RIK's Automated Court-Decision Anonymiser — Estonia's ML Tool for Redacting Personal Data from Nearly 80,000 Rulings

Estonian court decisions must have personal data removed before publication in the public Court Information System, but manually redacting long rulings — some running to more than 50 pages with multiple names, ID codes, birth dates, addresses, phone numbers, IBANs and email addresses — is slow and error-prone. The Centre of Registers and Information Systems (RIK), the government IT agency under Estonia's Ministry of Justice and Digital Affairs, commissioned language-technology company TEXTA OÜ to automate the process.

TEXTA built a hybrid pipeline combining a statistical Conditional Random Fields model, a neural Stanford Stanza named-entity-recognition model, and rule-based regex extractors for structured identifiers such as emails, IBANs and ID codes, trained on a custom-annotated corpus of Estonian court texts. RIK used the tool to anonymise and republish nearly 80,000 historic court decisions in the Court Information System, a backlog that would have been impractical to clear manually.

In TEXTA's own benchmark, the resulting model tagged 99% more personal names and roughly three times more organisation names than Estonia's existing open-source EstNER model on the same court-case dataset. In January 2026, the European Commission's Interoperable Europe Portal featured the tool in its Justice & AI Toolbox collection as a reusable good practice for other EU member states building similar court-transparency systems.

No independent, third-party accuracy audit of the anonymisation model — covering, for example, its miss rate on rare name formats or non-standard layouts — has been published by RIK, TEXTA or the Ministry, and the reported precision/recall figures come from TEXTA's own internal benchmark rather than a court-ordered evaluation, so the residual risk of an unredacted personal detail slipping through is not independently quantified.