Year: 2026
Type: other
Abstract: PDF is one of the most widely used file formats for saving and sharing data globally, and accurate parsing of its content is crucial for process automation in data-intensive industries such as Finance and Health, as well as in research fields such as NLP. Many recent studies have focused mainly on generic documents and simple extraction tools. This research compares commercial-grade AI extraction tools and LLMs on complex, real-world financial tables with borderless layouts, overlapping columns and cell values, multipage tables, rotated pages, and noisy data. It also uses a traditional PDF parser as a baseline to prove why AI/LLM-based solutions are necessary for complex files. The tools analyzed for this research were PyMuPDF, Docling, AWS Textract, Google Document AI, Mistral AI, GPT-5, and Claude Sonnet. We assessed table extraction capabilities of these tools on scanned and digital PDF files. We evaluated each tool for extraction accuracy, processing speed, cost-effectiveness, and robustness to document complexity. This evaluation used a dataset of 50 insurance commission statements, both scanned and digital, with varying levels of complexity. The evaluation results varied according to the complexity of the documents. In general, Large Language Model-based tools performed better overall in terms of accuracy with Mistral offering the strongest accuracy-to-cost trade-off among them. Computer Visionbased tools showed mixed results, with good performance on low to medium complexity and poor performance on highly complex documents. Rule-based tools performed poorly on borderless and shaded layouts even on nominally simple documents. This research provides a guideline for academic and industry researchers seeking an optimal, cost-effective solution for complex, unstructured table data.
Source: Doria (University of Helsinki)
Author: Ahmad, Muhammad
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162905742