This paper investigates the task of automatic topic classification for Vietnamese educational text data. We contribute a novel dataset annotated with nine prevalent topics discussed among Vietnamese students on social media platforms. The dataset inherently suffers from class imbalance issues, where some topics are significantly less frequent than others. To address this challenge, instead of using the conventional cross-entropy loss, we propose utilizing an alternative Dice loss function, a function typically used in image segmentation, for imbalanced learning. We employ a pre-trained language BERT model as the backbone for text feature extraction. Experimental results on the newly built dataset demonstrate the effectiveness of our proposed method in handling imbalanced class distributions within the Vietnamese education domain. Specifically, using the Dice loss increases the F1 score by 79.72% compared to using CE loss.