식 모형을 활용한 텍스트 가명처리가 널리 사용되고 있으나, 문맥을 충분히 반영하지 못하거나 데이터 유용성과 개인정보 보호 간 균형을 맞추는 데 어려움이 존재한다. 본 연구는 공간적으로 대구광역시, 시간적으로 2023년 12월부터 2025년 3월까지 수집된 교통 민원 텍스트 4,889건을 대상으로 현업에서 활용 가능한 안전한 가명처리 방법을 제시하고, 이를 정량적으로 비교·평가하고자 한다. 비교 대상 모형은 ① 개체명 인식(NER) 기반 치환 모델, ② 다국어 언어모델 기반 합성데이터 생성(Qwen 2.5), ③ 한국어 특화 언어모델 기반 합성데이터 생성(EXAONE 3.5), ④ 민원 제목을 ‘의도’로 간주하고 본문 내 핵심 명사·동사를 재조합하여 문장을 재작성하는 ‘의도 재조합’ 기반 한국어 언어모델(EXAONE 3.5)이다. 데이터 유용성과 개인정보 보호 관점에서 분석한 결과, 개체명 인식 방법은 원문 구조와 단어 분포를 가장 유사하게 유지하였으나, 개인정보 및 비속어 잔여 비율이 상대적으로 높게 나타났다. 반면 제안한 한국어 언어모델 기반 ‘의도 재조합’ 방식은 잔여 개인정보·비속어 비율이 5.27%로 가장 낮게 나타났고, 원문 문장의 의미 유사도도 약 94% 수준으로 유지되는 것을 확인하였다. 다국어 언어모델의 경우 문맥 왜곡과 외국어(중국어) 혼합 등이 발생하여 실제 현업에 적용하기에는 한계가 있는 것으로 나타났으며, 이러한 결과는 한국어 특화 언어모델과 의도 기반 재구성 접근이 지자체 교통 민원 데이터의 안전한 데이터 활용에 있어 주요한 기초 연구가 될 것으로 판단되며, 향후 도시내 다양한 분야로 일반화되기 위해 후속 연구의 필요성도 함께 제시하였다.

키워드
민원 텍스트
대규모 언어 모델
개인정보 보호
가명처리
합성데이터
References
1

An S., Bae K., Choi E., Choi S. J., Choi Y., Hong S. et al. (2024), EXAONE 3.0 7.8B Instruction Tuned Language Model, arXiv e-prints, arXiv-2408.

2

Asimopoulos D., Siniosoglou I., Argyriou V., Karamitsou T., Fountoukidis E., Goudos S. K., et al. (2024), Benchmarking Advanced Text Anonymisation Methods: A Comparative Study on Novel and Traditional Approaches, In 2024 13th International Conference on Modern Circuits and Systems Technologies (MOCAST), 1-6, IEEE.

10.1109/MOCAST61810.2024.10615642
3

Bae K., Choi E., Choi K., Choi S. J., Choi Y., Hong, S., et al. (2025), EXAONE Deep: Reasoning Enhanced Language Models, arXiv e-prints, arXiv-2503.

10.1158/1538-7445.AM2025-2503
4

Bhati B. S., Ivanchev, J., Bojic, I., Datta, A., Eckhoff, D. (2021), Utility-Driven K-Anonymization of Public Transport User Data. IEEE Access, 9, 23608-23623.

10.1109/ACCESS.2021.3055505
5

Chen H., Xiong C., Xie J. M., Cai M. (2020), Privacy Protection Method for Vehicle Trajectory Based on VLPR Data, J. Adv. Transp., 2020(1), 6026140.

10.1155/2020/6026140
6

Chibelushi C., Sharp B., Salter A. (2005), Transcript Segmentation Using Utterance Cosine Similarity Measure, In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science (NLUCS 2005), 78-90.

10.5220/0002560900780090
7

Conneau A., Khandelwal K., Goyal N., Chaudhary V., Wenzek G., Guzmán F., et al. (2019), Unsupervised Cross-Lingual Representation Learning at Scale, arXiv preprint, arXiv:1911.02116.

10.18653/v1/2020.acl-main.747
8

Ghasemzadeh M., Fung B. C., Chen R., Awasthi A. (2014), Anonymizing Trajectory Data for Passenger Flow Analysis, Transp. Res. Part C: Emerg. Technol., 39, 63-79.

10.1016/j.trc.2013.12.003
9

Hwang H. (2020), Issues and Challenges in the Use of Pseudonymized Information: Focusing on the Use of Pseudonymized Sensitive Data, Korea Insurance Research Institute.

10

Jang S., Cho Y., Seong H., Kim T., Woo H. (2024), The Development of a Named Entity Recognizer for Detecting Personal Information Using a Korean Pretrained Language Model, Appl. Sci., 14(13), 5682.

10.3390/app14135682
11

Kim B., Yoo M., Park K. C., Lee K. R., Kim J. H. (2021), A Value of Civic Voices for Smart City: A Big Data Analysis of Civic Queries Posed by Seoul Citizens, Cities, 108, 102941.

10.1016/j.cities.2020.102941
12

Kim H. T., Jang G. Y. (2023), Current and Alternative Methods of Data Alias Processing (Pseudonymization) Techniques, KIRI Research Report, 2023(7), Korea Insurance Research Institute, 1-85.

13

Kim J. S. (2024), Development and Evaluation of De-Identification Pipeline for Korean Clinical Notes Using Natural Language Processing, Master's Thesis, Seoul National University, Seoul, South Korea.

14

Kim M., Lee S. (2014), Measures of Abnormal User Activities in Online Comments Based on Cosine Similarity, J. Korea Inst. Inf. Secur. Cryptol., 24(2), 335-343.

