Automated analysis and decision support in administrative law using deformable convolution network optimized by improved hippopotamus algorithm

The ever-growing complexity and increasing volumes of legal documents in administrative law make the need for robust and interpretable automated analysis systems even greater. In practice, many machine-learning approaches do not cope with the irregularities of the given text and demand extensive manual hyperparameter tuning, which limits their scalability and applicability in sensitive legal contexts. This study proposes a new framework incorporating a Deformable Convolutional Network (DCN) for adaptive feature extraction with a Decision Tree classifier optimized using the Improved Hippopotamus Optimizer (IHO). The DCN adaptively adjusts the receptive fields to capture the complex dependencies inherent in legal texts, whereas the metaheuristic IHO, achieves better classification accuracy by adapting the search strategy along with PSO to refine Decision Tree parameters. The proposed system is evaluated on the EUR-Lex dataset, a benchmark corpus of European Union legal documents, attaining the classification accuracy of 95.1%, which outperformed six state-of-the-art methods, including Legal-BERT and Gradient Boosting Machines. Detail-oriented quantitative as well as qualitative evaluations, including ablation studies and error analyses, further lend support to the competence of the system to equalize predictive performance with interpretability-an imperative requirement for an administrative law legal decision support application.