This paper presents a work on recognizing emotional feelings of students on social media platforms expressed via posts or comments in Vietnamese, a low-resource language. We first introduce a manually annotated corpus using the seven main emotion types. Based on this corpus, we propose a framework which exploits a novel pre-trained language model, namely BERT, as a backbone for text representation. The key point of this framework is that instead of using traditional cross-entropy as the loss function, we combine CE loss and Focal loss to handle imbalanced data. Experimental results on this dataset show that combining multiple loss functions can be more effective than using them individually. The findings can guide emotion-aware educational technology, enhancing student emotion comprehension and advancing emotionally intelligent virtual assistants.