-1

I’m using: EmbeddingIntentClassifier for Intent Classification.

According to RASA’s documentation, EmbeddingIntentClassifier requires dense_features and/or sparse_features.

In my NLU pipeline, I have CountVectorsFeaturizer and RegexFeaturizer and they both creates sparse_features.

I want to know which one of these two components' output: CountVectorsFeaturizer and RegexFeaturizer is used as an input for EmbeddingIntentClassifier?

What’s the role of each sparse_features?

Thank you,

Wiktor Stribiżew
  • 607,720
  • 39
  • 448
  • 563
kogha
  • 61
  • 1
  • 5

1 Answers1

1

All the featurizers that you put in your NLU pipeline are used as input features. So both CountVectorsFeaturizer and RegexFeaturizer are used. Sparse or dense define the way the outputs are represented in the code. We use sparse matrices where possible to save memory. Here is the video explaining how DIET is working: https://www.youtube.com/watch?v=vWStcJDuOUk&list=PL75e0qA87dlG-za8eLI6t0_Pbxafk-cxb&index=2

Vova Vv
  • 104
  • 3