본문 바로가기
  • Home

Injecting Hanja Radical Structure into Korean BERT through LoRA-based Contrastive Learning

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(7), pp.45~54
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : May 6, 2026
  • Accepted : June 17, 2026
  • Published : July 31, 2026

Euna Song 1 Beak-Cheol Jang 1

1연세대학교

Accredited

ABSTRACT

Modern Korean vocabulary comprises approximately 66–70% Sino-Korean words (漢字語)[1] whose sub-character semantic structures encoded in Hanja radicals (部首) are not explicitly modeled by existing Korean pre-trained language models. We propose a parameter-efficient fine-tuning framework applying LoRA to KLUE BERT-base with InfoNCE contrastive loss on 19,396 word pairs from 3,128 Sino-Korean words (0.98% of total parameters updated), using Hanja radicals as supervision signals. The proposed method achieves k-NN Acc@1 of 0.7007 on a hold-out structural generalization task— outperforming all baselines including KoSimCSE (0.5758)—and WSD accuracy of 0.8402 on the auxiliary task, demonstrating that Hanja radical structure serves as an effective supervision signal for semantic clustering in Korean embedding spaces.

Citation status

* References for papers published after 2025 are currently being built.