Korean AI That Preserves Court Ruling Context While Masking Only Personal Data

According to AI TIMES, a research team led by Professor Lee Jae-jin (이재진) at Seoul National University unveiled on September 29 a de-identification model called Thunder-DeID+ that precisely masks only personal information while preserving as much context as possible in Korean documents. Trained on 5,872 court rulings, the model raised the separation rate for personal-information boundaries to 99.7% and focused on more accurately identifying expressions attached to Korean particles and endings. The core of the technology is to more finely distinguish where personal information begins and ends, instead of deleting the particles attached to names or institution names as well, so that only the necessary parts are processed. The research team said that after training on court rulings, the model also showed the potential to be applied to general sensitive documents with only a small amount of additional training. It aims to expand its use beyond legal documents to areas such as administrative documents and corporate work documents, where de-identification is needed before external sharing or AI input. Even with improved automation performance, the approach assumes that people will review parts that are highly risky or difficult to judge.