Configure a Trainable Evaluator
Use the procedure in this topic to configure a Trainable Evaluator.
The Trainable Evaluator trains the correct alternative from the input locators for each subfield. This training information is applied during extraction, and for each subfield, the best alternative is chosen from their respective input locators.
During configuration, it is important to add representative training documents for each subfield.
The input locators should not restrict the delivered alternatives to the correct values. The Trainable Evaluator also needs false alternatives to enhance the extraction results. Using the same input locator for different subfields is possible and can also enhance the extraction results.
Before you can configure your Trainable Evaluator settings, verify the following steps are complete:
-
A document set that contains a selection of documents from all classes is added to your project.
-
Any input locators in your project are fully functional and successfully tested.
-
Open the
Trainable Evaluator
Properties.
The properties for the Trainable Evaluator are displayed.
-
On the
General tab, use the
Add,
Delete,
Up, and
Down buttons to configure
Subfields table. Each subfield has the following column.
Setting
Description
Name
A descriptive name for the subfield.
Double-click in this field to edit the subfield name.
Input Locator
Select the input locator from the list. This list includes already existing locators and evaluators only, and must be defined above this Trainable Evaluator in the project hierarchy.
Micro Layout
This setting indicates where this evaluator searches for useful keywords in relation to a candidate. During training, the candidate is the trained value. At runtime, the candidate refers to each of the input locator alternatives as they are evaluated.
The following values are available for this setting:
- Table based:
-
Keywords are expected at the top or to the left of the candidate (N and W). Use this value when searching for content located in tables. (Default: Selected)
- Left top oriented:
-
Keywords are expected at the top and left of the candidate (N, NW, and W). For example, if the keyword is located in the middle of the page, the candidate should be found in the top left quadrant of the document. Use this value when searching for content located in captions or before colons.
- Symmetric:
-
Keywords are expected in all directions around the candidate. Use this value when searching for content that is located in running text.
- Line oriented:
-
Keywords are expected to the left and to the right of the candidate. (W and E). Use this value when searching for content in the same line.
If you change the value for the Micro Layout for a subfield that has already been trained, you must train the project again to ensure the knowledge base has the correct information.
- Open a test document and select Test to test your settings.
- Optional. Select Close to close the locator properties window.
- Save the changes to your project.