Any way to recreate BinarySingleStateFeaturizer performance with TEDPolicy?


I’m currently on Rasa 1.10.20, will be upgrading to 2.0 soon :crossed_fingers:

Made the switch over to the TED policy and I’m noticing that the predictions are not really following the stories all that well. A lot of my utterances use very similar vocabulary, so my hunch is that could be causing the incorrect predictions.

This is my current config:

    language: en

  - name: MultiProjectImporter

  - projects/Alternative Medicine
  - projects/Appt Followup
  - projects/Appt Prep
  - projects/Basic
  - projects/Best Treatment Quiz
  - projects/Conception
  - projects/Conditions
  - projects/Donor Eggs or Sperm
  - projects/Embryo Grading
  - projects/Endometriosis
  - projects/Entryway - Actively Trying
  - projects/Entryway - Exploring Treatment
  - projects/Entryway - Fertility Preservation
  - projects/Entryway - Intro Utterances
  - projects/Entryway - Preconception
  - projects/Entryway - Undergoing Treatment
  - projects/Exercise
  - projects/Fertility Preservation
  - projects/General Fertility Treatment
  - projects/General Infertility Info
  - projects/Genetic Testing
  - projects/IUI
  - projects/IVF
  - projects/Male Fertility
  - projects/Medications
  - projects/Mental Health
  - projects/Nutrition
  - projects/OI
  - projects/Other or Outside Current Scope
  - projects/PCOS
  - projects/Symptoms
  - projects/Testing
  - projects/Treatment Cost & Insurance

  - name: WhitespaceTokenizer
  - name: RegexFeaturizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: char_wb
    min_ngram: 2
    max_ngram: 5
  - name: DIETClassifier
    epochs: 300
    BILOU_flag: true
    use_masked_language_model: true
    number_of_transformer_layers: 4
    embedding_dimension: 50
      text: [256, 128]
      label: [256, 128]
    tensorboard_log_directory: logs
    tensorboard_log_level: epoch
  - name: EntitySynonymMapper
  - name: DucklingHTTPExtractor
    url: http://duckling:8000
      - distance
      - number
      - time
    locale: en_US
    timezone: America/New_York

  - name: TwoStageFallbackPolicy
    nlu_threshold: 0.5
    ambiguity_threshold: 0.01
    core_threshold: 0.1
    fallback_core_action_name: action_default_fallback
    fallback_nlu_action_name: flag_conversation_for_review
    deny_suggestion_intent_name: incorrect
  - name: AugmentedMemoizationPolicy
    max_history: 5
  - name: MappingPolicy
  - name: FormPolicy
  - name: TEDPolicy
    epochs: 150
    max_history: 5
    batch_size: [64, 128]
    random_seed: 6586
    tensorboard_log_directory: logs
    tensorboard_log_level: epoch

What I’m guessing is the culprit is the LabelTokenizerSingleStateFeaturizer and the way it featurizes the action labels and action text.

Is there a way to improve prediction and reduce the amount of featurization that happens on the action label/action text side?

I saw in the new docs that there’s an option for dense_dimension where I might be able to tune-down (or disable) action_text and label_action_text, but I’m unsure if that’s something the 1.10.20 TEDPolicy would be able to accept, since it’s not covered in the docs.

Hi @niveK TED in Pre-2.2 versions does not featurize the text of actions. It was only made possible with the 2.2 version which released end to end training.

With that in mind, can you post some stories that you expect to work and why do you expect them to work(for e.g. are there exact same/similar stories in the training data too?)