Hello, recently i tried to find out how the tf embedding model works in rasa. However, I have a couple of unresolved issues. The developers write on the blog that the model learns word embeddings, but we send all the messages in tf-idf style in the network. Where in this case are word embeddings formed? Is the same dictionary used for encoding intents in a vector for feeding into the network as for user messages? Or we just one-hot intents?
word embeddings are the outputs of the embedding layer inside the classifier