Skip to main content
To train and test a skill, you need labeled documents whose field locations and data types are explicitly marked. The fastest way to get a labeled set is to label documents manually, but that’s time-consuming, especially for Deep Learning activities that need large training sets. To save time, you can reuse labeled documents from skills you’ve already built, from manually reviewed processing results, or from ABBYY FlexiCapture.

​
Label documents manually

For each uploaded document, mark out the field locations and specify the data type each field is expected to contain. Follow the labeling guidelines for reliable training results.

​
Reuse labeled documents from another source

You can import labeled documents from three sources. In each case, export the labeled documents from the source to a folder, and then import that folder into your skill on the All Documents set.

​
Skill training set

When you publish a skill for production, you usually remove the training set and keep only a few sample documents in the published version. To reuse the same training set for a new version of the skill, export it before publishing. To export, click the more icon next to the document set name and choose Export Set with Labeling. The destination folder contains:
  • documentdefinition.json
  • skillsettings.json
  • classificationassignments.json (only for skills with classification activities)
  • A <Document name> subfolder for each document, containing the document images, documentinfo.json, and labeling.json

​
Manually reviewed processing results

Manual review produces labeled documents as a side effect. Set up exports of:
  • Field data to JSON with the Values, metadata, and field structure for each document option enabled.
  • Document images in any supported image format.
The destination folder contains one subfolder per transaction. Each <Transaction ID> subfolder contains:
  • <Applied skill name>.json with the field data.
  • Exported images: <Applied skill name>.pdf, <Applied skill name>.tiff, or a <Applied skill name>/Pages subfolder with page_*.jpg files.

​
FlexiCapture

You can reuse documents labeled earlier in ABBYY FlexiCapture. For the format and procedure, see Import labeled documents from FlexiCapture.

​
Import the labeled documents

After you have an export folder from any of the sources above, import it into your skill:
1

Open the All Documents set

On the Documents tab, select the All Documents set.
2

Choose Import Labeled Documents From Folder

Click the drop-down arrow next to Upload and choose Import Labeled Documents From Folder….
3

Select the export folder

Select the folder you created with the export procedure for your source.
Don’t change the folder that was created during export. Renaming files or altering the subfolder structure can break the import.
If a document with the same name already exists, Advanced Designer asks how to handle the duplicate:
  • Update Labeling – Overwrites field locations and settings on identically-named fields, and adds any fields present only in the imported document.
  • Import As New Documents – Renames the duplicates and imports them with their labeling intact.

​
Export and import labeling for Splitter skills

Export and import with labeling also works for Splitter (Assembling) skills, not just Extraction skills. Use the same Export Set with Labeling and Import Labeled Documents From Folder actions described in Import the labeled documents. When you export a skill that has classification activities, the export’s root folder includes a classificationassignments.json file. It stores the skill’s classification ground truth, recording the class assigned to each document in each classification activity. Advanced Designer re-applies those assignments automatically when you import the documents into another skill.
Import restores the classification assignments, not a trained classification model. Retrain the classification activity after importing. For more information, see Train a classifier and view classification results.
Classification handling has these limits:
  • Unreachable classifications are never included in an export.
  • Import fails when an imported classification conflicts with the target: a matching classification name with a different output field, or a matching field name with different or additional class values.
  • If a classification doesn’t exist in the target skill, Advanced Designer creates it but leaves it unreachable. Modify the pipeline to connect it.
  • Advanced Designer matches classifications and fields by name, not by internal ID. If names collide between the source and target skills, an import can land in the wrong classification, so keep names distinct.