Using Azure AI Document Intelligence Studio Custom Extraction Model [GCast 176]

Introduction

In this article, we'll explore how to utilize Azure AI Document Intelligence Studio, previously known as Forms Recognizer, to create a custom extraction model for processing invoices. This tool has undergone several updates and improvements, making it an effective solution for document analysis.

Getting Started

To begin using the Azure AI Document Intelligence Studio, navigate to documentintelligence.azure.com. This platform supports a variety of document formats, including PDF invoices, JPEGs, and Office documents.

Document Overview

When we initiate the process, we will work with a set of invoices that share a common format, although the text content varies. Key data points we aim to extract from these invoices include the customer name, address, invoice number, and date.

Creating a Custom Extraction Model

1. Project Setup:

   • After logging in, we can create a new custom extraction model by selecting the “Custom Extraction Model” option.
   • Name your project (e.g., “GCast Document Intelligence Custom”) and continue through the navigation prompts.

2. Service Configuration:

   • Specify the Azure subscription and resource group. In our instance, we'll create a new resource group named “gcast_doc_intops_RG.”
   • A document intelligence service will be configured with a chosen name, “gcast_doint,” set in the East US region with a standard pricing tier.

3. Storage Configuration:

   • Store documents in Azure Storage Blob Containers. We will create a new storage account, “gcast_doint_store,” also in the East US region, along with a blob container named “invoices.”

4. Document Labeling:

   • Once the project is created, we navigate to the label data tab to upload invoice documents. These can be added either by dragging and dropping or browsing files.
   • We choose to run layout analysis and manually label the documents one by one to identify key fields: vendor name, vendor address, customer name, and customer address.
   • We also create a table for line items in the invoice, specifying it as a dynamic table to accommodate varying row counts.

5. Training the Model:

   • After labeling the necessary fields, we initiate model training. A neural model is selected for simplicity, and we can use a single invoice sample for training due to the consistent layout.

6. Testing the Model:

   • Post-training, the model status updates to "succeeded." We test the model by uploading a new invoice to see if it accurately extracts fields such as the vendor address, customer name, and invoice number.
   • The model produces a confidence score indicating the accuracy of the extraction.

Conclusion

Once tested successfully, the Azure AI Document Intelligence Studio can streamline document analysis processes. A future article will delve into how to implement API calls to further interact with this model, offering expanded capabilities.

Keywords

• Azure AI
• Document Intelligence
• Custom Extraction Model
• Forms Recognizer
• Invoice Processing
• Cloud Storage
• Model Training
• Document Analysis
• Neural Model
• Confidence Score

FAQ

What is Azure AI Document Intelligence Studio? Azure AI Document Intelligence Studio is a cloud service that uses AI to extract information from documents, previously known as Forms Recognizer.

What types of documents can I process with Azure AI Document Intelligence Studio? You can process various document formats including PDFs, JPEGs, and Office documents.

How do I create a custom extraction model? You can create a model by setting up a project in the Document Intelligence Studio, labeling the required fields in your documents, and then training your model.

What is the purpose of labeling data in a custom extraction model? Labeling data is necessary to train the model on where to identify the required information in documents, allowing it to learn patterns and improve extraction accuracy.

What are confidence scores in the context of model testing? Confidence scores indicate the model's certainty in its predictions. Higher scores suggest greater accuracy in the extraction process.