A Large-Model-Based Dual-Stage Reasoning QA Approach with Multi-Hop Path Selection

Abstract

Knowledge reasoning and question answering (QA) is  significant in contemporary artificial intelligence. It simulates the logical reasoning processes of humans to derive new information from existing knowledge. However, current knowledge reasoning QA systems still face two challenges. On one hand, when dealing with multi-hop questions, the system needs to associate and reason across multiple pieces of knowledge paths, which not only increases computational complexity but also leads to inaccurate reasoning results. On the other hand, reasoning models often lack generalizability, they struggle to adapt to diverse real-world application contexts. These issues are particularly pronounced in industry and healthcare. For the industrial sector, equipment fault diagnosis and predictive maintenance require accurate understanding of complex technical terms and domain-specific jargon. These concepts are often highly specialized and abstract, making them difficult to reason. For healthcare domain, the high accuracy and reliability of QA systems are necessary. Any minor error in disease diagnosis could lead to serious consequences. Moreover, knowledge in these fields evolves rapidly, necessitating the system to update the knowledge base in real time. To address the aforementioned challenges, we propose a secondary reasoning QA algorithm enhanced by large language models for multi-hop questions. Our model leverages attention mechanisms to iteratively optimize path selection and utilizes the semantic understanding capabilities of LLM to further excavate the semantic background. Through experimental validation on four datasets, our model has demonstrated remarkable performance on multihop QA tasks, significantly enhancing the accuracy and generalization of the QA system.