To quickly identify users' hot topics in offensive comment texts and solve the problems of insufficient semantic description, loss of contextual information, and weak topic coherence of traditional topic models when dealing with comment texts, as well as the sensitivity of K-value and initial centroid of K-means clustering algorithm. The CoSENT (cosine sentence) model was used in this paper to obtain sentence-level vector features of comment texts containing offensive language. An improved K-means algorithm based on Canopy +was used for class clustering on the vector-matrix after dimensionality reduction through the UMAP(uniform manifold approximation and projection) model. c-TF-IDF (class term frequency-inverse document frequency) was used to identify the thematic features of each thematic cluster for thematic modeling. The validity of the method is verified through experiments comparing the offensive comment text dataset as well as the ordinary comment dataset. The results show that the method in this paper can get better topic consistency.