<p>Text classification is a crucial task in natural language processing, encompassing a wide range of sub-tasks such as sentiment analysis, topic categorization, news classification, and question answering. Deep learning models have demonstrated exceptional performance in this domain. However, many deep neural network models struggle to handle complex and imbalanced datasets, and their adaptability across multiple scenarios remains limited. To tackle these challenges, this paper proposes DT-GCNN, a model that integrates GRU for capturing sequence information and CNN for extracting local features. Furthermore, the model incorporates an adaptive soft-margin triplet loss function that dynamically adjusts triplet margins, thereby enhancing the learning quality of the embedding space. During the training process, intelligent algorithms are periodically employed to dynamically reconstruct triplets, thereby enhancing the model’s generalization capability. This study conducts extensive experiments on eight datasets from various categories. The results demonstrate that DT-GCNN outperforms most existing baseline models, showing notable superiority in handling complex and imbalanced category tasks. The proposed method also significantly enhances generalization ability and stability, exhibiting excellent performance across diverse datasets.</p>