LABDT (Lexi AdaBoost Decision Tree): An Approach for Legal Outcome Prediction Fusing Lexical, Semantic, and Similarity-based features

Legal outcome prediction is a complex task due to the nuanced vocabulary and reasoning embedded in judicial documents. This study proposes LABDT (Lexi AdaBoost Decision Tree), a novel hybrid framework that integrates lexical features (TF-IDF), semantic embeddings (Sentence-BERT), and similarity-based metrics for robust and interpretable legal decision classification. The model addresses class imbalance through SMOTE and reduces feature dimensionality via principal component analysis (PCA). LABDT was evaluated on a real-world dataset of approximately 18,000 cases from the Federal Court of Australia case records, spanning eight outcome classes. The results demonstrate that LABDT outperforms traditional and state-of-the-art classifiers with an accuracy of 92%, precision of 91%, recall of 91%, and F1-score of 91%. Notably, it achieved highly accurate classification on certain outcome classes like ‘approved’ and ‘related’. LABDT offers a superior balance between predictive performance and model interpretability compared to baseline models. The system’s explainable design and high classification reliability position it as a viable tool for AI-driven legal analytics and judicial decision support.