I know there is LanguageFeaturizer that allows us to add token-level features from different models described in the rasa documentation. But even after experimenting with those, I couldn’t find much improvement in the accuracy.
The new sentence transformers models from https://www.sbert.net/ seems to have rich semantic embeddings in terms of sentence similarity. wanted to check if there is any way we can use the features from these models and improve the intent classification accuracy and fallback intent accuracy.