Author ORCID Identifier
Tao Yang: 0000-0003-4676-3006
T. Robert Yu: 0000-0002-2248-5577
Huimin Zhao: 0000-0002-6471-9837
Abstract
Organizations increasingly need to analyze large volumes of unstructured but information-rich narrative data that are important for decision-making. Topic-modeling methods can identify underlying themes at scale, but their outputs often require substantial interpretation before being used in downstream domain-specific analysis. End-to-end LLM analysis can assist with interpretation, but it raises concerns about transparency and reproducibility. This study develops and evaluates a multi-stage framework that combines embedding-based topic modeling, domain-guided alignment, and bounded LLM-assisted interpretation. The framework implements three design principles: functional decomposition, complementary capabilities, and inspectability. A cognitive cost-benefit perspective provides a rationale for this design logic. We apply the framework to a corpus of CEO ESG narratives from U.S. publicly traded firms. Structural diagnostics and human evaluations indicate strong semantic coherence, interpretability, and thematic distinctiveness across the extracted themes. We further illustrate how the extracted topics can be operationalized in a capital-market setting. The study shows how a decomposition-based framework can transform semantically rich but weakly standardized narratives into structured and interpretable representations for subsequent analysis.
Recommended Citation
Yang, T., Yu, T., & Zhao, H. (In press). A Decomposition-Based Framework for LLM-Augmented Topic Analysis of Weakly Standardized Narratives. Communications of the Association for Information Systems, 59, pp-pp. Retrieved from https://aisel.aisnet.org/cais/vol59/iss1/72
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
