The NLU Fallback Intent - Avoiding False Positives

NLU systems can be complex since utterances can come in many forms. Being able to tune which utterances should not match to any intent is important. To do this, you need to extend the Fallback intent.

The Fallback intent (sometimes called None) is how many NLUs detect irrelevant information or information you don’t want your NLU to match against. The Fallback intent helps eliminate False Positives. Let's walk through some examples on how adding to the Fallback intent can improve your conversational design and experience.

Example #1

You are building a project about banking, you offer debit cards but not credit cards. Credit cards and debit cards are very similar so the NLU will likely recognize both, even though that is not your desired state.

Solution: Add credit cards to the None intent

And now we match to the correct intent.

Example #2

You are building a dentistry app, and don’t want the NLU to match to other medical procedures. 

Solution: Add Back, Leg, Arm to None intent

Prebuilt Fallback

The way the Voiceflow NLU works is we add utterances that don’t fit the themes of your project to our Fallback intent. So your Fallback intent should work out of the box for most projects.

Using the Fallback Intent vs Another intent

One of the important things when designing chatbots is understanding how users will interact with them. Imagine you’re building a restaurant bot and a customer says the name of a rival burger like “Big Chat Burger”, how should your bot respond?

Below are three valid options:

1. “Sorry we don’t have Big Chat Burgers here, but we can offer you AI Burgers”

2. “Sure you can have a AI Burger”

3. ”Sorry I did not get that, what would you like?”

Each option is valid, and it is up to the conversational designer, product manager or business analyst to decide which approach is correct.

Implementing Option 1

To implement option one, you should create a rival burger intent, and add all the names of burgers inside. We've added Big Chat Burgers and a Grandpa Chat Burger into Competitor_Burger Intent.

Implementing Option 2

To implement option two we add ChatBurger into our AI_Burger intent so it will have a higher chance to match.

Implementing Option 3

We add ChatBurger into the Fallback Intent.