The invention belongs to the technical field of natural language processing text classification, in particular to a multi-source multi-label text classification method based on an improved seq2seq model and a system thereof. The method comprises the following steps: data input and pretreatment, word embedding, encoding, encoding and splicing, decoding, model optimization and prediction output. Themethod of the invention has the following beneficial effects: adopting a seq2seq depth learning framework, constructing a plurality of encoders, and combining the attention mechanism to be used for atext classification task, so as to maximize the use of multi-source corpus information and improve the classification accuracy of the multi-label; In the error feedback process of decoding step, according to the characteristics of multi-label text, an intervention mechanism is added to avoid the influence of label sorting, which is more in line with the essence of multi-label classification problem. The encoder adopts the circulating neural network, which can learn according to the time step effectively. The decoding layer adopts one-way loop neural network and adds attention mechanism to highlight the learning focus.