SymCLIP: Symmetric Prompting with Adaptive Semantic Labeling for Multi-Intent Recognition

Abstract Understanding intent in social media images is challenging in Multi-label Intention Understanding (MIU) due to the ambiguity and abstractness of semantic intentions. Although vision–language models like CLIP provide strong foundations, existing adaptation methods often fail to establish an explicit and synergistic connection between visual features and textual intent labels during the learning process, treating them in isolation. Moreover, the use of fixed intent categories lacks adaptability to dynamic and context-dependent expression scenarios, limiting the model’s interpretative flexibility. In this paper, we propose SymCLIP, a framework that enhances intent understanding through Symmetric Vision–Language Prompting (SVLP) and Dynamic Label Integration (DLI). SVLP couples the two modalities by generating visual prompts from language prompts, enabling semantically guided visual feature extraction. DLI replaces fixed class labels with trainable embeddings, allowing adaptive and fine-grained intent representation. Experiments on the Intentonomy dataset show that SymCLIP achieves a new state-of-the-art MicroF1 of 55.38%, outperforming previous methods by 9.6%. Additional tests on image emotion recognition further confirm its strong generalization across subjective visual understanding tasks.