Research map / Large language models / Retrieval and knowledge
HanjaBridge: Resolving Semantic Ambiguity in Korean LLMs via Hanja-Augmented Pre-Training
Seungho Choi, Sihyun Park, Minsang Kim, Chansol Park, Bongsu Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Hanja candidates (k=8) on top of distillation-only CPT (ablation) | Qwen2.5-3B CPT · KoBALT-Hard | avg accuracy: 0.1028 -> 0.1437 | Improved |
| add Hanja candidates (k=8) on top of distillation-only CPT (ablation) | Qwen2.5-3B CPT · KoBEST-General | avg accuracy: 0.5690 -> 0.6354 | Improved |
| add token-level distillation to CPT (ablation) | Qwen2.5-3B CPT · English average | avg accuracy: 0.6501 -> 0.6753 | Improved |
| add Hanja candidates (k=8) to distillation CPT (ablation) | Qwen2.5-3B CPT · English average | avg accuracy: 0.6753 -> 0.6792 | Improved |
| increase number of Hanja candidates from 8 to 16 (ablation) | HanjaBridge · KoBALT-Hard | avg accuracy: 0.1437 -> 0.1322 | Hurt |
| increase number of Hanja candidates from 2 to 8 (ablation) | HanjaBridge · KoBALT-Hard | avg accuracy: 0.0983 -> 0.1437 | Improved |
| plain full CPT on Korean data without distillation or Hanja | Qwen2.5-3B · KoBALT-Hard | avg accuracy: 0.1189 -> 0.1115 | Hurt |
| plain full CPT on Korean data without distillation or Hanja | Qwen2.5-3B · English average | avg accuracy: 0.6976 -> 0.6501 | Hurt |
| use Hanja candidates at inference in addition to training (ablation) | HanjaBridge · Hanja multiple-choice probe | accuracy: 0.568 -> 0.575 | No clear effect |
| train with HanjaBridge, Korean-only at inference, versus Full CPT (ablation) | Qwen2.5-3B · Hanja multiple-choice probe | accuracy: 0.389 -> 0.568 | Improved |
| longer training with Hanja augmentation | HanjaBridge k=8 · Hanja attention probe | correct-candidate attention accuracy: 0.102 -> 0.308 | Improved |
Closest papers by meaning
- Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought (ICLR 2026 · similarity 0.84)
- Soft Token Alignment for Cross-Lingual Reasoning (NeurIPS 2026 · similarity 0.84)
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.84)
- MKGL: Mastery of a Three-Word Language (NeurIPS 2024 · similarity 0.84)
- KaSA: Knowledge-Aware Singular-Value Adaptation of Large Language Models (ICLR 2025 · similarity 0.83)
- In-Context Editing: Learning Knowledge from Self-Induced Distributions (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
