Empathy prediction from texts achieves little progress due to the lack of sufficient labeled data， while the related task of text sentiment polarity classification has a large number of labeled samples. Since there is a strong correlation between empathy prediction and polarity classification， a transfer learning?based text empathy prediction method was proposed. Transferable public features were learned from the sentiment polarity classification task to assist text empathy prediction task. Firstly， a dynamic weighted fusion of public and private features between two tasks was performed through an attention mechanism. Secondly， in order to eliminate domain differences in datasets between two tasks， an adversarial learning strategy was used to distinguish the domain?unique features from the domain?public features between two tasks. Finally， a Hinge?loss constraint strategy was proposed to make common features be generic for different target labels and private features be unique to different target labels. Experimental results on two benchmark datasets show that compared to the comparison transfer learning methods， the proposed method has higher Pearson Correlation Coefficient （PCC） and coefficient of determination （R²）， and has lower Mean?Square Error （MSE）， which fully demonstrates the effectiveness of the proposed method.