Azure AI Document Intelligence - How do I incrementally train a custom extraction model?

Question

I am aware that custom extraction models do not support incremental training, only the custom classifiers. The previous workaround for this was to execute a model compose operation where we would combine the previously trained model with another model trained on the new samples. However, in v4.0 (GA), the model compose operation requires the training of a new explicit custom document classifier, in place of the implicit classifier that the previous iteration of the feature employed.

My issue is this — I need to train a custom neural extraction model, but I also need the ability to extend the extraction scope of the same to other templates in the future.

With v4.0 (GA), I cannot simply perform a model compose as I would need to train a custom document classifier for all the templates I am considering. Even if I were to do that, what would the process be for "incrementally training" the composed model?

Would I need to train another extraction model, incrementally train the custom classifier, and then perform another model compose ?

Also, are there any code samples for the incremental training operation on the custom classifier?

Answer 1

Hello ArkalShriyans-2355,

Welcome to the Microsoft Q&A and thank you for posting your questions here.

I understand that you like to train a custom neural extraction model and also need the ability to extend the extraction scope of the same to other templates in the future.

Regarding your scenarios and explanations, Azure AI Document Intelligence does not support incremental training for extraction models as of v4.0. The only viable approach is to use model compose with a classifier, which requires retraining extraction models for new templates. For true incremental extraction training, consider submitting a feature request to Microsoft: [Document Intelligence Feedback - https://feedback.azure.com/d365community/forum/093a6d87-1a24-ec11-b6e6-000d3a4f0f84

The Key Constraints in Azure AI Document Intelligence v4.0 was that:

• Custom extraction models cannot be updated incrementally (only classifiers can).
• Model compose combines multiple extraction models under a classifier but does not merge their capabilities.

Though, the classifier can be incrementally trained by adding new samples or classes. But when composing models, if a new extraction model is added, does the classifier need to be retrained incrementally to recognize the new template, and then composed with all existing extraction models? The process might involve retraining the classifier each time a new extraction model is added, which could be manageable but still requires careful versioning. If this workaround is what you deem fit that is a cool one.

The best advice would involve you following Microsoft's recommended approach for composing models - https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-model?view=doc-intel-4.0.0 and using the classifier's incremental capabilities while acknowledging the extraction model's limitations. - https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/incremental-classifier?view=doc-intel-4.0.0

I hope this is helpful! Do not hesitate to let me know if you have any other questions or clarifications.

---

Please don't forget to close up the thread here by upvoting and accept it as an answer if it is helpful.