LDA와 네트워크 분석의 융합을 통한 단어 의미 중심성의 시계열 추적

The recent surge in digital text data has highlighted the importance of quantitative text analysis techniques for identifying discourse patterns and extracting key terms at specific points in time. In particular, Latent Dirichlet Allocation (LDA)-based topic modeling has been widely used as an effective method for automatically extracting themes and representative keywords from document collections. However, most existing studies are limited to interpreting LDA results at individual time points or comparing the frequencies of top-ranked keywords. These approaches fall short in explaining semantic relationships between words or their structural position changes. Even in time-series analyses of textual data, the results are often presented as a simple listing of outputs from each point in time. Furthermore, there remains a lack of methodological approaches for systematically tracking how specific words are associated with different topics and semantically related words across time. To address these limitations, this study proposes a methodology that extends LDA topic modeling into a network-based framework by constructing word–word networks for each year and tracking changes in word centrality over time. Specifically, it selects meaningful words based on their probability distributions and filters weak edges to reduce noise, thereby constructing refined semantic networks for structural analysis. Centrality metrics are then analyzed in a time-series context to quantitatively identify how a word's role and semantic centrality change within the topic structure over time. This study offers a distinctive insight by combining LDA with network analysis to trace changes in the semantic centrality of words over time, moving beyond traditional frequency-based trend analysis. In particular, the proposed time-series and structural approach to text data analysis enables the detection of discourse turning points, the analysis of semantic expansion paths for key concepts, and the identification of shifting boundaries between topics, suggesting broad applicability across diverse subjects and domains.