All articles on Ultimate Guide to Cognigy NLU Training
The preceding sections concentrated on building the recognition baseline so that as many user inputs as possible trigger the proper Intent. In this article, we learn how to lower false positives /false recognition of Intents using various approaches.
Reducing False Positives
Once that baseline is established, we can have a look at reducing false positives / false recognition of Intents. Take the following example:
In our use case, these user inputs could result in some unfavorable screenshots, depending on the outputs we would provide. Other use cases could have additional reasons to avoid false positives, e.g., legal implications. We will go into the details on how to prevent false positives in the next sections:
Using the Reject Intent
The Reject Intent is the proper way to teach your model the use case boundaries. You train it with Example Sentences just like other Intents, but you only need to make sure that the Example Sentences do not map to the wrong Intent, its overlap with other Intents is perfectly fine:
You can see that adding the Reject Intent did solve our false positives issue quite efficiently. It is important to fill the Reject Intent with Example Sentences that use phrases that are also being used in our Intents but have a very different meaning overall. However, this will only cover the false positives that we have thought of so far, which brings us to the next section:
Using NLU Settings
You can also set the Confidence Threshold to higher levels than it was before, which will then prompt the Confirmation Question more often, allowing the user to provide more details:
This is a very time-efficient measure to take, please note though that this can negatively impact the user experience if your model is not sufficiently trained and has lower Intent scores in general.
We have covered all the strategies you can use to correctly map user inputs to your Intents and outlined best practices along the way. The only thing left to do is to implement them in your projects!
You can find all of the showcased Flows and Lexicons in the attached package: