Multi-Granularity Adapter Fusion with Dynamic Low-Rank Adaptation for Structured Privacy Policy Understanding

Privacy policies are legal documents. They describe how user data is collected, stored, and shared. These documents are highly structured and often use complex language. This makes automated understanding difficult. We propose PPAF-LLaMA (PrivacyPolicy AdapterFusion for LLaMA), a simple and scalable framework. It adapts large language models to structured document tasks. The method uses adapter blocks at the token, sentence, and paragraph levels. These adapters help capture different levels of meaning. We connect them using a cross-layer feature fusion method. This method brings together signals from different layers of the model. We also use a dynamic low-rank adaptation method. It changes the rank based on input complexity to improve both speed and performance. We train the model with a multi-task goal. It includes risk prediction and key information extraction. We also use regularization to make the model more stable. Results show that PPAF-LLaMA works better than fine-tuning or regular adapter methods for understanding privacy policies.