@article{ART003300051},
author={Jung-Kyu Shin and Beak-Cheol Jang},
title={Interpretability for Korean Language Models: Evidence from Attention Visualization},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={1},
pages={41-49}
TY - JOUR
AU - Jung-Kyu Shin
AU - Beak-Cheol Jang
TI - Interpretability for Korean Language Models: Evidence from Attention Visualization
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 1
PB - The Korean Society Of Computer And Information
SP - 41
EP - 49
SN - 1598-849X
AB - This study investigates how the agglutinative nature and morphological complexity of Korean are reflected in language model (LM) internal representations by fine-tuning KLUE RoBERTa Base on the NER task and conducting qualitative and quantitative analyses of attention maps. Our methodology includes a stable training design based on subword–label alignment and masking respecting character-level annotations, attention weight extraction, attention strength visualization, and pattern-specific attention distribution quantification. The analysis reveals three patterns: span-internal cohesion, where entity tokens attend to span boundaries; boundary alignment, where post-entity particles tagged O function as boundary cues; and long-distance dependencies, where distal arguments form semantically coherent links. These findings suggest that Korean linguistic characteristics are structurally organized at the attention layer and head level. This work enhances the interpretability of Korean LMs and establishes a foundation for interpretability research applicable to diverse downstream tasks.
KW - Language Model;Transformer;Attention;Visualization;Korean;Interpretability
DO -
UR -
ER -
Jung-Kyu Shin and Beak-Cheol Jang. (2026). Interpretability for Korean Language Models: Evidence from Attention Visualization. Journal of The Korea Society of Computer and Information, 31(1), 41-49.
Jung-Kyu Shin and Beak-Cheol Jang. 2026, "Interpretability for Korean Language Models: Evidence from Attention Visualization", Journal of The Korea Society of Computer and Information, vol.31, no.1 pp.41-49.
Jung-Kyu Shin, Beak-Cheol Jang "Interpretability for Korean Language Models: Evidence from Attention Visualization" Journal of The Korea Society of Computer and Information 31.1 pp.41-49 (2026) : 41.
Jung-Kyu Shin, Beak-Cheol Jang. Interpretability for Korean Language Models: Evidence from Attention Visualization. 2026; 31(1), 41-49.
Jung-Kyu Shin and Beak-Cheol Jang. "Interpretability for Korean Language Models: Evidence from Attention Visualization" Journal of The Korea Society of Computer and Information 31, no.1 (2026) : 41-49.
Jung-Kyu Shin; Beak-Cheol Jang. Interpretability for Korean Language Models: Evidence from Attention Visualization. Journal of The Korea Society of Computer and Information, 31(1), 41-49.
Jung-Kyu Shin; Beak-Cheol Jang. Interpretability for Korean Language Models: Evidence from Attention Visualization. Journal of The Korea Society of Computer and Information. 2026; 31(1) 41-49.
Jung-Kyu Shin, Beak-Cheol Jang. Interpretability for Korean Language Models: Evidence from Attention Visualization. 2026; 31(1), 41-49.
Jung-Kyu Shin and Beak-Cheol Jang. "Interpretability for Korean Language Models: Evidence from Attention Visualization" Journal of The Korea Society of Computer and Information 31, no.1 (2026) : 41-49.
