In addition to the trainable content classification that is provided by the Adaptive Feature Classifier, you can add explicit instructions to classes to handle exceptions. Instructions are words and phrases that can be combined using Boolean operations. This means that if a word is present on a document, the corresponding classification confidence changes.

Negative instructions can be used to restrict the classification of a class. If these negative words are present on a document, the corresponding classification result confidence is lowered. If the negative value is big enough, it results in an overall classification result confidence of 0%.

These instructions are processed after layout classification but at the same time as content classification is performed. This is because recognition is required before the instructions can be processed, and performing it with the content classification ensures that recognition is present. The content classifier result is combined with the Instruction Classifier result. If the Project Tree contains both parent and child classes, additional hierarchical rules are applied before the final classification result is determined.

These instructions are configured independently for each class, and are maintained within the Project Tree.

For example, use an Instruction Classifier to help distinguish between two types of letters. Since letters are not typically consistent in content or layout, a classification instruction can improve your classification results if specific words or phrases are present on the document.

You can maintain your instruction classifiers in the following ways:

• Add a classification instruction to a class

• Modify a classification instruction

• Delete a classification instruction

• Combine several classification instructions

• Import classification instructions

• Export classification instructions

• Import classification instructions using Import as New

Example: Using Instructions to Help with Classification

For example, if you have a project that processes letters where one is a mortgage inquiry and the other is a credit check report, your Project Tree might look as follows.

During classification, neither the Layout Classifier nor the Adaptive Feature Classifier can to classify these documents as anything other than a Letter. However, this project is configured so that the Letter class is not a valid classification result. To ensure that these letters are correctly classified, adding the following instructions to each child class can help ensure that these letters are classified.

By adding both positive and negative instructions to both classes, you can influence the final classification result. The following example shows the instructions and their relevance for each of the child classes.

In this example, the same instructions are used with different relevance values for each child class. Because these letters never contain both phrases, it is a good way to distinguish them. If a letter contains the word Mortgage then the letter is likely classified as MortgageInquiries. This is because the presence of that word significantly lowers the classification result for the CreditChecks class, yet increases the confidence of the MortageInquiries class. Similarly, if the document contains the phrase Credit check, the document is classified as CreditChecks because the classification confidence for MortgageInquiries is very low.