Retrieval-augmented Chain-of-thought Prompting Method for Chinese Frame Semantic Parsing

Chinese Frame Semantic Parsing (CFSP) seeks to identify the semantic frames evoked by lexical units and assign roles to sentence constituents. To improve large language models (LLMs) in CFSP, we propose a method that integrates RAG with chain-of-thought (CoT) reasoning. Experiments on the CFN2.1 dataset show improvements of 13.52% in frame identification accuracy, 2.24% in argument F₁, and 5.09% in role F₁ over the best existing methods, respectively.