# Why word embedding such as word2vec is not used as the output layer of a seq2seq decoder?

It would make sense to make the decoder predict a smaller embedding vector instead of softmax over a large dictionary. The word having the most cosine similarity with the output embedding could be chosen as the output word.