Training using Search (Refine)

Use the Search action as part of the model training process, for labels that are less frequently occurring and don’t regularly appear in clusters and/or shuffle mode.

If there are minimal initial training examples for a label, you can use Search sparingly for one or more terms, for a given label. This provides sufficient examples for theTeach action to be available (for example by displaying roughly half of the examples relevant for that label).

Note: Using the Search action too much can lead to annotating bias, and over-fitting the model’s understanding of a label concept into specific terms/phrases, instead of understanding the broader context and variability of the concept itself. This means that you can overuse search unless guardrails are provided by the platform.

Using Search recommendation in Validation

1. Go to Validation.
   
   

   

   
   
2. Select one of the recommendations.
   
   

   

   
   
   You are redirected to the Discover page once you select a recommendation.

3. Search for terms or expressions related to the label you are searching for.

   
   

   

   
   
   Note: Apply examples sparingly using search for any label to avoid annotating bias.
4. Add a label, then click the Apply labels button to bulk annotate messages:
   
   

   

   
   
   
   

   

   
   
   Attention: Remember to also apply all the other relevant labels to the messages while searching, to avoid partial annotating.

Note: See Training using Search (Explore) to better understand how to use Search in the Train tab.