Embedding-Driven Legal Outcome Prediction: Explainable AI and Dimensionality Reduction Techniques

The problem of legal case classification is a notoriously difficult task because of the feature text type, the nature of legal language, and demand for accurate and comprehensible models. This work solves the issues of forecasts of judicial decisions by considering the multilingual and high-dimensional cases and making the oriented predictions more explainable. To achieve this, we utilize text embeddings, T5, Multilingual BERT, and XLM-RoBERTa integrated with machine learning classifiers: Random Forest and SVM. To make computations less resource-intensive while maintaining relatively high accuracy, dimensionality reduction through principal component analysis is applied; to ensure the reasons for certain decisions made by the AI models are understood, the data analysis technique known as Local Interpretable Model-agnostic Explanation is used. The approach uses not only the abstractive and extractive techniques for summarizing the legal judgments. Our approach shows that Random Forest with T5 embeddings have a training accuracy of 98% and a testing accuracy of 53%. Demonstrating the result of this work, the effectiveness of using embeddings and explainability for enhancing judicial analytics and advancing the efficiency and transparency.