Strange Behavior in RASA NLU

Rasa Open Source
1

I am building bank intent chatbot, having a problem when numbers are added alone

intent called transfer

inside this intent, we have examples like

  • I want to transfer money
  • I want to transfer 20$

when I test the intent by 20 only, it interprets the intent to be transfer , strange behavior

NLU Pipeline

pipeline:

  • name: WhitespaceTokenizer
  • name: CountVectorsFeaturizer
  • name: DIETClassifier
  • name: rasa_addons.nlu.components.gazette.Gazette
  • name: EntitySynonymMapper
2

Hi Ahmed,

We have an example bot with just these types of intents. Take a look here and you should find just what you’re looking for.

Greg

3

Hello @stephens

The demo looks very interesting. I am having question-related to DIETClassifier if we have numbers in one of the intents then test it using just number the result is strange, it misclassifies the intent. Biasing the result due to the number while when I try EmbeddingIntentClassifier correctly identify the intents

4

Not sure what you mean by test it using just number. If you’re talking about a user utterance that is literally a number only, 93.00, then this is a different issue.

For that, we recommend an “inform” intent. You can also see this in the same example bot here. Typically, this type of intent might be used as part of a form like the transfer_form in this bot.

You’ll find a discussion regarding the “inform” intent idea in this NLU blogpost under the Similar Intents heading.

Greg

5

when the user says 5 it will be classified as inform_transfer my assumption is that it’s the only intent that contains numbered examples when using EmbeddingIntentClassifier this case doesn’t happen. What is your recommendations for use case

image
image1341×580 167 KB

Numbers are misclassified when tested alone. previously it was working fine.

** NLU Pipeline**

pipeline:
  - name: WhitespaceTokenizer
  - name: CountVectorsFeaturizer
  - name: EmbeddingIntentClassifier
  - BILOU_flag: true
    name: CRFEntityExtractor
    features:
      - [low, title, upper]
      - [low, bias, prefix5, prefix2, suffix5, suffix3, suffix2, upper, title, digit, pattern]
      - [low, title, upper]
  - name: rasa_addons.nlu.components.gazette.Gazette
  - name: EntitySynonymMapper

Thanks in advance