The paper proposes a two-phase framework called TnT-LLM to automate the process of end-to-end label generation and assignment for text mining using large language models, where LLMs produce and refine a label taxonomy iteratively using a zero-shot, multi-stage reasoning approach, and are used as data labelers to yield training samples for lightweight supervised classifiers. The framework is applied to the analysis of user intent and conversational domain for Bing Copilot, achieving accurate and relevant label taxonomies and a favorable balance between accuracy and efficiency for classification at scale.
