Abstract: The challenge of extracting complex textual features from Weibo posts has been tackled through the proposal of a text-sentiment analysis algorithm built on a lightweight pre-trained model. First, the loss function was refined by introducing a neural-network dropout regularization algorithm, through which supervised data augmentation was achieved. Second, at the algorithmic level, a dual-channel network combining convolution and temporal attention was incorporated, so that the efficiency of long-text feature extraction was enhanced. Finally, to mitigate the feature loss introduced by deep architectures, residual connections were employed to improve robustness. Experiments were carried out on a public hotel-review dataset and a Weibo-comment dataset. The results indicated that, relative to the baseline model, classification accuracy was increased by 1.33 and 1.41 percentage points on the two datasets, respectively, thereby demonstrating that the proposed algorithm effectively strengthens the feature-capturing capability and sentiment-analysis accuracy of lightweight models.