DIETClassifier is a neural Rasa model that jointly recognizes user intents and detects entities. It receives features produced by the NLU pipeline, rather than processing raw text alone. section 12.5.1
A typical pipeline prepares text before DIETClassifier runs:
- WhitespaceTokenizer splits the utterance into tokens. section 12.5.1
- RegexFeaturizer, LexicalSyntacticFeaturizer, and CountVectorsFeaturizer add signals for classification and entity detection. section 12.5.1
- Alternative pretrained embeddings from spaCy or Hugging Face Transformers can replace selected feature components. section 12.5.1
Rasa pipeline configuration
A Rasa configuration places DIETClassifier after tokenization and featurization, with settings such as epochs: 100 and constrain_similarities: true. Later components can map entity synonyms, select responses, and provide fallback behavior.
