From Responding to Co-Creating: Identifying Strategic CSR Processes Using Pre-Trained Language Models

Corporate social responsibility (CSR) is shifting from compliance-driven to strategically embedded practice, yet large-scale identification of strategic CSR (SCSR) remains methodologically challenging due to reliance on surveys or ratings. This study proposes a hybrid identification framework combining pre-trained language models (PLMs) and Latent Dirichlet Allocation to conduct paragraph-level semantic analysis on 3022 CSR reports from Chinese manufacturing firms. The approach effectively identifies and extracts the topics of both strategic and non-strategic CSR. Findings show that only 25.6% of CSR behaviors simultaneously possess strategic intent, business synergy, resource investment, and value output. SCSR practices are highly concentrated on value chain optimization and R&D collaboration, with significant variations across ownership types and industry environments. This study contributes a novel PLM-based methodology for SCSR identification and offers empirical evidence and practical tools for CSR quality assessment, regulatory oversight, and strategic decision-making in manufacturing sectors.