I have been experimenting with the Spacy pretrained pipeline and the Supervised Embeddings pipeline, along with the ResponseSelector component.
While I like the fact that Spacy -pretrained pipeline allows your to utilize pre-trained word embeddings, i find that the Supervised-Embeddings pipeline is more versatile (ResponseSelector uses this, supports multi intents, uses a more sophisticated classifier etc).
However, the limitation of the SupervisedEmbeddings pipeline as I see is that it does not let you utilize all the great work done by others (eg, use pre-trained word embeddings etc). Is there a workflow which would allow me to get the benefits of both worlds? I.e. can we initialize supervised embeddings with fast-text or other word vectors? I’m sure this could be done if I write my own classifier, but I’m trying to understand if there has any work been done in this direction? Also, is the ConveRTFeaturizer a tool which tried to achieve something similar?
I believe @dakshvar22 might be able to give me some insights on this