Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models

Gabriel Loiseau, Damien Sileo, Damien Riquet, Maxime Meyer, Marc Tommasi


Abstract
Accurate privacy evaluation of textual data remains a critical challenge in privacy-preserving NLP. Recent work has shown that LLMs can serve as reliable privacy evaluators, achieving strong agreement with human judgments; however, their computational cost and impracticality for processing sensitive data at scale limit real-world deployment. We address this gap by distilling the privacy assessment capabilities of Mistral Large 3 (675B) into lightweight encoder models with as few as 150M parameters. Leveraging a large-scale dataset of privacy-annotated texts spanning 10 diverse domains, we train efficient classifiers that preserve strong agreement with human annotations while dramatically reducing computational requirements. We validate our approach on human-annotated test data and demonstrate its practical utility as an evaluation metric for de-identification systems.
Anthology ID:
2026.legal-1.6
Volume:
Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (LEGAL2026 and CALD-pseudo 2026) @ LREC 2026
Month:
May
Year:
2026
Address:
Palma, Mallorca (Spain)
Editors:
Ingo Siegert, Maria Irena Szawerna, Khalid Choukri, Simon Dobnik, Paweł Kamocki, Therese Lindström Tiedemann, Pierre Lison, Ricardo Muñoz Sánchez, Ildikó Pilán, Lisa Södergård, Kossay Talmoudi, Elena Volodina, Xuan-Son Vu
Venues:
LEGAL | CALD-pseudo | WS
SIG:
Publisher:
ELRA
Note:
Pages:
53–61
Language:
External URL:
https://lrec.elra.info/lrec2026-ws-legal-06
DOI:
10.63317/3kv2f8gmz5rb
Bibkey:
Cite (ACL):
Gabriel Loiseau, Damien Sileo, Damien Riquet, Maxime Meyer, and Marc Tommasi. 2026. Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models. In Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (LEGAL2026 and CALD-pseudo 2026) @ LREC 2026, pages 53–61, Palma, Mallorca (Spain). ELRA.
Cite (Informal):
Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models (Loiseau et al., LEGAL-CALD-pseudo 2026)
Copy Citation: