KoBERT–LLM 합의 기반 기타 라벨 재라벨링 기법

초록

The “other” label in text classification aggregates heterogeneous and ambiguous samples, often limiting overall performance. We propose an automatic relabeling method that combines a KoBERT classifier with an LLM, using their agreement and confidence to reassign “other” documents to more suitable labels. Retraining on the relabeled data redistributes “other” samples and improves overall performance as well as the stability of “other” predictions. This suggests model-consensus relabeling as a practical way to mitigate the “other” label issue with minimal manual validation. Future work will extend this into a universal data refinement framework via hierarchical classification and prompt optimization.

키워드