딥러닝 기반 한국어 맞춤법 교정 모델의 세대별 성능 비교 연구 - LSTM부터 GPT-4o까지 -

Abstract

This study compares and analyzes Korean spelling correction models across different generations of deep learning architectures. The models include the pre-Transformer RNN-based LSTM, Transformer-based understanding models BERT and ELECTRA(with KoELECTRA selected and enhanced with a Copy Mechanism), and generative models such as NLLB, ET-5, KoBART, and GPT-4o. Quantitative evaluation used loss, accuracy, F1, and AUC for BERT and ELECTRA, and BLEU, GLEU, recall, precision, and F0.5 for the remaining models. ET-5 demonstrated the highest overall performance, followed by KoBART, NLLB, GPT-4o(zero-shot and 10-shot), LSTM, and KoELECTRA+Copy Mechanism. Qualitative evaluation focused on the five best-performing models. Errors were classified into four categories: spacing, punctuation, loanword spelling, and orthographic/standard language errors, with the last further divided into Replace, Addition, Removal, Unnecessary foreign word use, and Abbreviation/Neologism errors. ET-5 showed the fewest errors, while the others had similar error counts. Spacing errors were most common, followed by orthographic/standard language, punctuation, and loanword errors. This study contributes by empirically analyzing correction patterns and error-handling tendencies of each model, suggesting directions for developing future Korean spelling correction systems that balance practical usability and precision.