Abstract

The increasing volume of unstructured textual data across digital systems presents significant challenges in transforming raw information into meaningful insights that support decision-making. Generative artificial intelligence has been increasingly adopted to improve information accessibility through automated summarization. In previous work, an iterative prompting approach demonstrated its potential to improve semantic consistency and contextual accuracy in summarization tasks. However, while summarization produces more compact representations of information, it does not fully address the need for structured and statistically grounded insights required for large-scale analysis. Therefore, this study extends the iterative prompting framework from summarization toward dynamic topic structuring and data-driven reasoning.

The proposed approach applies a two-stage iterative process. In the first iteration, an initial prompt generates semantic summaries and preliminary topic groupings from textual inputs. These outputs are subsequently analyzed to extract key statistical signals, including topic frequency distributions, proportions of dominant and minor topics, keyword overlap, and indicators of thematic fragmentation. This information is then incorporated as structured feedback to refine the prompt in the second iteration, guiding the model to consolidate redundant topics, reduce semantic noise, and improve the organization of topic structures.

Unlike conventional topic modeling methods such as Latent Dirichlet Allocation and Non-negative Matrix Factorization, which rely on predefined distributional assumptions and static topic structures, the proposed framework adapts dynamically to the characteristics of the input data. The iterative prompting process is positioned as an adaptive refinement mechanism analogous to statistical estimators, where data-driven feedback supports the formation of more stable and interpretable topic representations. This integration of generative language models with statistical analysis represents a methodological innovation that bridges generative AI with modern statistical practices.

Experimental evaluation is designed to assess the role of prompt iteration in supporting the formation of more consistent and interpretable topic structures through comparisons across prompt iterations, analysis of topic structure variations, and observation of changes in semantic patterns.

More broadly, this study highlights the critical role of modern statistics in transforming unstructured textual data into operationally relevant insights. The proposed approach has the potential to support applied research, strengthen data-driven governance, enhance organizational decision-making processes, and promote data-driven innovation with implications across multiple sectors and everyday life.