<p>Knowledge graph reasoning is a critical research direction in the field of artificial intelligence. Current reasoning methods excessively rely on the topological structures of knowledge graphs to generate static rules, which often ignore the semantics of relations. Static rules have limitations in diversity and quality, making it difficult to adapt to complex reasoning tasks. To address these issues, this paper proposes ChatER, a dynamic rule generation framework that can dynamically generate logical rules with structural features and semantic features deduced by large language models (LLMs). A confidence scoring mechanism is adopted to quantify rule reliability, effectively alleviating logical brittleness. The optimized high-confidence rules are directly utilized for inference and collaboratively integrated with knowledge graph embedding (KGE) methods to establish a fact-consistency constraint mechanism. Finally, experiment evaluated ChatER’s performance. The performance of multiple benchmark datasets is superior to the current latest baseline methods.</p>