Siebel Smart Answer Administration Guide > About Siebel Smart Answer > Smart Answer Manager >

Training, Learning, and Feedback

Most customers use the training feature for initial KB creation and use the learning feature for ongoing maintenance. It is recommended that you use the learning feature to train new categories. If, however, you need to make extensive changes, you would probably create a new KB using the training feature.

Building Concept Models

During training and learning, the NLP engine builds a concept model to represent each category maintained by the KB file. The NLP and Statistical engines analyze texts using NLP and statistics, generating concept models for each category in the KB. NLP allows Smart Answer Manager to understand the intent of the text rather than just treating the text as a collection of randomly ordered words or strings. The NLP engine can locate unnecessary header information and allows Smart Answer Manager to disregard it when analyzing the incoming text.


Training begins by clearing all statistics from the KB, leaving only its structure, and then the KB populates the concept models of categories using a categorized collection of message texts (corpus).

For eMail Response, when training the corpus, you should submit emails in the same order that you might receive email on a typical day. You should not process a group of emails for the same category at the same time. The emails in the training corpus should be in a random sequence.

For Call Center and eService, you should train the KB based on the topics associated with the business objects that will be retrieved and displayed in response to users' natural language queries.

A corpus should contain examples of all categorized emails, or requests that you expect to receive in your production environment. The number of emails or requests in each category should be in proportion to the number that you receive in your production environment.

Categories in the KB will only return scores (confidence-level percentages) after they have gathered sufficient statistical information.


Learning is a gradual process that occurs as Smart Answer Manager receives feedback from processing emails or requests. This reshapes categories so that the concept models constantly evolve. As you process emails or requests, the NLP engine updates the concept models of existing categories.

When Smart Answer receives feedback, only positive feedback is applied to the learning process. Thus, a YES answer to a specific FAQ will be applied to the training of your KB, whereas a NO or NOT SURE will have no effect on KB training.


Feedback is processed when Smart Answer Manager receives text together with its correct classification. For eMail Response, this process can be as simple as a customer service representative who answers a customer inquiry by manually selecting a different category than the one that was suggested by Smart Answer. When the customer service representative clicks Send on the Communication Detail - Response view, Smart Answer Manager immediately modifies the concept models to reflect this new information.

In eService, for example, the KB receives feedback when the user clicks the Yes button in response to the "Does this answer your question" or "Was this selection helpful" questions.

Siebel Smart Answer Administration Guide