Technology: combining AI and people

The system recognizes more than 30 types of entities using a hybrid approach combining named entity recognition and NLP adapted to Slovak. The foundation was created by manually annotating approximately 500–1000 documents, on which the first model was trained, and it continues to improve on an ongoing basis. Anonymization is semi-automated: AI does most of the work and a human verifies it, with intervention needed in only about 40 % of documents. Although accuracy does not yet match a human’s, AI is consistent even on long texts, which helps unify practice regarding what is considered personal data.

Results in practice and what’s next

The average time for anonymization has fallen from tens of minutes to approximately five minutes per decision. The user interface built on a unified design manual and React also helps: it offers, for example, back/forward (Ctrl+Z), continuous saving, and resilience to connection outages. Thanks to clear navigation and the presentation of marked data, output quality increases and the risk of information leakage decreases.

The system has already processed tens of thousands of decisions, and almost all courts are gradually getting involved, including the new administrative courts. Looking ahead, the use of transformers and large language models is being considered for more accurate entity recognition and anonymization. Another step could be an intelligent editor for drafting decisions and recommending similar case law, so that the work of judges and judicial staff is more efficient.