단어의 확률적 정보와 임베딩 유사도를 통합한 그래프 뉴럴 네트워크 기반 토픽 모델

초록

Topic models have been widely utilized to systematically identify key themes in large-scale text data. Representative approaches include the probabilistic latent Dirichlet allocation (LDA) and the embedding-based BERTopic. However, probabilistic models often fail to capture semantic associations between words, whereas embedding-based models tend to be less interpretable. To address these limitations, this study proposes a new graph neural network-based topic model that integrates the numerical signals of both approaches. A keyword graph was constructed by assigning LDA-derived topic-word probabilities as edge weights and BERTopic embeddings as node features, followed by graph-based deep clustering to extract topics. The proposed model is validated through a case study using online text data related to generative AI-based English education. In this case, the model effectively distinguishes major discussion topics within educational discourse while capturing semantically related sub-concepts. As a result, it outperformed single-model baselines in both coherence and modularity metrics, demonstrating that the fusion of probabilistic structures and semantic similarities enables the extraction of more cohesive and interpretable topics.

키워드