FINTAGGING: Benchmarking Large Language Models for Extracting and Structuring Financial Information

Corporate filings contain large volumes of numerical information that must be organized into standardized, machine-interpretable records for search, comparison, auditing, and decision-making. Although XBRL (eXtensible Business Reporting Language) provides the reporting infrastructure, assigning numerical facts to appropriate US Generally Accepted Accounting Principles (US-GAAP) concepts remains difficult. Evidence for a correct tag is often distributed across narrative disclosures, tables, and accounting definitions, while the full taxonomy contains more than 17,000 semantically related concepts. Many financial NLP benchmarks and XBRL tagging datasets reduce this problem to flat classification over limited label sets, underrepresenting the structural cues that determine how facts should be interpreted.We introduce FINTAGGING, a benchmark for evaluating large language models (LLMs) on taxonomy-grounded financial information extraction from corporate filings. Rather than treating tagging as a single prediction problem, FINTAGGING decomposes it into two stages. FinNI (Financial Numeric Identification) identifies numerical facts and assigns value types from mixed text-table inputs. FinCL (Financial Concept Linking) links each fact to the most appropriate US-GAAP concept. This formulation separates local fact extraction from global taxonomy alignment and exposes model failures.Zero-shot experiments show that LLMs recover many numerical facts but remain weaker at concept linking, revealing a persistent gap between numerical extraction and taxonomy-grounded semantic alignment. The AI contribution of FINTAGGING is a staged benchmark that disentangles fact extraction from concept alignment to expose LLM capability boundaries. Its engineering application is automated financial disclosure processing for auditing, regulatory compliance, and investment decision-making.