An agentic framework for data augmentation in construction defect report classification

Construction defect report classification faces critical challenges in resource-constrained environments, including severe class imbalance, limited training data, and domain-specific linguistic complexity. While transformer-based models are widely adopted for text classification, their effectiveness diminishes in specialized domains with short, technical reports and limited computational resources. Traditional oversampling techniques like SMOTE fail to generate linguistically coherent synthetic data that preserves domain-specific terminology and reporting conventions.This paper introduces a multi-agent data augmentation framework that addresses these limitations through structured synthetic data generation. Our approach employs an ``Analyst'' agent to extract domain-specific patterns from minority-class reports, coordinating multiple ``Worker'' agents that generate diverse synthetic samples using controlled lexical and stylistic variation. A human-in-the-loop validation process ensures domain fidelity and practical applicability.Evaluated on 3,243 real-world construction defect reports across 10 imbalanced classes, our framework demonstrates that classical machine learning models enhanced with intelligent augmentation can match transformer baselines and outperform traditional oversampling methods. Support Vector Machines with proposed agentic augmentation achieve 0.74 F1-score compared to 0.61 for SMOTE augmentation matching BERT's performance (0.75 at 20 samples, 0.74 at 40 samples). Random Forest with augmentation (F1-score: 0.76 at 30 samples) outperforms both BERT (0.75) and RoBERTa (0.71) baselines. The framework maintains semantic diversity (SBERT similarity < 0.4) and achieves 97\% expert acceptance rate, demonstrating both technical effectiveness and practical viability.Our findings challenge the assumption that transformer models universally outperform classical approaches, particularly in resource-constrained, domain-specific settings. The multi-agent augmentation framework provides a scalable solution for organizations with limited computational resources, advancing practical NLP applications in construction quality management and similar specialized domains with severe class imbalance.