Sequence-to-sequence models have quite a few practical applications.

Perhaps the most practical application is machine translation. We can use machine translation to take a phrase in one language as input and output that phrase in another language. Machine translation is an important service that we depend on more and more. Thanks to advances in computer vision and machine translation, we can listen to a language we don't know, or look at a sign in a language we don't know, and have a pretty good translation almost immediately on our smartphone. Sequence-to-sequence networks really have gotten us very close to Douglas Adam's imagined babel fish from The Hitchhiker's Guide to the Galaxy.

Question-answering can also be accomplished in whole or in part by sequence-to-sequence models, where we can imagine the question as an input sequence and the answer as an output sequence. The most generalized application of question-answering is chat. If you support an enterprise with a call center, you have thousands or maybe millions of question/answer pairs that pass over the phone every day. That's the perfect training set for a sequence-to-sequence chat bot.

There are several nuanced forms of this question-answering idea that we can exploit. Every day, I get roughly 3.4 billion emails. Of those, I probably only need to read 20-30 (and that's a classification task); however, my responses to those emails are rarely novel. I could almost certainly create a sequence-to-sequence network that would write my emails for me, or at least draft a response. I think that we are beginning to see behavior like this built in to our favorite email programs already, and more fully automatic responses are sure to come.

Another great use of sequence-to-sequence networks is in automatic text summarization. Imagine a set of research papers or a big stack of journal articles. All those papers probably have an abstract. This is just another translation problem. We can use a sequence-to-sequence network to generate an abstract, given some paper. The network can learn to summarize documents in this manner.

Later in the chapter, we will implement a sequence-to-sequence network to do machine translation. Before we do that though, let's understand how this network architecture works.