Research Of Automatic Charge Prediction In Judicial Field

In recent years,the cross integration of artificial intelligence and judicial field has been widely concerned by academia and industry.The application of judicial artificial intelligence will not only help improve the efficiency of the judicial department,but also help lower the threshold for the use of judicial services,which is conducive to promoting justice and transparency.Automatic charge prediction is one of the most critical tasks for judicial artificial intelligence.It aims to predict charges committed by some suspects based on corresponding criminal facts.This paper focuses on the automatic charge prediction in the judicial field.The main work and results are as follows:For modelling the semantic differences of words in different cases,we propose representing the criminal facts based on the semantic importance differences of words.In fact encoding part,we combine the context information to get the attention weight of each word based on bidirectional gated recurrent neural network and self-attention mechanism.The extensive experiments show that the proposed model gives better representation performances,therefore,we get better prediction results than other comparison models.And finally we achieve 88.0% on F_union in CAIL dataset.For modelling the problem of multiple accusations for one defendant,we propose combined punishment charge prediction model by conversion strategies.In this model,we firstly transform the problem into a multi-label text classification task.Then we use binary relevance strategy to decompose multi-label charges into multiple independent single-label charges so as to achieve the binary classification of each charge based on the Sigmoid classifier.The experimental results demonstrate that the proposed model can effectively solve the problem of combined punishment for several charges prediction.And we obtained a 4.2% performance improvement on the CAIL dataset.For modelling the intergration of external legal knowledge,we propose a model based on article and fact attentive interaction mechanism.In this paper,inspired by legal process that the law articles are the authoritative legal basis for human judges to arrive at an appropriate sentence based on criminal facts,we propose a novel fact-article attentive interaction mechanism to automatically measure the compatibility between the hidden representations of articles and facts.Extensive experimental results demonstrate that the proposed model significantly outperforms five state-of-the-art methods and consistently across three benchmark datasets,in terms of both accuracy and interpretability.In conclusion,this paper proposes corresponding methods for dealing with the task of automatic charge prediction in the judicial field.The major contributions are as follows:?1?We improve the encoding mechanism of criminal facts and increased the accuracy of the fact semantic representation.?2?Based on the conversion strategy,the model can predict combined punishment cases,which improves the universality of the charge prediction model.?3?we propose a model based on article and fact attentive interaction mechanism for the intergration of external legal knowledge,which improves interpretability of the prediction process.The exploration in this paper can lay a technical support foundation for subsequent research.