With the continuous advancement of Natural Language Processing (NLP) technologies,sentiment analysis has been extensively applied in fields such as social media monitoring,customer service,and market research. However,when dealing with fine-grained sentiment classification tasks,existing methods often rely on a flat classification strategy,failing to fully capture the inherent hierarchical relationships among emotions,which consequently limits classification accuracy. Here,we propose a novel fine-grained text sentiment classification method based on a hierarchical framework. This method constructs a hierarchical emotional label system by grouping fine-grained labels with shared emotional meaning into broader coarse-grained categories,thereby enhancing the model's ability to comprehend emotional hierarchies. Architecturally,the model leverages feature representations from different Transformer layers of the RoBERTa (Robustly optimized BERT Approach) encoder to capture both low-level and high-level semantic information,which supports the training of distinct classifiers of coarse- and fine-grained levels. We implement two hierarchical strategies:a multi-task learning one where the coarse-grained classification task serves as an auxiliary to regularize and improve the generalization of the primary fine-grained classifier;a parallel coarse-to-fine strategy where each fine classifier computes its loss independently,then incorporates prior information from the coarse-grained classifier to optimize feature selection,enabling progressive refinement of sentiment predictions. Furthermore,the advantages of both strategies are combined via a probability threshold mechanism to further boost performance. Experimental results on the GoEmotions and Empathetic Dialogues datasets show that both proposed strategies outperform the baseline model in terms of F1-score and accuracy. The final fused model achieves further improvements,with accuracies reaching 64.6% and 59.4% on the respective datasets,which surpasses the performance of existing state-of-the-art methods.