Five AI models were tested to evaluate their ability to extract invoices and parse tables from real business documents. The tasks focused on invoice field extraction and structured table parsing using 20 invoices and 20 tables. The evaluation criteria included accuracy, speed, cost, and stability against messy inputs. AWS Textract emerged as a top performer in invoice extraction, achieving 91.3% accuracy, and proficiently managing standard fields without making assumptions. While it excelled with flat structures, it faced challenges with more complex table formats.
AWS Textract excelled in extracting structured fields from invoices with a 91.3% accuracy rate, avoiding hallucination and ensuring strict adherence to input data.
When dealing with complex tables, Textract achieved 82.1% accuracy, demonstrating a stronger performance than GPT-4o and Azure, particularly in flat structures.
Collection
[
|
...
]