Inconsistent results in RASA1.1.3

After training RASA , i observed a specific text is mapping to wrong intent .So i did train again then its mapping to correct intent .Such inconsistency exists everytime i train , although the intents are of different context and dont share common utterances .

Results from rasa shell : Next message: X_Replacement_License { “intent”: { “name”: “smalltalk.pleasantry.you_are_a_chatbot”, “confidence”: 0.6313276290893555 } This intent is wrongly mapped . The intent “smalltalk.pleasantry.you_are_a_chatbot” dont have the above utterance trained.

Results After training again : X_Replacement_License { “intent”: { “name”: “X_Replacement_License”, “confidence”: 0.853678822517395 }

This is the expected intent .

So intents are switching for the specific text every time i train between those 2 intents .This scenario happens for other utterances too .

How to fix this ?

This blog post addresses some common issues in intent classification: Discover Actionable Ways to Improve your Training Data with Insights. Maybe it could help.

When two intents have training examples containing very similar language, the model has trouble distinguishing between them. Sometimes this happens during annotation, when a training example is labeled with the wrong intent by mistake. Or, intent confusion can result when intents overlap. In these cases, intent classification can be improved by merging multiple similar intents into one that’s more general.