<p>Document-level sentiment analysis has gained prominence as data volumes grow, yet there often lacks a precise definition and efficient processing methods for genuinely long texts. This study addresses these challenges by proposing DocSentiNet. This novel adaptive model architecture combines an improved TaylorAttention mechanism, Text Convolution, and Longformer. Our approach effectively captures both global semantic dependencies and local features in extended texts, significantly reducing computational complexity while maintaining performance. We introduce a Taylor expansion-based attention mechanism and a multi-level feature fusion strategy to achieve this. Evaluated on multiple document-level sentiment analysis datasets, DocSentiNet demonstrates superior performance in handling long text sequences, achieving substantial improvements over existing state-of-the-art models with only a marginal increase in parameters. This research advances document-level sentiment analysis and provides insights for enhancing pre-trained language models in long text understanding tasks.</p>