Injury degree appraisal of large language model based on retrieval-augmented generation and deep learning

Abstract

Large Language Models (LLMs) have shown impressive performance in various natural language processing tasks. However, their application in specialized domains like forensic injury appraisal remains challenging due to the lack of domain-specific knowledge and the need for accurate retrieval of relevant information. This study proposes a novel approach that combines Retrieval-Augmented Generation (RAG) with graph-based knowledge bases and deep learning to enable LLMs to conduct injury appraisals based on China’s Standards for Assessing the Extent of Bodily Injuries (SAEBI). We create a dataset of 26,199 real-world injury appraisal cases and develop a RoBERTa-CNN model for accurate classification of injury locations and severity levels. By integrating this model with a graph-based knowledge base, our RAG strategy significantly improves the performance of nine state-of-the-art LLMs in injury appraisal tasks, with accuracy gains ranging from 21 to 59 percentage points compared to traditional retrieval methods. The additional experiments on crime classification also show that our method has good transferability in different domains. Our approach showcases the potential of combining domain-specific knowledge, advanced retrieval techniques, and deep learning to enhance the performance of LLMs in specialized domains like forensic injury appraisal.