Notice-Augmented Real-World Audit Report Generation by Large-Scale Complex Tabular Data Understanding and New Fields Discovery

Abstract

With the recent advances in large language models (LLMs), many commercial table-to-report generators have been released. However, existing systems rarely consider (i) mining potential audit items and (ii) incorporating data-collection notices, both of which are crucial for understanding the table context and the semantics of indices and values. To address this gap, we decouple tabular data understanding into a five-step sequential pipeline, including report framework initialization, table structure parsing, new field discovery, content analysis, and report generation. Empirical experiments and expert assessment show that our prompt-based pipeline can interpret notice files and understand tabular data, thereby generating audit reports. This workflow has been deployed in a data management system to support periodic report generation. Our demo video can be found at: https://youtu.be/9GTAAhoLu8Q.