The current NN architecture in embedding intent classifier doesn’t work at a word embedding level. Each sentence is represented as a Bag of Words representation and that BOW vector is embedded in the network. So you can’t get word level embeddings in this architecture.