HYBRID NLP AND STACKING APPROACH FOR ROBUST DATA EQUILIBRIUM IN CLEARANCE ADJUDICATION SYSTEMS

Abstract

The rapid growth of digital records in legal and defence domains has created a pressing need for intelligent systems capable of extracting meaningful insights from large volumes of unstructured textual data. Security clearance adjudication, which involves analyzing case digests, categorical records, and appeal decisions, requires consistent, unbiased, and accurate decision support. To address these challenges, this study proposes an integrated framework that combines Natural Language Processing (NLP), data balancing techniques, and ensemble learning strategies for enhanced predictive performance in defence clearance adjudications. The research focuses on two binary classification tasks: Favourable Decision (approval or rejection) and Decision Upheld (maintained or reversed). Baseline models such as K-Nearest Neighbors (KNN), Logistic Regression (LR), and Multinomial Naïve Bayes (MNB) are initially evaluated but exhibit limitations in handling imbalance and contextual variability. The proposed approach incorporates NLP-based preprocessing with the Synthetic Minority Over-sampling Technique (SMOTE) to balance class distribution. A stacked classification model is then employed, where Random Forest (RF) acts as the base learner and Logistic Regression serves as the meta-classifier. This hybrid design combines robust feature learning with optimized decision boundaries, resulting in improved classification accuracy and balanced performance. Experimental results, validated using confusion matrices, ROC curves, and accuracy metrics, demonstrate the superiority of the proposed model. The framework is adaptable to other domains requiring reliable binary decision-making with unstructured data.