Abstract
This paper presents a multi-task learning approach for multi-label emotion classification in Kazakh-language news articles using transformer-based models. We constructed a labeled dataset of 600 news articles by collecting and analyzing real user reactions from the official Kazakh news platform NUR.KZ. Each article was annotated with one or more emotions based on aggregated user feedback. A pre-trained transformer model was fine-tuned to predict these emotional reactions, demonstrating effective performance in detecting multiple emotions simultaneously. Our model achieved low loss values during training and validation, indicating good convergence, and high accuracy across the emotion categories. These results highlight the effectiveness of using naturally occurring user feedback for emotion modeling in low-resource languages and suggest promising applications in other domains such as social media and public opinion analysis.