An Improved Study of Fine-Grained Text Sentiment Analysis Based on Semantic Map Convolutional Networks

The challenge of semantic dynamics and feature coupling difficulties in fine-grained sentiment analysis tasks. Therefore, we innovatively propose an improved framework based on dynamic semantic selection and multi-scale feature enhancement, i.e., RDGCN+ATOSS+GAT, which overcomes the challenge of hierarchical semantic representation of complex sentences and effectively solves the dilemma of parsing multiobjective sentiment conflicts by constructing an adaptive task oriented semantic selection module, ATOSS, and an attentional enhancement architecture of graphs, GAT. The experimental results show that the improved model performs excellently on the fused dataset under the weighted evaluation system, with a weighted accuracy as high as 83.21% and a weighted F1 score of 78.65%. Compared with the baseline model RDGCN, the weighted accuracy is improved by 2.73% and the weighted F1 score by 4.73%, while the logarithmic loss is reduced by 10.94%. The advantages of ATOSS in handling long datasets are fully verified in the long-short dataset comparison experiments. The ablation experiments further confirm that the ATOSS module has a significant hierarchical semantic decoupling advantage in long text scenarios (3.82% reduction in weighted F1 scores), while the GAT module plays a central role in capturing fine-grained sentiment features by virtue of the multi-head attention mechanism (3.55% reduction in weighted F1 scores). With the help of a dynamic weight fusion strategy, the framework deftly balances the cross-domain generalisation capabilities and demonstrates excellent robustness in complex semantic scenarios such as contradictory modifier clauses. The method proposed in this study opens up a new technical paradigm in the field of opinion analysis, which is of great value in e-commerce review mining and opinion analysis applications.