Analysis of K-culture in Reddit data Using FFN Concept Representations

초록

This paper applies FFN Concept Representations (FFNCR)-a representation-centered computational framework that enables corpus insight beyond keyword-based topic clustering-to analyze Korean cultural discourse. We examine discursive continuity at the level of communicative functions rather than general lexical stability using a time-aware transformer trained on Korean-related Reddit data, analyzing feed-forward network (FFN) dimensions in vocabulary space across temporally aligned checkpoints. The results show that FFN dimensions form stable semantic structures that persist over time while their lexical realizations shift, reflecting gradual semantic reorganization. At the corpus level, we observe that domains such as platform interaction and media discourse persist across years while their lexical expressions evolve, including shifts toward more automated interaction patterns and platform-specific media signals. Compared with document-level topic models, FFNCR captures this persistent semantic-functional structure and reveals how interactional practices re-emerge and evolve. These findings demonstrate that concept-level representations provide a robust framework for analyzing cultural discourse and diachronic linguistic behavior in online communities.

키워드