A text classification method in Natural Language Processing （NLP） was introduced into the field of criminal psychology to scientifically and intelligently grade the violent tendencies of prisoners. A Criminal semantic Convolutional Hierarchical Attention Network （CCHA-Net） based on the joint modeling of two channels of improved HAN （Hierarchy Attention Network） and TextCNN （Text Convolutional Neural Network） was proposed to complete the violent criminal temperament grade by separately mining the semantic information of crime facts and basic information of prisoners. Firstly， Focal Loss was used to simultaneously replace the Cross-Entropy function in both channels to optimize the sample size imbalance problem. Secondly， in the two-channel input layer， positional encoding was simultaneously introduced to improve the perception of positional information. The HAN channel was improved by using max-pooling to construct salient vectors. Finally， global average pooling was used to replace the fully connected method in all output layers to avoid overfitting. Experimental results show that compared with 17 related baseline models such as AC-BiLSTM （Attention-based Bidirectional Long Short-Term Memory with Convolution layer） and Support Vector Machine （SVM）， the indicators of CCHA-Net reach the best， the micro-average F1 （Micro_F1） is 99.57%， and the Area Under the Curve （AUC） under the macro-average and the micro-average are 99.45% and 99.89%， respectively， which are 4.08， 5.59 and 0.74 percentage points higher than those of the suboptimal AC-BiLSTM. It can be verified that the violent criminal temperament grade task can be effectively performed by CCHA-Net.