SOUTH KOREAResearch
arXiv: Seoul National University's Thunder-DeID+ sharpens de-identification of Korean court judgments
Seoul National University researchers introduced Thunder-DeID+, a system for masking personal and case-specific information in Korean court judgments, which must be redacted before public release. A boundary-aware tokenizer lifted word-level strict typed F1 scores on full judgments from 0.76-0.77 to 0.87-0.88 across four model sizes, and the judgment-trained model reached 0.93 on broader sensitive documents after fine-tuning on 1,000 examples.
Key points
- The team gathered 27,992 judgments from the Ministry of Government Legislation and the courts' Judicial Information Disclosure Portal.arXiv
- After annotation and filtering, the corpus holds 5,872 full judgments with 193,004 annotated personal-information spans.arXiv
- A redesigned three-tier taxonomy captures indirect identifiers, such as workplaces and incident locations, that can re-identify parties.arXiv
- The strongest general open-weight model tested, Qwen 3.6 27B, scored 0.483 and altered the source text in 18.91% of judgments.arXiv
- Korean particles attach to names without spaces, so standard tokenizers often merge personal details with neighboring text.arXiv