NEURAL NETWORK METHODS FOR ANALYZING TEXT FEEDBACK FROM STUDENTS: ARCHITECTURE, LEARNING, AND INTERPRETATION OF RESULTS

This article presents a comparative analysis of neural network architectures LSTM, BiLSTM, and RuBERT applied to the automatic classification of student feedback. Particular attention is paid to data preprocessing, manual annotation with consideration of sentiment and thematic aspects, as well as corpus reliability assessment using Cohen’s kappa coefficient (0.82). The experimental results show that RuBERT achieves the highest Accuracy (0.87) and F1-score (0.85), which is statistically confirmed by t-test results. At the same time, BiLSTM demonstrates higher efficiency compared to LSTM due to its ability to capture bidirectional context. Error analysis revealed that recurrent models most often confuse neutral and negative feedback, while the transformer-based architecture performs better in handling ambiguous expressions and subtle emotional nuances. The practical significance of the study lies in the possibility of integrating the proposed approach into LMS and digital educational platforms to automate the analysis of student feedback. Future research directions include expanding the corpus with Kazakh-language texts and applying advanced Transformer-based models (RoBERTa, DeBERTa, ChatGLM).