Knowledge Representation And Application Of Law Prediction Research

Automatic legal provision prediction is an important sub task in the field of judicialintelligence.Its purpose is to predict the corresponding legal provision according to the case description.Traditional law prediction methods are based on single model,and it can not obtain enough knowledge for law prediction.Therefore,how to obtain additional knowledge and integrate these knowledge into the law prediction model is the key of this task.Aiming to solve this problem,we optimize the techniques by introducing the external knowledge and multi-task learning.The contributions of this paper mainly includes the following aspects:(1)In order to deal with the redundant information in the encoding state of the recurrent neural network and the confusing cases in the case description,we propose an attention and filter gate based model for law prediction.This model introduces the gate structure and attention mechanism into the law prediction method.The key of the model is to use the gate structure to filter the hidden layer state of the recurrent neural network after encoding,and then perform the attention calculation between the context vector of the law label and the filtered hidden state.After that,the model can learn the similarity between each law and case description.The similarity relationship is used to enhance the knowledge representation ability of the classification model and improve the effect of the law prediction.(2)In order to solve the problem that different level crimes are cited in different legal pro-visions,we propose a method of law prediction based on position embedding vectors and multivariate feature.This model aims to use the convolutional neural network for feature extraction,adding position vectors to overcome the problem that convolutional neural networks cannot learn long-short term dependency.At the same time,by incor-porating the(key,value)features extracted in advance by the knowledge extractor,the model can effectively improve the understanding of the case description and obtain more information that helps to predict the law.(3)To solve the data imbalance problem,we propose a law prediction model based on label dependence and few-shot learning.This method infers the attributes and legal provision simulaneously,and calculates the attention weights between the attribute vectors and text embedding to obtain additional knowledge.At the same time,the graph convolution network is used to model the dependence of the label space,and then the correspondence between the label and the text is obtained to assist the prediction of less samples.In addition,the proposed methods of law prediction are applied to the WeChat applet.