Dynamic knowledge correction via abductive for domain question answering

Abstract

Domain question answering with large language models (LLMs) often relies on previously learned domain knowledge. Previous methods typically used large language models for direct reasoning to obtain results, which have poor reasoning ability due to complexity or timeliness of domain knowledge. In this paper, we propose an abductive-based dynamic knowledge correction for large language models reasoning framework (AKC). Specifically, we first identify domain knowledge sources based on task relevance to construct a domain-specific knowledge base. Then, we decompose the initial results generated by the large language model into individual elements and perform minimal inconsistency reasoning in conjunction with the domain knowledge base to dynamically correct erroneous reasoning outcomes. Experiments on three domain-specific datasets-law, traditional Chinese medicine, and education-demonstrate that the AKC framework significantly improves LLM accuracy in domain-specific question answering.