How to resolve a model build error.
Introduction
As you train your chatbot you should rebuild its natural language processing (NLP) model over several iterations. This will increase the accuracy of intents and produce a stronger F1 score. Sometimes errors related to utterances prevent a model build.
Utterances
Most of the time a problem associated with utterances prevents me from initiating a model build in the Bot Builder. It's either that an intent does not meet the minimum threshold of 20 utterances so Einstein should not be toggled on, or there are duplicate utterances. Identify intents that have a low number of utterances and manually insert examples to boost the count. Also, export utterance records (Intent Utterance [MIIntentUtterance object]) via Dataloader to identify and delete duplicates.
Dataloader
Process the exported records to identify duplicates. Set aside the record ids of the duplicate utterance records in a separate CSV file for deletion via Dataloader.
Sample Excel formula: =IF(COUNTIF($C$2:$C$21812,C2)>1,"Dupe")
- Using conditional formatting on a large quantity of rows in Excel to identify duplicates can result in slow processing performance, or even crash the application.
After the delete operation has complete, check for errors. Inspect intents that may have a low utterance count from the deletion process. It may take some time for the utterance update to propagate, sometimes a few hours. Return to Model Management in the Bot Builder and attempt to build a model.
Model build in progress