<p>Emotion classification from text is a crucial task in affective computing, with applications in social informatics, human-computer interaction, and urban data analysis. This task involves identifying emotional states and answering questions such as, “Is the writer or reader happy, angry, or fearful about the target?” While deep neural network (DNN)-based models are effective at this task, they frequently require large labelled datasets, which are expensive and impractical, especially for dynamic, irregular data seen on platforms such as social media. To address these challenges, an efficient weakly-supervised learning framework is proposed for text emotion classification that reduces reliance on large-scale annotations. A Transformer-based model and anchor-based similarity computations are incorporated within a self-training framework to generate weakly labeled data. A compact neighbor consistency regularization (CNCR) mechanism is introduced to improve classification robustness across diverse datasets. Additionally, heuristic rules are designed to filter raw data, manage noise, and ensure better model generalization across different datasets. Extensive experiments were conducted on three real-world datasets, demonstrating that the proposed method is both efficient and comparable to previous semi-supervised learning algorithms.</p>