Experimental Comparison of Three Topic Modeling Methods with LDA, Top2Vec and BERTopic

Text clustering can automatically analyze hot events from large-scale data and provide key feature descriptions of event topics. This paper selected #ChatGPT as the text training object, and conducted a comparative analysis of topic clustering on domestic Weibo and overseas Twitter social platforms. Weibo and Twitter are characterized by a large number of active users, high discussion heat, and wide public attention. Therefore, we chose the text in the #ChatGPT topic of these two platforms as the main data for data analysis in the experimental part. This study aims to evaluate the independence (separateness) between topic clusters. Finally, through the 3-D scatterplot of cosine similarity and Pearson’s correlation coefficient, we conclude that BERTopic has better topic separation, more independence between the eight topics, relatively clear semantics, and better understanding of text data structure and content. Compared with the other two methods (LDA and Top2Vec), the BERTopic model in the experiment is at least 34.2% better than the other algorithm models in Chinese and English clustering, and a better topic clustering effect is obtained.