Intelligent Judicial Research Based On BERT Word Vector And Attention-CNN

Nowadays,artificial intelligence is developing rapidly,and intelligent justice is an important application field of artificial intelligence.On the one hand,the rapid development of deep learning provides technical support for intelligent judicial research.On the other hand,a large number of open judgement documents solve the big data need of intelligent judicial research.In the construction of intelligent justice,achieving accusations prediction and relevant law articles prediction with artificial intelligence can help judges to make judgments and save a lot of resources,which is of great significance for the construction of intelligent justice.This paper realizes the accusations and relevant law articles prediction by solving the multi-label text classification tasks of accusations and relevant law articles.This paper utilizes the CAIL2018-Small dataset to conduct research on criminal cases of single-person with multiple accusations and multiple relevant law articles.The data in the dataset are all derived from the public criminal case judgment documents on the China Judgments Online.All data in the dataset has only one criminal.There are 202 accusations and 183 relevant law articles in the dataset.One piece of data may contain multiple accusations and multiple relevant law articles.This paper utilizes the micro-average F1 value(F1micro)and the macro-average F1 value(F1macro)as evaluation metrics.This paper utilizes 100 times the average of F1 micro and F1 macro as the score of each task.The main research work of this paper is as follows:(1)Comparing three average word vector models,including the average word2 vec word vector model,the average BERT word vector model,and the average BERT-word2 vec word vector model.Compared with the two single word vector models,the average BERT-word2 vec word vector model with different word vector features has higher scores in the multi-label text classification tasks of accusations and relevant law articles.(2)Adding different levels of Attention mechanism to the multi-core CNN model and combining the BERT sentence vector features to propose the BERT-ACNN model.Compared with the three average word vector models,the four RNN models and the CNN model,the BERT-ACNN model has the highest scores in the multi-label text classification tasks of accusations and relevant law articles.Finally,this paper utilizes the methods of oversampling and increasing the number of convolution layers to improve the performance of the BERT-ACNN model.