ERCcR: A Clustering Topic Model Integrating Statistical and Semantic Features for Enhanced Topic Representation

Topic models aim to automatically discover themes from large collections of text. However, most popular clustering-based topic models primarily rely on statistical weighting strategies for topic representation, overlooking the rich semantic information present in the documents. To fully leverage document semantics and improve the weighting strategy, enhancing both the consistency and diversity of the generated topics, a semantic-enhanced clustering topic model (ERCcR) is proposed. This model follows the clustering topic model paradigm to obtain topic clusters, and introduces an innovative weighting strategy that comprehensively considers both statistical and semantic information in topic representation. The strategy first integrates a variant of Term Frequency-Inverse Document Frequency (TF-IDF) to capture statistical information, and then incorporates word-level semantic features through a word similarity matrix. Furthermore, document semantic information is extracted via nested clustering and incorporated into the weight calculation. Evaluation experiments were conducted on four commonly used topic model datasets, and ERCcR achieved strong results in terms of topic coherence, with scores of 0.755, 0.815, 0.772, and 0.678, respectively. In terms of topic diversity, ERCcR outperformed other clustering-based topic models, with scores of 0.909, 0.921, 0.901, and 0.867. These results demonstrate that ERCcR effectively leverages semantic information to enhance its ability to represent topics.