Microblog, as an important platform for people to express their opinions, has become a research hotspot of text sentiment analysis. There are some problems in micro-blog texts, such as casual expression, mixed Chinese and English, and difficult identification of emotional words on the Internet, which make it difficult to determine emotion classification. This paper proposes a BERT_LSTM model based on attention mechanism. Firstly, the Roberta model is used to extract semantic features and input them into the LSTM module, and then the semantic features are transferred repeatedly and finally input into the Attention layer to obtain weighted features. Finally, Softmax linear function is used to achieve sentiment classification. The comparative experimental results show that the accuracy of Roberta_LSTM model based on attention mechanism proposed in this paper is 2.1% higher than that of single-language Roberta model, which shows that this model can more accurately capture the emotional tendency of microblog short texts.