Sequence-to-sequence models for generating emotion labels have been identified as an effective method for multi-label emotion classification, utilizing RNNs to capture emotion relationships. However, while these sequence generation models implicitly learn label correlations, they may not adequately capture the strong correlations that exist among nuanced emotions. Moreover, traditional RNN-based sequence generation models may struggle to extract textual semantic information as effectively as pre-trained language models. To address the above problems, this paper proposes a novel Multi-Label Emotion Classification Based on Pre-trained Sequence Generation Model (EmoBART). EmoBART leverages a pre-trained generative language model, BART, as the network skeleton for generating emotion labels and incorporates a correlation network to explicitly model fine-grained emotion correlations. The EmoBART architecture comprises an encoding module for semantic information extraction, a decoding module for predicting emotion label sequences through generative labeling chains, and a correlation network module that integrates the CorNet network to learn emotion correlations and predict emotion labels. Extensive comparative experimental results show that EmoBART outperforms existing models in terms of multi-label emotion recognition accuracy.