Research on Artificial Intelligence Assisted Judge Decision-Making—From the Perspective of Sentencing Deviation Identification

Sentencing is the ultimate embodiment of penal justice.To achieve the goal of “making people feel fairness and justice in every judicial decision,” the Chinese Supreme People’s  Court continues to reform the standardization of sentencing.For this purpose,the Supreme People’s Court issued the “Sentencing Guidance for the People’s Court” (referred to as the Guidance) in 2008,which has since been revised six times.The Guidance provides comprehensive guidelines for the basic methods and steps of sentencing,the scope of common sentencing circumstances,and the sentencing of common crimes.However,real-world scenarios are complex,and the Guidance cannot cover all situations.Furthermore,differences in regional economic and social development levels,divergent rulings among individual judges,and the personal characteristics of defendants may lead to inconsistent sentences for similar cases.This may lead to a low rate of settlement,as seen in 2020 when the Supreme People’s Court heard a total of 1.12 million first-instance cases,of which about 11% and 2% went through a second trial and remand for retrial,respectively.This indicates that numerous cases have not yet been settled (without appeal or reverse appeal),and many of these may be controversial cases where judges may hold divergent opinions on specific sentencing such as imprisonment or fines.This low rate of settlement may also influence the judicial process and is detrimental to safeguarding the authority and credibility of the law.Some scholars have suggested that the individual judge’s discretion in sentencing should be compared with the collective experience of judges in sentencing.Judges who make rulings that reflect the collective experience should be supported and respected for their discretion,while judges who deviate significantly from the collective experience should have their decisions identified and corrected.Since 2016,the Chinese Supreme Court has vigorously promoted the construction of smart courts,hoping to use big data and artificial intelligence technology to discover judicial consensus.This would improve the accuracy and fairness of case acceptance and trials,and enable the judicial system to make fair and consistent rulings.

From the perspective of trial supervision,this article proposes a technical method for automatically detecting sentencing deviations.Since 2021,China Judgments Online (https://wenshu.court.gov.cn/) has released more than 100 million legal judgment documents to the public.This undertaking has provided a massive data foundation not only for research related to judicial judgments,but also for developing advanced machine learning algorithms that can automatically detect deviations in sentencing.In this article,we take the criminal judgment documents from 2018 as the sample and analyze a total of 460486 legal documents based on 62 charges.We then propose a method that can accurately detect abnormal situations of sentencing deviation in judicial trials.Specifically,the model includes the following three aspects.First,using the sentence as the dependent variable and the text extracted from the “as determined through trial” and “considered by the court” in the legal documents as the description of case facts,four deep learning models (LSTM,TextCNN,BiLSTM,Transformer) are constructed for sentence prediction.Second,based on the predicted results of the model,the difference between the predicted sentence and the actual sentence is calculated.Finally,we propose a heterogeneity index that is used to identify the charges that have deviations in sentencing.Inspired by the coefficient of variation in statistics,the heterogeneity index is constructed as

Heterogeneity index of the Rth crime =( Median of the absolute residual value of the R^th crime)/(Median of the sentence of the R^th crime)

In this article,the heterogeneity index is used to measure the degree of inconsistency between the model judgment and the judge’s judgment.Specifically,for the heterogeneity index of the Rth crime,the denominator “median sentence” represents an average sentence level for the crime in past judicial practices and can be roughly considered the average sentencing by judges.The numerator measures the average difference between the model judgment and the judge’s judgment for each case of the crime,which is very similar to the standard deviation in the construction of the coefficient of variation.Thus,the heterogeneity index can to some extent represent the degree of inconsistency between the model judgment and the judge’s judgment.In this context,if the prediction values given by the model and the judge’s ruling are basically consistent,it can be concluded that no dispute between human and machine judgments exists for that particular sentence.However,if the difference between the two is large,then inconsistency exists between human and machine judgments,and sentencing bias may be present.The calculation shows that the crimes of producing,copying,publishing,selling,spreading obscene materials for profit,crimes of harboring and sheltering,and crimes of misappropriation of funds are the top three charges with the highest heterogeneity index,which indicates that these three charges are most likely to result in cases of sentencing deviation.

To further quantify the factors that influence sentencing deviation and identify the judicial characteristics of such cases,this article takes the crime of misappropriation of funds as an example and conducts related empirical analysis.The results of regression analysis show that mitigating circumstances,such as confession,and statutory or discretionary sentencing factors have a mitigating effect on sentencing.The size of the misappropriated funds is directly proportional to the length of the sentence.However,there are also some phenomena that contradict judicial interpretations.For example,the defendant’s purpose for misappropriating funds has an impact on sentencing,which may be due to the prominent heterogeneity of this specific crime.The goal of this article is not to replace the judge’s ruling by accurately predicting the length of the sentence through establishing a model.Rather,it is to identify cases of sentencing deviation based on the established sentencing mechanism,provide reference and support for sentencing judicial practice,and enhance the normativity of the judicial system.