10.13089/JKIISC.2014.24.2.335
15

Korea Institute of Science and Technology Information (KISTI) (2017), Final Report on the Field Application Project for Creation and Distribution of De-Identified Personal Information.

16

Lee H. G., Yoo G. D. (2024), Integrated Data Safe Zone Prototype for Efficient Processing and Utilization of Pseudonymous Information in the Transportation Sector, J. Korean Soc. Intell. Transp. Syst., 23(3), 48-66.

10.12815/kits.2024.23.3.48
17

Lee H. S., Song J. H. (2016), A Research on De-Identification Technique for Personal Identifiable Information, Software Policy & Research Institute, 8.

18

Ministry of the Interior and Safety (2025), 2025 Public Data Provision and Data-Driven Administration Evaluation Guidelines.

19

Nahid M. M. H., Hasan S. B. (2024), SafeSynthDP: Leveraging Large Language Models for Privacy-Preserving Synthetic Data Generation Using Differential Privacy, arXiv preprint, arXiv:2412.20641.

10.32388/PJIL3E
20

Nergiz M. E., Atzori M., Saygin Y. (2008), Towards Trajectory Anonymization: A Generalization-Based Approach, In Proceedings of the SIGSPATIAL ACM GIS 2008 International Workshop on Security and Privacy in GIS and LBS, 52-61.

10.1145/1503402.1503413
21

Park G. C. (2020), Big Data Analysis of Civil Complaints Using Text Mining Techniques: Gangnam-gu Office, Seoul Digital Foundation.

22

Park J. H., Kang M. G. (2024), Spatial Information Extraction and Basic Analysis from 120 Dasan Call Civil Complaint Texts Through Named Entity Recognition Modeling, J. Korea Plan. Assoc., 59(7), 169-180.

10.17208/jkpa.2024.12.59.7.169
23

Personal Information Protection Commission, Korea Internet & Security Agency (KISA) (2024), Guideline for Pseudonym Information Processing.

24

Rakhimzhanov D., Belginova S., Yedilkhan D. (2025), Automated Classification of Public Transport Complaints via Text Mining Using LLMs and Embeddings. Infor., 16(8), 644.

10.3390/info16080644
25

Sang E. F., De Meulder F. (2003), Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition, arXiv preprint, cs/0306050.

26

Seo D. K., Kim K. W., Kim J. Y., Lee D. H. (2020), Personal Information Detection and De-Identification System Using Sentence Intent Classification and Named Entity Recognition, KIPS Annual Conference Proceedings, 27(2), 1018-1021.

27

Shafaeipour N., Stanciu V. D., van Steen M., Wang M. (2024), Understanding the Protection of Privacy when Counting Subway Travelers Through Anonymization, Comput. Environ. Urban Syst., 110, 102091.

10.1016/j.compenvurbsys.2024.102091
28

Thetbanthad P., Sathanarugsawait B., Praneetpolgrang P. (2025), Automated Redaction of Personally Identifiable Information on Drug Labels Using Optical Character Recognition and Large Language Models for Compliance with Thailand’s Personal Data Protection Act, Appl. Sci., 15(9), 4923.

10.3390/app15094923
29

Vakili T., Henriksson A., Dalianis H. (2024), End-to-end Pseudonymization of Fine-Tuned Clinical BERT Models: Privacy Preservation with Maintained Data Utility, BMC Med. Inform. Decis. Mak., 24(1), 162.

10.1186/s12911-024-02546-838915012PMC11197357
30

Vaswani A., Shazeer N., Parmar N., Uszkoreit J., Jones L., Gomez A. N., et al. (2017), Attention is All You Need, In Advances in Neural Information Processing Systems (Vol. 30).

31

Walter M., Beskorovajnov W., Lieberwirth F., Sürmeli J., Zwick P., Heinrich R. (2023), Mobility Data Anonymization - A Literature Review and an Industry-Driven Survey, Karlsruhe Reports in Informatics, 2023(3). Karlsruhe Institute of Technology (KIT).

32

Wandelt S., Zheng C., Wang S., Liu Y., Sun X. (2024), Large Language Models for Intelligent Transportation: A Review of the State of the Art and Challenges. Appl. Sci., 14(17), 7455.

10.3390/app14177455
33

Wang S., Sun X., Li X., Ouyang R., Wu F., Zhang T., et al. (2025), GPT-NER: Named Entity Recognition via Large Language Models, In Findings of the Association for Computational Linguistics: NAACL 2025, 4257-4275.

10.18653/v1/2025.findings-naacl.239
34

Won J., Yoo S. (2016), Pattern Analysis for Civil Complaints of Local Governments Using a Text Mining, J. Korean Soc. Surv. Geod. Photogramm. Cartogr., 34(3), 319-327.

10.7848/ksgpc.2016.34.3.319
35

Yermilov O., Raheja V., Chernodub A. (2023), Privacy- and Utility-Preserving NLP with Anonymized Data: A Case Study of Pseudonymization, arXiv preprint, arXiv:2306.05561.

10.18653/v1/2023.trustnlp-1.20
Information
  • Publisher :Korean Society of Transportation
  • Publisher(Ko) :대한교통학회
  • Journal Title :Journal of Korean Society of Transportation
  • Journal Title(Ko) :대한교통학회지
  • Volume : 44
  • No :1
  • Pages :99-113
  • Received Date : 2025-11-24
  • Revised Date : 2025-12-17
  • Accepted Date : 2025-12-23
  • DOI :https://doi.org/10.7470/jkst.2026.44.1.099
Journal Informaiton Journal of Korean Society of Transportation Journal of Korean Society of Transportation
Archives
: 2015 Vol.34~
  • NRF
  • KOFST
  • crossref cited-by
  • crosscheck
  • orcid
  • open access
  • ccl
Journal Informaiton Journal Informaiton - close