Approaches of Neuro-Symbolic Integration: Large Language Models and Knowledge Bases

Abstract

The paper reviews modern approaches to the integration of neural network and symbolic artificial intelligence, including architectures using generative models in combination with knowledge bases and agents. Special attention is paid to the limitations of large language models (LLMs) in solving problems related to long-term context, as well as RAG-type mechanisms and modern agent-based systems are analyzed. An architecture of intelligent systems based on OSTIS Technology is proposed, in which LLMs are integrated into the problem solver. The necessity of formal description of generative AI methods as part of the problem solver is substantiated. The authors emphasize the advantages of moving the knowledge base from the position of a source of context for LLMs to the position of a shared semantic memory that combines different types of knowledge, from problem descriptions and solution models to the history of agent interaction. This approach provides automatic verification of knowledge, accumulation of experience and transparency of decisions.