DIET replacement?

Rasa Open Source
1

Hi all,

I have much love for the RASA community, but DIET hasn’t been sota for a long time now. Is the community working on something new?

According to the massive text embedding benchmark, SGPT-5.8B-nli or ST5-Large are performing far better than the LaBSE (which is the base model available for diet) at classification tasks. https://paperswithcode.com/paper/mteb-massive-text-embedding-benchmark/review/?hl=72261

2

Yes, you should try some alternatives for intent classification and entity recognition. Once simple change to improve performance is to disable DIET for entity extraction and try the CRF entity extractor:

- name: CRFEntityExtractor
- epochs: 100
  name: DIETClassifier
  entity_recognition: false

Another alternative is to replace DIET with your choice of featurizer and use the logistic regression classifier:

- name: LanguageModelFeaturizer
  model_name: "bert"
  model_weights: "sentence-transformers/all-MiniLM-L6-v2"
- name: CRFEntityExtractor
- name: LogisticRegressionClassifier
  max_iter: 100
  solver: lbfgs
  tol: 0.0001
  random_state: 42
  ranking_length: 10
3

Thanks Stephens,

will try both and give you some feedback.