BERT-Based Prediction Method of Imprisonment Term of Criminal Case

Abstract

Term of imprisonment prediction is to predict the result of sentence according to the case facts description in criminal cases.To solve the problem that the superficial features and the dependence relationship between subtasks are not sufficient to distinguish confusable cases with the same charge and law article but different imprisonment terms, and considering that the prediction results of imprisonment term are affected by the case elements from case description and the distinguishable attribute features, this paper proposes a BERT-based prediction method of imprisonment term of criminal case (namely, BERT + LSTM + Attention Mechanism method).This method first uses the long and short memory network model(LSTM) with attention mechanism to obtain the distinguishable attribute representation vector, and then uses the BERT model to obtain the case element representation vector, and finally combines the two kinds of vectors so as to obtain the prediction results through a softmax classifier. The experimental results on the CAIL dataset show that the proposed method has good experimental results in predicting the term of imprisonment of criminal cases.