Aiming at the poor prediction effect in the crime prediction task, a crime prediction model incorporating quantitative information is proposed. According to the numerical basis of judicial sentencing, quantify the values in the text, use the self-attention mechanism and maximum pooling to highlight the weight of quantitative information and keywords, then use the bidirectional long and short memory cycle network to extract the joint features between quantitative information and keywords, and use the capsule network to extract the depth features; Due to the imbalance of data, a priori weight is used in the loss function to balance the data. Experiments show that this method has better prediction effect than other models.