With the rapid development of short video platforms, their content security governance is facing severe challenges. This article conducts a data-driven modeling study on the identification of illegal information in platform comment texts. The study first cleaned and integrated large-scale authentic comment texts, and constructed a diastase containing both violation and non violation samples; Furthermore, exploratory analysis was conducted on the diastase through word length distribution and word frequency statistics, and it was found that high-frequency words for illegal comments were concentrated in sensitive words such as “black”, “discriminatory”, and “disgusting”; On this basis, the LDA topic model was used to mine the violation information, and the optimal number of topics was determined to be 9 through consistency scores, revealing core violation topics such as “racial discrimination” and “gender opposition”; Finally, a support vector machine (SVM) classification model was constructed for automated recognition, and the AUC value of the model reached 0.930 in testing, demonstrating excellent classification performance. This study confirms the effectiveness and practicality of text mining technology in the field of content security, providing reliable technical solutions and theoretical basis for platform side automated auditing.