LegalBERT++: Domain-Adaptive Pretraining for Explainable Legal Document Understanding and Judgment Prediction

The exponential growth of legal texts necessitates intelligent methods for their efficient analysis and interpretation. This paper introduces LegalBERT++, a domain-adaptive deep learning framework designed for explainable legal document understanding and judgment prediction. The framework combines a shared transformer-based encoder with a multi-task learning strategy that simultaneously addresses citation tracking and outcome classification, enabling effective transfer of legal reasoning patterns across tasks. Feature selection incorporates both textual and structural elements, including legal terminology, citation counts, and catchphrase keywords, thereby enriching contextual representations. Experimental evaluation on the Legal Text Classification Dataset (25,000 cases) demonstrates that LegalBERT++ consistently outperforms traditional machine learning and baseline transformer approaches. For citation tracking, the framework achieved an F1-score of 87.3%, surpassing vanilla BERT (81.5%) and LegalBERT (84.9%). For outcome classification, it attained an accuracy of 89.4% and a macro-averaged F1-score of 88.1%, improving over prior baselines by more than 2.3%. An ablation study confirmed the effectiveness of the multi-task learning paradigm and the incorporation of citation catchphrases, both of which contributed significantly to performance gains. In addition to quantitative improvements, interpretability was enhanced through attention visualization and SHapley Additive exPlanations (SHAP), which revealed that the model consistently emphasized legally salient terms such as precedents and outcome-defining expressions. This transparency strengthens the framework’s reliability and facilitates its integration into practical legal workflows. The findings establish LegalBERT++ as a scalable, interpretable, and high-performing solution for legal informatics. Future extensions will explore hierarchical architectures, multilingual corpora, and user-centered evaluations to further advance the role of artificial intelligence in legal research and judicial decision support.