Abstract
Corporate culture plays a pivotal role in shaping firm outcomes, yet existing wordembedding methods lack the semantic depth for sentence-level analysis. Large language models bridge this gap but at prohibitive cost, posing a fundamental trade-off between analytical accuracy and scalability. We propose an LLM-Enhanced Two-Stage Culture Measurement Model (LLM-TSCMM) that combines large language models with active learning to address this challenges. The model identifies culturerelated text and classifies it into specific dimensions through a two-stage pipeline, with a portfolio theory-based active learner optimizing batch selection at each stage to maximize annotation utility under budget constraints. We evaluate our framework on annual reports from over 4,000 Chinese listed firms spanning 2001 to 2024. Empirical results show that PT-AL significantly outperforms alternative methods under identical cost constraints, and that the derived culture metrics capture economically meaningful cross-firm variation. This work contributes to the literature by introducing a novel, efficient approach for large-scale corporate culture measurement, and establishes portfolio-based active learning as a cost-effective solution for massive LLMdriven text analytics.