근대 잡지 <개벽>의 데이터과학적 언어 분석

Abstract

This paper presents a data-scientific analysis of the entire corpus of the modern Korean periodical Gaebyeok, examining both metadata and main article texts. Metadata analysis reveals that unspecified pseudonyms account for about 35% of entries—mostly announcements or miscellaneous items not essential for identifying pen names—despite the periodical’s history of forced discontinuation and author censorship. Main text analysis, using deep learning techniques— subword tokenization, fastText static embeddings, and BERTopic dynamic embeddings—produced notable results. Subword tokenization effectively segmented tokens such as ‘아니-하면,’ ‘무엇-을,’ ‘경제-적,’ and ‘생각-을’ without a conventional morphological analyzer. fastText captured subwords as character n-grams in modern Korean texts with orthographic differences from contemporary Korean, positioning functionally related words such as ‘잇다,’ ‘업다,’ ‘하다,’ and ‘것이다’ in close semantic space. BERTopic identified latent topics, with top clusters including ‘people, arts’, ‘thoughts, mind’, ‘Gaebyeok, readers’, ‘song, sound’, ‘school, Gyeongseong’, ‘Joseon, society’, ‘Japanese, Korean, and ‘conference, government’. An important subcluster on ‘Cheondokyo’, the founding religion of the periodical, should also be incorporated into the topic results.