A Novel Approach to Enhance the Topic Modeling Using Graph with BERT Embeddings

This paper introduces a novel topic modeling framework that integrates graph-based representations with BERT embeddings to enhance semantic understanding and structural analysis of document collections. Traditional methods often struggle to capture the interplay between document relationships and semantic context. We address this by constructing a document graph based on co-citation, co-reference, or semantic similarity, and augment it with BERT embeddings. Through fusion techniques, we combine these representations, resulting in improved topic coherence and interpretability. Extensive tests across various datasets show our method effectively identifies hidden topics and improves tasks like summarizing text and finding information. Our approach uses BERT and a graph-based topic model