Multi-task Learning Model Of Legal Judgment Prediction Based On Model Fusion

Legal judgment prediction refers to the prediction of the judicial document through the case description in the given criminal legal instruments,which can not only provide reference for judicial officers to promote judicial efficiency effectively,but also enhance the judicial transparency and fairness.Among them,charge prediction and law article recommendations are two important sub-tasks of legal judgment prediction.These two sub-tasks are closely related and interact each other.However,previous studies have usually analyzed them as two independent tasks that are analyzed separately,and can't capture relationship between sub-tasks,resulting in poor prediction performance.Another difficulty of this task is that it is difficult to extract relevant features from the case text that can distinguish confusing charges with high similarity in case description text.In order to solve these problems,this thesis presents a multi-task learning model based on model fusion using deep learning technology.The main research contents are as follows:(1)In order to capture the logical dependence between charge prediction and law article recommendations,this thesis adopts the model structure of multi-task learning.In this structure,the feature extraction module extracts key semantic information,the charge keyword representation and the global case description representation,and sends it to the multi-task prediction module;the multi-task prediction module stitches the obtained representation features to construct a multi-task learning model for joint modeling of charge prediction and law article recommendations,and uses multiple binary classifications to simultaneously predict the charge and the law article.This model structure of multi-task learning not only has considerable advantages in learning multiple related tasks through shared representation,but also improves the generalization ability of the model.(2)In order to fully extract the features of the case text,this article designs the feature extraction module and builds the models separately.In order to obtain the key semantic information in the case text,this thesis proposes an attention neural network that combines Transformer Encoder and Deep Pyramid Convolutional Neural Network(DPCNN).Transformer Encoder obtains the representation of case description,and at the same time uses statistical methods to extract the relevant K law articles from the case description,input it into the DPCNN to obtain the relevant law articles representation,and uses the attention mechanism to combine the relevant law articles representation and the case description representation to obtain the Key semantic information representation.Aiming at the problem of confusing charges,this thesis uses an unsupervised method to construct a charge keywordtable based on large-scale legal data,and then extracts the charge keywords that help distinguish confusing charges from the case description,and integrates the charge keyword representation into the model to solve the problem of confusing charges.In order to obtain a global case description representation,this thesis builds a text graph based on the word co-occurrence in the case description and the word co-occurrence in the corpus,and then learns the graph convolution network based on the text graph to obtain discontinuous and long-distance semantics in the corpus co-occurrence of global vocabulary.Multiple comparative experiments conducted on the CAIL2018-Small legal data set show that the model proposed in this thesis can significantly improve the performance of the two tasks of charge prediction and law article recommendations.