The Test Workbench results are dependent on the bot definition and its corresponding test set. You can generate a test set with the information from the bot definition to pinpoint areas that need improvement. Create a test dataset with examples that you suspect (or know) will be challenging for the bot to interpret correctly considering the current bot design and your knowledge of your customer conversations.
Review your intents based on learnings from your production bot on a regular basis. Continue to add to and adjust the bot’s sample utterances and slot values. Consider improving slot resolution by using the available options, such as runtime hints. The design and development of your bot is an iterative process that is a continuous cycle.
Here are some other tips for optimizing your test set:
-
Select the most common use cases with frequently used intents and slots in the test set.
-
Explore different ways a customer could refer to your intents and slots. This can include user inputs in the forms of statements, questions, and commands that vary in length from minimal to extended.
-
Include user inputs with a varied number of slots.
-
Include commonly used synonyms or abbreviations of custom slot values supported by your bot (for example, “root canal”, “canal”, or “RC”).
-
Include variations of built-in slot values (for example, “tomorrow”, “asap”, or "the next day").
-
Examine the bot robustness for spoken modality by collecting user inputs that can be misinterpreted (for example, “ink”, “ankle”, or "anchor").
