딥러닝 기반 한국 고전한문 표점 추론 자동화 모델의 구축과 활용 = Development and Application of a Deep Learning–Based Model for Automated Punctuation Inference in Korean Classical Chinese

다국어 초록 (Multilingual Abstract)

This study compiles and refines collated and punctuated Classical Chinese texts accumulated through prior research and projects to construct a database of approximately 3.4 million items (≈420 million characters). Building on this resource, we develop a punctuation inference model specialized for Korean Classical Chinese by fine-tuning the pretrained deep learning language model Chinese-RoBERTa into a multi-label token classification architecture. The training corpus—covering eight genres including annals, collected works, and diaries—was preprocessed and standardized to seven punctuation marks (, 。 · ? ! 《 》). The final model achieves an overall F1 score of 0.9050 on held-out validation data. On unseen corpora containing only traditional ring-dot punctuation (Hanguk Munjip Chonggan and Ilseongnok), the model attains F1 scores of 0.8784 and 0.9065, respectively, for punctuation-position matching. By punctuation type, question marks, commas, periods, and middle dots exhibit high performance, whereas book-title brackets (《》)— which require long-range dependencies in paired structures—and exclamation marks—sparse in the data—show lower recall. We release an open-source integrated system—including model weights, training data, source code, and GUI/CLI batch processing—to support records and information services and research workflows using natural-language analysis, such as text preprocessing, indexing and search, translation preprocessing, and OCR postprocessing. Future work includes a dual-path architecture for paired punctuation, genre-adaptive modules, and multi-task integration with sentence-structure analysis and named-entity recognition.