DEEP LEARNING PREDICTING MODEL FOR COURT VERDICTS

Abstract

This paper presents a deep learning-based prediction model for court verdicts, developed using historical datasets of court cases from multiple countries. The datasets were subjected to standard pre-processing techniques, after which the cleaned data was partitioned into training and testing sets for model development and validation, respectively. The model was implemented using Bidirectional Long Short-Term Memory Networks (Bi-LSTM) and N-gram models. During training, the Mean Square Error (MSE) function was employed as the loss function to monitor the variation between the predicted and actual outputs, optimized through the backpropagation algorithm. To handle complex nonlinear relationships and enhance convergence speed and stability, the ReLU activation function was applied. Furthermore, the Adam optimizer was integrated into the model to improve learning efficiency. The results achieved were highly promising, with the model attaining 99.07% accuracy on the training set and 98.01% accuracy on the validation set, alongside very low error losses of 0.002% and 0.003%, respectively. The findings demonstrate the model’s ability to accurately predict court verdicts and crime types, showcasing its potential to serve as a valuable tool for both legal practitioners and individuals in anticipating case outcomes prior to trial.