NLU Training

Use training data to improve your Speech Assets

Overview

Enabling the Training Data option when creating a new release allows the NLU to retain and utilize the new utterances provided by users. These utterances are first checked against existing training data for an exact match to ensure that they are assigned to the correct intent as intended during their initial training. If there is no exact match, the system then uses predictive analytics to determine the intent. This ensures higher accuracy and consistency in intent recognition.

Enabling and Collecting Training Data

Activation

To enable Training Data, toggle on the Enable Training Data option during the fourth step of the Release process.

Once activated, the NLU system conducts two primary checks for each new utterance:

• It searches for an exact match within the existing training data. If a match is found, the utterance is directly assigned to the related intent.

• If there is no match, and the utterance meets certain criteria, it is added to the Training table unless it is already included in the Speech Assets.

Managing Training Data

Navigation

To view and manage collected utterances:

1. Click on the Deployments tab, located in the upper right corner of your screen.

2. Select Training from the dropdown menu to access the Training table.

   

Filtering Entries

In the Training tab, you can view and evaluate the collected utterances. Exact matches with retained training data are automatically linked to their respective intents. For non-matching utterances, the system offers options to integrate them into Speech Assets, thus improving the database and refining the predictive matching process.

Confidence Level

The Confidence Level enables you to filter utterances by the confidence score (from 0% to 100%) assigned to their detected intents.

Intents

The Intents filter enables you to display utterances linked to specific intents by selecting from this filter.

Date

The Date filter enables you to filter utterances collected within a specific date range.

Manage the Collected Training Data

Below the filters, you can see the table where the collected utterances are listed. You can also order them by date or by their assigned intents by clicking on the respective columns. The rightmost column is called 'Actions,' which offers three different actions: Add, Ignore, Context:

Add

By clicking on the respective utterance will be added to the selected intent. By default, the intent is selected, which was triggered by this particular utterance, but you can also select another intent to which you want to add the utterance:

The number in parentheses behind the pre-selected intent indicates the confidence level with which the intent was picked when the listed utterance was handled.

Ignore

Clicking on will delete the respective entry from the table. Be careful with doing that since an entry cannot be restored afterward.

Context

You can gain more information on how the respective utterance was originally used by clicking on .

As a result, a sidebar will open on the right showing the exact user's conversation in which the utterance was first used. This certain utterance will be highlighted in dark blue in the shown dialog:

The text beneath a user's utterance states the intent the utterance was assigned to.

Data Anonymization

To protect user privacy, all utterances that have PII data displayed in the Training tab are anonymized, meaning that PII such as names, birthdates, addresses, and other sensitive information is replaced with fictitious data. This ensures that user-sensitive data is anonymized before it is used for training the NLU speech recognition model. For more information, you may refer to the Anonymizing Personally Identifiable Information page.

Continue to the next page for a comprehensive guide to training best practices.