Abstract
Modern text data processing and classification methods require extensive use of machine learning and neural networks. Categorizing text into different classes has become a crucial task in many fields. This paper presents a multi-class text classification model utilizing a Modified TF-IDF (MTF-IDF) approach in combination with Long Short-Term Memory (LSTM) neural networks, XGBoost, and MLPClassifier algorithms. Additionally, the study explores the integration of TF-IDF and CountVectorizer (MTF-IDF) methods for text vectorization, aiming to enhance classification efficiency. The research findings indicate that the LSTM model achieved the highest accuracy rate of 89%, demonstrating superior performance. The MLPClassifier model achieved 85% accuracy, while XGBoost obtained 81% accuracy. Moreover, the integration of TF-IDF and MTF-IDF methods significantly improved the detection of rare but essential words, enhancing the overall performance of the models. This study is dedicated to addressing the problem of automated detection of harmful content in the Kazakh language. Hate speech in the digital space refers to any online material that harms individuals or communities through aggression, manipulation, discrimination, or the intentional spread of socially damaging narratives. The results provide a solid foundation for future research aimed at the early identification and mitigation of hate speech in the digital space, contributing to a safer online environment.