Enhancing Reasoning Pathways for Power Defect Analysis Using CoT and Role-Play Prompt

Modern Large Language Models (LLMs), such as ChatGPT, demonstrate exceptional capabilities in text classification and reasoning. The categorization of severity levels for descriptions of Power Defects and the inference of Defect Causes present an innovative and challenging task aimed at providing comprehensive and accurate reasoning pathways to power grid workers. In this study, a comparison is made among three Chain-of-Thought (CoT) prompting methods and the Role-Play prompting method using a Power Grid dataset. It is observed that the manually designed Manual-CoT method achieves the best results, with other methods showing significant improvements in classification accuracy and the coherence of reasoning pathways. This further highlights the potential for substantial enhancement of Large Language Models’ reasoning abilities in specialized domains through expert-guided template pathways.