A new paper introduces Thunder-DeID+, a de-identification model developed for Korean court judgments, and tests whether it can be applied to a wider range of sensitive documents. The authors assess the model's transferability, paying particular attention to how well it handles boundary precision, label granularity, and recall when moving beyond its original domain.

The study is motivated by the need to protect personal information in legal texts and other sensitive materials. Because Korean is an agglutinative language, where words are formed by chaining morphemes, identifying and masking personal identifiers requires careful boundary detection. The paper's focus on boundary precision suggests that errors in splitting or merging tokens could undermine de-identification performance.

The abstract does not report specific results, but it frames the evaluation around three dimensions: how precisely the model identifies the start and end of sensitive spans, how finely it distinguishes between different types of labels, and how completely it recovers all relevant instances. The authors position Thunder-DeID+ as a potential tool for extending de-identification beyond court records to other domains, though the outcome of that transfer remains the central question of the study.