A Dynamic Retrieval-Augmented Generation Framework for Border Inspection Legal Question Answering

Border inspection legal question answering (LQA) is designed for specific legal scenarios, aiming to address legal questions related to border inspections and provide accurate and practical legal guidance to the public and border inspection departments. Current research requires a retrieval process regardless of whether the necessary knowledge already exists, and does not evaluate the legal information in the generated responses, leading to errors and low-quality content. we propose the Dynamic Retrieval-Augmented Generation framework for Border Inspection LQA (DRAG-BILQA), which generates responses first and then dynamically controls retrieval based on confidence scores. In this framework, we innovatively introduce a legal factor recognition module to ensure the legal accuracy of generated answers and improve the quality of the generation. We also present the BorderLegal-QA dataset, filling the gap in border inspection legal datasets. We conducted extensive experiments on both the BorderLegal-QA and the general LQA dataset JEC-QA. Experimental results show that DRAG-BILQA outperforms Sota models in terms of legal factor recognition accuracy and retrieval performance, achieving a METEOR score of 39.5 and improving generation quality. Case analysis also demonstrate that DRAG-BILQA effectively Augmented the legal accuracy of the generated content.