This paper proposes a novel method to enhance the performance of two key aspects of natural language processing: aspect-based sentiment analysis (ABSA) and span detection, specifically for the Vietnamese language. The proposed approach leverages the strengths of pre-trained language models with a two-stage task decomposition strategy. In the first stage, the model identifies opinion spans and their associated aspects within the text, functioning as a sequence tagging task. In the second stage, it classifies the sentiment polarity related to each identified aspect, treating it as a hierarchical classification task. This approach effectively captures both linguistic dependencies and sentiment relationships within Vietnamese text by structuring the learning process in this manner. Experimental evaluations demonstrate the superiority of this approach, achieving a micro F1 score of 76.27%, which significantly outperforms prior baselines. This highlights the model’s capability to handle hierarchical classification and sequence tagging tasks, particularly in low-resource language environments.