Skip to main content
Verification operators can improve data extraction by retraining NLP models. When FlexiCapture fails to detect a field or confuses one field for another, the operator indicates the correct field and retrains the model, and FlexiCapture then uses the retrained model for more accurate extraction.
Additional training is not available for NLP models loaded into Document Definitions.

​
Start training during verification

Initiate training in one of two ways:
  • Add a training stage after the verification stage. Training starts when the conditions set for the training batch are met. For more information, see Workflow setup.
  • Send documents to the training stage manually: right-click the document in the working batch and select Train.

​
How training works

  • When training starts, FlexiCapture creates a generic training batch (if one does not already exist) and copies every document for the Document Definition into it, regardless of variant.
  • Each document is assigned the For training or For testing status.
  • Documents marked For training are used to build a new NLP model.
  • The new model is tested against the documents marked For testing.
  • If the new model performs no worse than the existing one, it replaces it. Otherwise, it is rejected.
If some documents share the same fields but place them very differently, create separate training batches for each variant to improve recognition.

​
Create a training batch for a specific vendor or variant

1

Open the project

On the Project Setup Station, open the project with the NLP model. For more information, see Create NLP models.
2

Open Field Extraction Training Batches

Select Fields Training → Open Field Extraction Training Batches, press Ctrl+Alt+B, or select Field Extraction Training Batches on the shortcut menu.
3

Create the batch

Select File → New Batch (or press Ctrl+N). Choose the Document Definition and variant, and then select NLP Batch on the shortcut menu.
4

Add documents and train

Add your documents, recognize them, edit the order of sections, and start training with Train on the shortcut menu, Ctrl+F7, or the Train Batch button on the toolbar.
The quality of a trained NLP model depends on the number of training documents and the quality of their markup:
  • Mark up every field described by the Document Definition in the training documents.
  • Include 100 to 500 documents in each training batch. This range lets FlexiCapture choose the best parameters without slowing training.
When operator feedback is used for training, new documents are added to both the training batch and the variant batch:
  • For a variant that has its own training batch, the model created for that batch is used.
  • For all other variants, the model created for the generic training batch is used.
If a document identical to one already in a training batch is added from the same source, the new document replaces the old one, and this is recorded in the background task log. FlexiCapture uses the document registration parameters to decide whether a document is a copy.

​
Training batch settings

After creating the batch, select Show NLP Batch Settings to set the following options.

​
Export training statistics

After training finishes, you can export statistics for an NLP model, including:
  • The training batch settings.
  • Information about the new and old NLP models.
  • The training time.
  • The version of the NLP component used.
  • Document and field training statistics.
  • How recent the exported data is. If the isActual parameter is false, the batch was modified after training (documents added or removed, markup changed, and so on). Launch training again for up-to-date statistics.
To export the log, right-click the batch, click Export Field Extraction Statistics on the shortcut menu, and specify where to save the CSV file.