Problem with config after update and conversion of files

[Deprecated] Rasa X Community Edition
1

Hey guys, in the past I have always worked with Rasa X 0.41.1 and now updated to the newest version. I also needed to convert my files (domain, config, …). I did this with the automated migration. All the files look fine and also my mappings in the domain were converted to rules very well. After I built a model with the new files, the bot identifies every intent as an nlu_fallback with high probability – no matter what words I type. I guess there is some problem with my new config compared to the old one. My bot was mostly based on the mapping policy and the memoization policy. It looked like this:

language: de
pipeline: supervised_embeddings
policies:
  - name: MemoizationPolicy
    max_history: 2
  - name: KerasPolicy
  - name: MappingPolicy
  - name: FallbackPolicy
    nlu_threshold: 0.35
    core_threshold: 0.35
    ambiguity_threshold: 0.05
    fallback_action_name: action_default_fallback

The new transformed config is:

language: de
pipeline:
  - name: SpacyNLP
  - name: SpacyTokenizer
  - name: SpacyFeaturizer
  - name: RegexFeaturizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: char_wb
    min_ngram: 1
    max_ngram: 4
  - name: DIETClassifier
    epochs: 1
  - name: EntitySynonymMapper
  - name: ResponseSelector
    epochs: 1
  - name: FallbackClassifier
    threshold: 0.35
    ambiguity_threshold: 0.05
policies:
  - name: RulePolicy
    core_fallback_threshold: 0.35
    core_fallback_action_name: action_default_fallback
  - name: MemoizationPolicy
    max_history: 2

Does someone has an idea about what could be the problem here and why the nlu always identifies the fallback intent?

Thanks in advance!

2

You need to set the epochs for DIETClassifier, ResponseSelector to a higher number of epochs (e.g. 100)

3

Thanks, it helped :). With 100 it took very long to build the model and led to a timeout in my system, but I now tried it with 10 and this value works great for my case!