Abstract
This research paper considers methods for automatic detection of profanity and aggressive comments in Kazakh social networks. The relevance of the study is due to the increasing volume of profanity comments online and the lack of effective automated solutions for resource-limited languages. The Kazakh language is characterized by complex morphology and contextual specificity, which makes it difficult to use traditional natural language processing methods. The proposed study introduces a hybrid model based on the BERT transformer architecture, combining multihead attention and convolutional neural networks. This architecture enables the effective detection of offensive texts, including both explicit and subtle forms of comments exchanged among cadets and officers online. To train and test the model, a corpus of Kazakh-language comments annotated by experts was constructed, including opinions from cadets, commanders, and personnel. The quality of the annotation was assessed using Fleiss's Kappa coefficient, which gave 0.6649, with an overall agreement level of 91.67%, which indicates high reliability of the data. Experimental results showed that the proposed model outperforms the baseline methods (Naive Bayes, LSTM, FastText) in terms of accuracy, F1-score, and ROC-AUC. The resulting ROC-AUC value was 0.95. The research results confirm the effectiveness of transformer models with an attention mechanism for detecting profanity in the Kazakh language and can be used in the development of automated content moderation and information security monitoring systems.