Hi,
I watched the video about the DIETClassifier. I dont understand how tokens can be processed by pretrained embeddings.
In the default (non-english) DIET-Pipeline, there are no pretrained embeddings like spaCy oder BERT configured for usage.
Is the output of the pretrained embeddings step an empty vector in such a pipeline?
DIET is a plug n play model you can use any custom pretrained embeddings. Though Rasa provides Convert and Bert embeddings in this blog How to Use BERT in Rasa NLU And i think if there is no pretrained embeddings defined in the pipeline then there wouldn’t be any empty vector either.
Yup, if there are none provided they simply won’t be used; if they are provided, the token will be “looked up” in the pretrained embeddings to get a representation (vector) which will be added to the sparse features.