Large Language Models for Economic Relationship Inference: Analyzing Environmental Regulation and Corporate Innovation

Understanding the complex interplay between environmental regulation and corporate innovation remains an important yet elusive issue in environmental economics and management. Traditional econometrics, bounded by the constraints of structured data, is limited in accounting for the complexity and variability of innovation, especially expressed in different textual forms. Although large language models (LLMs) are a promising way to structure information from unstructured data, its implementation for specific and domain-related economic work has been problematic in terms of trustworthiness, accuracy, and reliability. This paper introduces the Prompt-Enhanced GPT-4 Framework (PE-GPT4), a new model tailored for systematically and rigorously assessing environmental regulation's effect on corporate innovation performance. The proposed framework applies a multi-stage approach to prompt engineering that guides GPT-4 through determining regulation degree, categorizing innovation outcomes, and making causal inferences, within a system of checks. The research developed EcoInnovate-1K, a new dataset based on 1,000 case descriptions of enterprise-level cases uniquely and expertly annotated from a range of real-world sources. We benchmarked PE-GPT4 against other leading LLMs (including a GPT-4 model as baseline), Qwen3-7B, Claude 3 Opus, and Gemini Ultra, on EcoInnovate-1K as an evaluative method, and demonstrated PE-GPT4's clear superiority in evaluation metrics over all models.