Enhanced Privacy Policy Comprehension via Pre-trained and Retrieval-Augmented Models

Privacy policies are crucial for informing users about how their personal information is collected, stored, and used by organizations. However, privacy policies are lengthy, information-dense, and filled with legalese, making them difficult for users to comprehend. Although previous studies have attempted to improve the readability of privacy policies via traditional machine learning techniques, these methods overlook or oversimplify critical information or rely on predefined question-answers that cannot adapt to user personalized queries. Inspired by the fact that large language models perform well in terms of text comprehension and text question-answering, we propose a novel large language model-based privacy policy question-answering framework, which aims to help users understand privacy policies more intuitively and effectively. Specifically, our framework consists of two modules: the personal data practice disclosure module, which performs a pre-training and fine-tuning approach to extract structured information about data categories and corresponding data operations from privacy policies. The retrieval-augmented question-answering module integrates sparse and dense retrievers to find the most relevant evidence from the privacy policy and generate responses corresponding to user queries. The experimental results across two representative datasets demonstrate the superiority of our method over other prior methods.