[Objective] An emotion-cause pair extraction method based on multi-label Seq2Seq model is proposed to improve the F1 value. [Methods] BERT pre-training is used to obtain semantically rich word vectors. Bi-GRU and LSTM are used to encode the global features and local features of the text respectively. The hybrid attention mechanism is introduced to realize the fusion of the two and improve the integrity of text semantic feature capture. [results] For the joint extraction of emotional cause pairs, compared with the latest model, the F1 value of emotional cause pairs in this paper is increased by 0.98 % and 0.81 % on the two data sets, the subtasks of emotional extraction are increased by 0.68 % and 0.78 % respectively, and the subtasks of cause extraction are increased by 1.1 % and 2.31 % respectively. [limitations] The model mainly considers explicit emotion-cause pairs and does not discuss implicit emotion-cause pairs. [Conclusion] Experimental results show that the proposed model improves the F1 of emotion-cause pairs.