Document Intelligence custom neural model

Important

• Document Intelligence public preview releases provide early access to features that are in active development.
• Features, approaches, and processes may change, prior to General Availability (GA), based on user feedback.
• The public preview version of Document Intelligence client libraries default to REST API version 2023-10-31-preview.

This content applies to: v4.0 (preview) | Previous versions: v3.1 (GA) v3.0 (GA)

Custom neural document models or neural models are a deep learned model type that combines layout and language features to accurately extract labeled fields from documents. The base custom neural model is trained on various document types that makes it suitable to be trained for extracting fields from structured, semi-structured and unstructured documents. Custom neural models are available in the v3.0 and later models The table below lists common document types for each category:

Documents	Examples
structured	surveys, questionnaires
semi-structured	invoices, purchase orders
unstructured	contracts, letters

Custom neural models share the same labeling format and strategy as custom template models. Currently custom neural models only support a subset of the field types supported by custom template models.

Model capabilities

Custom neural models currently only support key-value pairs and selection marks and structured fields (tables), future releases include support for signatures.

Form fields	Selection marks	Tabular fields	Signature	Region
Supported	Supported	Supported	Unsupported	Supported 1

¹ Region labels in custom neural models use the results from the Layout API for specified region. This feature is different from template models where, if no value is present, text is generated at training time.

Build mode

The build custom model operation supports template and neural custom models. Previous versions of the REST API and SDKs only supported a single build mode that is now known as the template mode.

Neural models support documents that have the same information, but different page structures. Examples of these documents include United States W2 forms, which share the same information, but can vary in appearance across companies. For more information, see Custom model build mode.

Supported languages and locales

See our Language Support—custom models page for a complete list of supported languages.

Tabular fields

With the release of API versions 2022-06-30-preview and later, custom neural models will support tabular fields (tables):

• Models trained with API version 2022-08-31, or later will accept tabular field labels.
• Documents analyzed with custom neural models using API version 2022-06-30-preview or later will produce tabular fields aggregated across the tables.
• The results can be found in the analyzeResult object's documents array that is returned following an analysis operation.

Tabular fields support cross page tables by default:

• To label a table that spans multiple pages, label each row of the table across the different pages in a single table.
• As a best practice, ensure that your dataset contains a few samples of the expected variations. For example, include samples where the entire table is on a single page and where tables span two or more pages.

Tabular fields are also useful when extracting repeating information within a document that isn't recognized as a table. For example, a repeating section of work experiences in a resume can be labeled and extracted as a tabular field.

Supported regions

As of October 18, 2022, Document Intelligence custom neural model training will only be available in the following Azure regions until further notice:

• Australia East
• Brazil South
• Canada Central
• Central India
• Central US
• East Asia
• East US
• East US2
• France Central
• Japan East
• South Central US
• Southeast Asia
• UK South
• West Europe
• West US2
• US Gov Arizona
• US Gov Virginia

Tip

You can copy a model trained in one of the select regions listed to any other region and use it accordingly.

Use the [REST API](/rest/api/aiservices/document-models/copy-model-to?view=rest-aiservices-2023-10-31-preview&preserve-view=true &tabs=HTTP) or Document Intelligence Studio to copy a model to another region.

Input requirements

• For best results, provide one clear photo or high-quality scan per document.

• Supported file formats:

  Model	PDF	Image: JPEG/JPG, PNG, BMP, TIFF, HEIF	Microsoft Office: Word (DOCX), Excel (XLSX), PowerPoint (PPTX), and HTML
  Read	✔	✔	✔
  Layout	✔	✔	✔ (2023-10-31-preview)
  General Document	✔	✔
  Prebuilt	✔	✔
  Custom	✔	✔

  ✱ Microsoft Office files are currently not supported for other models or versions.

• For PDF and TIFF, up to 2000 pages can be processed (with a free tier subscription, only the first two pages are processed).

• The file size for analyzing documents is 500 MB for paid (S0) tier and 4 MB for free (F0) tier.

• Image dimensions must be between 50 x 50 pixels and 10,000 px x 10,000 pixels.

• If your PDFs are password-locked, you must remove the lock before submission.

• The minimum height of the text to be extracted is 12 pixels for a 1024 x 768 pixel image. This dimension corresponds to about 8-point text at 150 dots per inch (DPI).

• For custom model training, the maximum number of pages for training data is 500 for the custom template model and 50,000 for the custom neural model.

• For custom extraction model training, the total size of training data is 50 MB for template model and 1G-MB for the neural model.

• For custom classification model training, the total size of training data is 1GB with a maximum of 10,000 pages.

Best practices

Custom neural models differ from custom template models in a few different ways. The custom template or model relies on a consistent visual template to extract the labeled data. Custom neural models support structured, semi-structured, and unstructured documents to extract fields. When you're choosing between the two model types, start with a neural model, and test to determine if it supports your functional needs.

Dealing with variations

Custom neural models can generalize across different formats of a single document type. As a best practice, create a single model for all variations of a document type. Add at least five labeled samples for each of the different variations to the training dataset.

Field naming

When you label the data, labeling the field relevant to the value improves the accuracy of the key-value pairs extracted. For example, for a field value containing the supplier ID, consider naming the field supplier_id. Field names should be in the language of the document.

Labeling contiguous values

Value tokens/words of one field must be either

• Consecutive sequence in natural reading order without interleaving with other fields
• In a region that don't cover any other fields

Representative data

Values in training cases should be diverse and representative. For example, if a field is named date, values for this field should be a date. synthetic value like a random string can affect model performance.

Current Limitations

• The model doesn't recognize values split across page boundaries.
• Custom neural models are only trained in English. Model performance is lower for documents in other languages.
• If a dataset labeled for custom template models is used to train a custom neural model, the unsupported field types are ignored.
• Custom neural models are limited to 20 build operations per month. Open a support request if you need the limit increased. For more information, see Document Intelligence service quotas and limits

Training a model

Custom neural models are available in the v3.0 and later models.

Document Type	REST API	SDK	Label and Test Models
Custom document	Document Intelligence 3.1	Document Intelligence SDK	Document Intelligence Studio

The build operation to train model supports a new buildMode property, to train a custom neural model, set the buildMode to neural.

https://{endpoint}/documentintelligence/documentModels:build?api-version=2023-10-31-preview

{
  "modelId": "string",
  "description": "string",
  "buildMode": "neural",
  "azureBlobSource":
  {
    "containerUrl": "string",
    "prefix": "string"
  }
}

Next steps

Learn to create and compose custom models:

Build a custom model Compose custom models