Unlocking Document Insights with Amazon Textract

We have been exploring the Google & Azure Document Intelligence features lately, in this article we will focus on AWS document intelligence offerings (mainly Amazon Textract). Intelligent Document Processing (IDP) on AWS consists of AI services that automate the extraction of data from documents. It offers managed machine learning services to handle tasks like OCR, form data extraction, table parsing, document classification and other document analysis features. These services can be used standalone or in combination to build end-to-end document processing workflows. Here is a quick overview of the set of services on AWS which can help with document capture tasks.  

Amazon Textract is the main AWS service for document text extraction. It goes beyond basic OCR and gives layout structure and relationships in the document. Textract ML models have been trained on millions of documents enabling it to detect and extract printed text, handwriting, form fields, and tables from scanned documents. All results include confidence scores and coordinate data for each extracted element so that developers can locate the text on the page and understand accuracy confidence.  

[Key capabilities of Textract] 

[1] Optical Character Recognition (OCR) - Reads both printed and handwritten text from documents and images and returns the text layer. 

[2] Form fields Extraction - Identify key value pairs in forms. 

 [3] Table Extraction - Recognizes tables and spreadsheets in documents. 

[4] Signature Detection - Detect the presence and location of signatures in documents. 

[5] Query-Based Extraction – Allows to ask questions of a document in natural language. 

[6] Specialized Document Support – Certain common document types like Invoice, receipts or tax documents models are pre-trained to extract entity fields which are specific to these document types. There is a Lending API also which is fine tuned for Loan/Mortgage packages.

[7] Layout - Ability to extract layout elements such as paragraphs, titles, lists, headers, footers, and more from documents. 

In upcoming articles we will explore more on how to use these services and build our understanding on Amazon Document Intelligence offering.  

[Summary] 

Textract is a key part of AWS Intelligent Document Processing (IDP) services to automate data capture from documents. AWS provides a collection of machine learning models for various document capture related tasks, including OCR, form and table data extraction, and document classification. To an extent it goes beyond basic OCR to extract layout and structural information, Textract is trained on millions of documents and can accurately detect and extract printed text, handwriting, form fields and tables. It also provides confidence scores and coordinate data, enabling developers to assess accuracy and locate text. Key capabilities of Textract include - OCR for reading printed and handwritten text, form field extraction for identifying key-value pairs, table extraction for recognizing tables and spreadsheets, signature detection, query-based extraction using natural language, specialized support for document types like invoices, receipts..., and layout analysis to extract elements like paragraphs and headers.  

Using these services most of the insights about the document can be captured and associated, however this much of data could be a lot and filtering these can be challenging in production IDP use cases. Also form fields, table and entity data can contain lot of noise, key is how to filter these data points and associate only important information to documents to push it to repositories and then make it searchable /actionable for downstream business applications.