Automated multi-label classification of risk clauses in construction contracts using GPT-driven data augmentation

Contract review is an important method to prevent risks. Prior research mainly utilizes single-label classification for risk clauses in contracts, with the data imbalance problem neglected. This research proposes an automated model for classifying risk clauses in contracts using Generative Pre-trained Transformer (GPT). To address data imbalance, a dual-stage approach is applied, where GPT-4o and GPT-2 are used for data generation and quality evaluation, respectively. The method is validated through 3411 sentences from international Engineering, Procurement, and Construction (EPC) contracts and 2807 generated sentences. Experimental results show that the GPT-based model outperforms the four baseline models. Data augmentation improves the model performance, as shown by a reduction in Hamming loss across all models, with the GPT model improving by 10.16%. This research contributes to advancing the application of GPT in contract risk analysis and provides an improved approach for addressing data imbalance in multi-label classification tasks. • The model identified 10 risk clause categories. • GPT-4o is an effective data augmentation strategy for addressing sample imbalance. • GPT-2 can evaluate augmented text quality to enhance sample quality. • The model achieved a Hamming loss of 0.0495 in multi-label classification task.