Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Information Extraction From Fiscal Documents Using LLMs

Vikram Aggarwal, Jay Kulkarni, Aditi Mascarenhas, Aakriti Narang, Siddarth Raman, Ajay Shah, Susan Thomas

TL;DR

The paper tackles the problem of extracting structured data from multi-page, hierarchical fiscal PDFs using LLMs. It introduces a multi-stage pipeline that converts pages to high-resolution images, leverages sequential context across pages, applies multi-level validation based on fiscal hierarchies, and uses meta-prompting and semantic cleaning to produce machine-readable CSVs. Validation relies on internal numerical consistency and Tree Edit Distance Similarity (TEDS), achieving 73–96% structural accuracy across Karnataka's 2020-21 volumes and identifying precise failure locations for correction. The approach demonstrates that LLMs can process real-world, multilingual, complex tabular data, offering a scalable method for producing research-ready datasets and motivating reproducible research and cross-state deployment in developing economies.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in text comprehension, but their ability to process complex, hierarchical tabular data remains underexplored. We present a novel approach to extracting structured data from multi-page government fiscal documents using LLM-based techniques. Applied to annual fiscal documents from the State of Karnataka in India (200+ pages), our method achieves high accuracy through a multi-stage pipeline that leverages domain knowledge, sequential context, and algorithmic validation. A large challenge with traditional OCR methods is the inability to verify the accurate extraction of numbers. When applied to fiscal data, the inherent structure of fiscal tables, with totals at each level of the hierarchy, allows for robust internal validation of the extracted data. We use these hierarchical relationships to create multi-level validation checks. We demonstrate that LLMs can read tables and also process document-specific structural hierarchies, offering a scalable process for converting PDF-based fiscal disclosures into research-ready databases. Our implementation shows promise for broader applications across developing country contexts.

Information Extraction From Fiscal Documents Using LLMs

TL;DR

The paper tackles the problem of extracting structured data from multi-page, hierarchical fiscal PDFs using LLMs. It introduces a multi-stage pipeline that converts pages to high-resolution images, leverages sequential context across pages, applies multi-level validation based on fiscal hierarchies, and uses meta-prompting and semantic cleaning to produce machine-readable CSVs. Validation relies on internal numerical consistency and Tree Edit Distance Similarity (TEDS), achieving 73–96% structural accuracy across Karnataka's 2020-21 volumes and identifying precise failure locations for correction. The approach demonstrates that LLMs can process real-world, multilingual, complex tabular data, offering a scalable method for producing research-ready datasets and motivating reproducible research and cross-state deployment in developing economies.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in text comprehension, but their ability to process complex, hierarchical tabular data remains underexplored. We present a novel approach to extracting structured data from multi-page government fiscal documents using LLM-based techniques. Applied to annual fiscal documents from the State of Karnataka in India (200+ pages), our method achieves high accuracy through a multi-stage pipeline that leverages domain knowledge, sequential context, and algorithmic validation. A large challenge with traditional OCR methods is the inability to verify the accurate extraction of numbers. When applied to fiscal data, the inherent structure of fiscal tables, with totals at each level of the hierarchy, allows for robust internal validation of the extracted data. We use these hierarchical relationships to create multi-level validation checks. We demonstrate that LLMs can read tables and also process document-specific structural hierarchies, offering a scalable process for converting PDF-based fiscal disclosures into research-ready databases. Our implementation shows promise for broader applications across developing country contexts.

Paper Structure

This paper contains 16 sections, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. The Challenge
  3. The Opportunity
  4. LLM approaches
  5. Methodology
  6. Key Innovations
  7. Implementation: Karnataka Finances 2020-21
  8. Precise Problem Statement
  9. Approach
  10. Output Schema
  11. Cleaning
  12. Validation Results
  13. Computation Costs
  14. Future Directions
  15. Reproducible Research
  16. ...and 1 more sections

Figures (1)

  • Figure 1: Hierarchy in Karnataka finances