Now you can create custom Classifiers in Caravel with no code and no training data.

With the addition of Zero Shot Classification and Ensembles, we’ve drastically improved the ease of creating Classifiers in Caravel and, in turn, made it easier for you to use NLP and AI for your specific use case using Caravel's user interface.

Zero Shot Classification

With this release we're introducing a new concept to Caravel, Zero Shot Classification. Zero Shot Classification is based on Zero Shot Learning, which, put simply, means the ability to learn with zero training data.

One of the most significant barriers to custom AI for any company is the requirement of lots of annotated training data. Zero Shot Classification eliminates this barrier. However, until now, it hasn’t been made accessible in a user interface.

So, how does it work? With Zero Shot Classification in Caravel, you come up with a label name. That’s it. There is no training data required.

Zero Shot Classification is available now within Ensemble type classifiers.

We encourage you to try it out. Since you don't need training data, it's simple to get started and test. Start with a template from our Prebuilt Classifier library. Just look for anything that mentions “Template” in the description. Or, if for the more ambitious, you can create your own from scratch.

Adding a zero shot label is done by inputting a Natural language definition


This brings us to the addition of another concept in Caravel, Ensembles. In Machine Learning, ensembles are a strategy where you use multiple models to serve as a single model. Usually, this technique leads to better results than training a single model.

In Caravel, Ensembles are similar in that you can use more than one model to classify a label. Dissimilarly, you can use more than just models.

With Ensembles, you can combine Zero Shot Classification with Rules & Conditions classify your text. Using Rules & Conditions, you can classify text using a combination of Prebuilt models, sentiment, and even various text matching algorithms from simple exact search to more advanced queries and approximate "Fuzzy" search.

To create your first Ensemble, you can pick a template from the Prebuilt Classifier library or choose Create ensemble as your method when creating a new custom Classifier.

New Create ensemble option available from custom classifier creation flow

Adding Rules & Conditions to an ensemble

Multi-Label Classifiers

With the elimination of training, Zero shot makes creating custom Classifiers very easy. However, we knew we could take another big step to make it even easier, give Classifiers the ability to have multiple labels.

This also adds the benefit of keeping your classifications a lot more organized.

New Classifier pill design in Caravel with multiple labels under one Classifier

Classifier Previews

Predictions are only as good as the inputs into the algorithm. We continually strive to bring transparency to what is happening under the hood and make it easy to understand so that you can get the best predictions out of your Classifiers.

When you create a Classifier, you'll have an interactive Preview area to test any text you want and see what your Classifier thinks of it. We also provide score logs that give you insight into your Classifiers confidence of each label and whether or not your text passed your Rules & Conditions.

Viewing scoring logs from the new Classifier preview area

Emotions Prebuilt Classifiers

Emotion data allows you to draw correlations between your customer's emotional responses and your behaviors. With data on emotions mapped across your customer's journey, you can tie their emotional reactions to different KPIs and performance metrics.

For example, you can identify areas your users get confused while onboarding and benchmark your onboarding experience against confused users.

You can also use emotional data to measure your customers' responses in their 1:1 interactions and coach your team to be more empathetic.

We offer two Emotion classifiers: Emotions (All) and Emotions (Basic).

With Emotions (All) you can predict 27 different emotions:

Admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, and surprise

With Emotions (Basic) you can predict 6 basic emotions

Happy, surprised, sad, angry, fearful, and disgusted.

Emotions prebuilt classifiers are available for use today in the prebuilt library

You can try out these new updates by going to and creating your free account.

Head over to our updated docs to read all about how Classifiers work.

Other improvements & fixes

  • New and improved classifier pill design
  • Ability to ignore any classification within your feed
  • Simplified classifier creation flow
  • Removed the need to select intent, topic, or redaction
  • Removed the step of jump-starting. Now available within the Tagging classifier type when editing it's trained sentence list.
  • Improved historical analysis options - select one or multiple sources to analyze and a date range.
  • Ability to turn sentiment prediction on/off for each Classifier
  • Ability to edit the confidence of each custom classifier
  • Customer Experience, Bug, and Apparel classifier templates in the prebuilt library
  • Classifier help docs