If they want to translate some input to sensible output they would be better off using a translation model that has two LSTMs and a ton of data. See: https://gi

If they want to translate some input to sensible output they would be better off using a translation model that has two LSTMs and a ton of data. See: https://github.com/harvardnlp/seq2seq-attn https://github.com/harvardnlp/seq2seq-attn They are also not giving credit for the graphic they are using: http://colah.github.io/posts/2015-08-Understanding-LSTMs/ http://colah.github.io/posts/2015-08-Understanding-LSTMs/