한국어의 언어적 특성을 반영한 이종 그래프 기반의 단문 분류 모델

초록

Short text classification poses inherent challenges due to insufficient context and irregular sentence structures. The agglutinative nature and flexible word order of the Korean language further complicate the analysis of semantic units. Consequently, research focused specifically on short text classification in Korean has been quite limited. In this study, we propose K-SHINE, a classification model designed for Korean that builds upon the SHINE framework to address these challenges. The model constructs graphs that incorporate morpheme, part-of-speech, and named entity information, while dynamically learning document graphs for classification. To effectively capture linguistic characteristics of Korean, we utilize the Kiwi morphological analyzer, KLUE/RoBERTa embeddings, and the KPF-BERT-NER named entity recognition model. Experimental results show that K-SHINE outperforms existing comparison models in multi-class Korean short text classification.

키워드