A cognitive hierarchy-guided abstract forest knowledge base based on cognitive load theory

The Retrieval-Augmented Generation (RAG) framework mitigates hallucination issues arising from questions beyond its knowledge boundaries by retrieving relevant knowledge to expand the knowledge scope of Large Language Models (LLMs). However, as the cognitive agent in the question-answering process, the RAG framework overlooks the implicit knowledge demands in complex questions, leading to a decline in the recall rate of pertinent knowledge. Furthermore, existing RAG frameworks neglect the intricate relationships among knowledge entries in the knowledge base, resulting in their reasoning processes lacking effective knowledge support. These issues indicate that the RAG framework may be misaligned with Cognitive Load Theory proposed by John Sweller during the cognitive process, thereby introducing additional cognitive load that hinder effective reasoning. To address these problems, we propose a Cognitive Hierarchy-Guided Abstract Forest Knowledge Base (CH-AFKB) based on Cognitive Load Theory. This approach conceptualizes knowledge in unstructured knowledge bases to expand the implicit knowledge within the existing knowledge base content. It also iteratively constructs an abstract forest knowledge base with ”concept-instance” characteristics from the bottom up to model relationships among knowledge entries. In experiments across three downstream tasks, the RAG framework with CH-AFKB demonstrated superior performance in most scenarios compared to strong baselines. Compared to existing RAG and LLM baselines, the CH-AFKB exhibited remarkable advantages in retrieval efficiency, knowledge integrity, and factual consistency.