Продолжаем научные публикации 🚀
Принята статья к публикации и презентации на конференции IDEAL2026 - 27th International Conference on Intelligent Data Engineering and Automated Learning.
📍 Manchester, UK 📅 19-21.11.2026 🛜 online
▶️ Classification of Pre-Sanitised Citizen Complaints by Cloud LLMs 👥 Pavel Kalinin, Petr Sedelnikov, Leonid Volgin, Dmitriy Fedorov
Automated topic routing of citizen complaints is a high-value public-sector task that is constrained by data-protection law: complaint texts routinely contain personal data, and Russian Federal Law No. 152-FZ forbids transferring such data to external cloud providers, which excludes the most capable commercial large language models (LLMs) from deployment and at the same time impedes reproducibility. We show that de-identification dissolves this tension, and we study it empirically on a real corpus of 24,607 Russian-language complaints over a taxonomy of 743 topics. The work makes three contributions. First, we describe and evaluate a hybrid de-identification pipeline that combines deterministic detectors validated against the official checksums of national identifiers with neural recognition of personal names, and we report span-level precision and recall together with a privacy-oriented character-level recall indicator. Second, we measure the privacy–utility trade-off: running a deployed embedding-centroid classifier on de-identified rather than raw text leaves quality statistically unchanged (Δtop-1 within ±0.7 points across three label-set sizes, with confidence intervals containing zero). Third, we show that de-identification opens a lawful path to cloud LLMs: on de-identified complaints GPT-5.5 attains 90.5% top-3 accuracy over 34 topics, far above the local classifier (44%). We release the de-identification tool, a reproducible synthetic PII benchmark and the evaluation scripts.
