Comment by siavosh

Does someone know how the "semantic" embeddings are learned? That seems like perhaps the main technical challenge here.