With the development and applications of the Smart Court System(SCS)in China,the reliability and accuracy of legal artificial intelligence have become focal points in recent years.Notably,criminal sentencing predictio...With the development and applications of the Smart Court System(SCS)in China,the reliability and accuracy of legal artificial intelligence have become focal points in recent years.Notably,criminal sentencing prediction,a significant component of the SCS,has also garnered widespread attention.According to the Chinese criminal law,actual sentencing data exhibits a saturated property due to statutory penalty ranges,but this mechanism has been ignored by most existing studies.Given this,the authors propose a sentencing prediction model that combines judicial sentencing mechanisms including saturated outputs and floating boundaries with neural networks.Building on the saturated structure of our model,a more effective adaptive prediction algorithm will be constructed based on the fusion of several key ideas and techniques that include the utilization of the L1 loss together with the corresponding gradient update strategy,a data pre-processing method based on large language model to extract semantically complex sentencing elements using prior legal knowledge,the choice of appropriate initial conditions for the learning algorithm and the construction of a double-hidden-layer network structure.An empirical study on the crime of disguising or concealing proceeds of crime demonstrates that our method can achieve superior sentencing prediction accuracy and significantly outperform common baseline methods.展开更多
基金supported by the National Natural Science Foundation of China under Grant Nos.T2293773,72371145,and 12288201the Special Funds for Taishan Scholars Project of Shandong Province,China under Grant No.tsqn202211004National Key Research and Development Program under Grant No.2022YFC3303000.
文摘With the development and applications of the Smart Court System(SCS)in China,the reliability and accuracy of legal artificial intelligence have become focal points in recent years.Notably,criminal sentencing prediction,a significant component of the SCS,has also garnered widespread attention.According to the Chinese criminal law,actual sentencing data exhibits a saturated property due to statutory penalty ranges,but this mechanism has been ignored by most existing studies.Given this,the authors propose a sentencing prediction model that combines judicial sentencing mechanisms including saturated outputs and floating boundaries with neural networks.Building on the saturated structure of our model,a more effective adaptive prediction algorithm will be constructed based on the fusion of several key ideas and techniques that include the utilization of the L1 loss together with the corresponding gradient update strategy,a data pre-processing method based on large language model to extract semantically complex sentencing elements using prior legal knowledge,the choice of appropriate initial conditions for the learning algorithm and the construction of a double-hidden-layer network structure.An empirical study on the crime of disguising or concealing proceeds of crime demonstrates that our method can achieve superior sentencing prediction accuracy and significantly outperform common baseline methods.
