Updated single candidate optimizer combined with convolutional neural network for interpretable legal document classification

Lawyers are increasingly insisting on artificial intelligence systems that are predictive and yet can be questioned and traced with understandable reasoning, with predictive and transparency commonly viewed as opposing each other in transformer-based models. In order to fill in this gap, a new USCO-enriched TextCNN architecture, namely updated single candidate optimizer (USCO) is proposed in this study that can perform interpretable classification of legal documents. The model, which combines jurisprudentially informed constraints, e.g. n-gram span bounds and an explanation-informed optimization, to promote adherence to patterns of legal reasoning, was validated on the CaseHOLD benchmark of 53,408 U.S. federal court holdings. The system was strictly compared to seven state-of-the-art baselines such as Legal-BERT and PSO optimized CNNs critically evaluated with stratified five-fold cross-validation and human interpretability evaluation through legal experts. Findings show that USCO-TextCNN is superior in Macro-F1 (88.9) and accuracy (89.7) and also yields much better results in the clarity (4.2/5) and actionability (4.0/5) of its clause-level explanation scores than transformer-based models. This paper confirms the possibility of harmonizing performance and transparency by domain-constrained metaheuristic optimization, which is a feasible way forward to the deployment of deployable, trustworthy AI in legal practice.