Abstract
This article examines methods to identify and mitigate offensive and abusive language in online material on social networks. The study seeks to address a critical issue: the identification of violent and damaging information on Kazakh social media platforms. At present, Kazakh is regarded as a resource-deficient language, and the quantity of automated systems processing its texts is restricted. The study introduces a novel hybrid model that employs contemporary machine learning and deep learning techniques tailored for Kazakh texts. The proposed model utilizes BERT transformer embeddings and integrates multi-head self-attention with convolutional neural network layers. This methodology enables the successful detection of both explicit and implicit hate speech. The stages of natural language processing (NLP), including pre-processing, lemmatization, and tokenization of textual data, are examined, and a specialized Kazakh dataset is created. The data is categorized into classes by voluntary annotators. The model's results demonstrated superior accuracy and stability relative to conventional approaches (e.g., Naive Bayes, LSTM, FastText). The proposed architecture was particularly adept at identifying implicitly conveyed texts with significant emotional depth. This study's findings will enhance digital security in the Kazakh language and foster ethical conversation among social media users. In the future, the model is intended to be utilized in real-time and augmented with multilingual data.