It’s been a remarkable few months, and I couldn’t be prouder of my PhD students. Their hard work has paid off with three new publications spanning text anonymization, privacy evaluation, and intent detection in digital communications. Huge congratulations to Gabriel Loiseau and Senaid Popovic: this one is for you. đ
Below is a short summary of each paper, with links to read more.
DIDECO: Detecting Intent in Digital Communications #
Led by Senaid Popovic, DIDECO is the first annotated dataset built specifically to detect both explicit and implicit intents in digital communications. It tackles a real cybersecurity gap with a taxonomy, grounded in Speech Act Theory and persuasion psychology, that separates explicit communicative goals (what is requested) from implicit persuasion mechanisms (how compliance is engineered), covering 20 intent categories. The team annotated 220 LLM-generated spear-phishing emails with a multi-label protocol and six trained annotators, producing 2,162 intent annotations. The analysis shows that sophisticated attacks layer multiple intents, combining explicit goals with implicit persuasion, paving the way for intent-aware detection of social engineering.
🇬🇧 DIDECO: An Annotated Dataset for Intent Detection in Digital Communications. Senaid Popovic, Damien Riquet, Maxime Meyer, Fabien Lauer, Yannick Parmentier. The 15th biennial Language Resources and Evaluation Conference (LREC 2026). [link]
Adaptive Text Anonymization #
Anonymizing text is highly context-sensitive: the right privacy/utility balance shifts with the domain, the privacy objective, and the downstream task. Led by Gabriel Loiseau, this work introduces a framework for task-specific prompt optimization that automatically builds anonymization instructions for language models, adapting to different privacy goals, domains, and usage patterns. Across five datasets, it consistently achieves a better privacy/utility trade-off than existing baselines while staying computationally efficient on open-source models, and it even discovers novel anonymization strategies along the privacy/utility frontier.
🇬🇧 Adaptive Text Anonymization: Learning Privacy-Utility Trade-offs via Prompt Optimization. Gabriel Loiseau, Damien Sileo, Damien Riquet, Maxime Meyer, Marc Tommasi. Findings of the Association for Computational Linguistics: ACL 2026. [link]
Distilling Human-Aligned Privacy Sensitivity Assessment #
Also led by Gabriel Loiseau, this paper distills the privacy-assessment capabilities of Mistral Large 3 (675B) into compact encoder models of roughly 150M parameters. Using a large-scale dataset of privacy-annotated texts spanning 10 diverse domains, the resulting classifiers keep strong agreement with human annotations while dramatically cutting compute, and double as an evaluation metric for de-identification systems.
🇬🇧 Distilling Human-Aligned Privacy Sensitivity Assessment from Large Language Models. Gabriel Loiseau, Damien Sileo, Damien Riquet, Maxime Meyer, Marc Tommasi. Joint Workshop on Legal and Ethical Issues in Human Language Technologies (LEGAL2026) and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (CALD-pseudo 2026). [link]
