CRF with dense features

Rasa Open Source
21

Ah! I think I may have found it. And I stand corrected. It does seem like the CRF model ignores features. It seems to even ignore the sparse features in the pipeline.

22

Yeah this is definately a bug.

These two pipelines yield the same results.

pipeline:
  - name: WhitespaceTokenizer
  - name: LanguageModelFeaturizer
    model_name: "roberta"
    model_weights: "roberta-base"
  - name: CRFEntityExtractor

pipeline:
  - name: WhitespaceTokenizer
  - name: CRFEntityExtractor

Making a note on the GitHub issue as well.

23

Thank you for the time command, I didn’t know it and it will be really helpful for my research.

I have already noticed that the sparse features are not taken into account but I tought that the CRF module generates them automaticaly from the given tokens.

So do you think that the dense features are being ingored (in that case, why would it be different when using DIET without transformer?) or is it simply a bug?

24

From what I undestood the time command give the time of the whole testing, is there any command to have the time for the prediction only?

Also, it makes sence that the LM config takes more time in testing because, even if it’s not taken into account, the LM Featurizer in still trained which is computationly expensive.