Gmail’s Inbox companion app now uses machine learning to automatically generate email replies for you

If you haven’t yet transitioned over to Google Inbox, Google’s neural networked-powered email client that runs alongside Gmail, the latest “Smart Reply” feature may be your best incentive. The feature aims to make responding to emails simple and intuitive by automatically generating three short context-appropriate responses for you to pick from and send; it’s like up-scaling auto-correct to generate complete responses.

At first glance the feature may appear to parallel the quick replies found in most messaging apps, but Smart Reply is far more complex, leveraging a cloud-powered deep neural network tuned for natural language processing similar to the underlying network responsible for improved voice search functions and YouTube’s video suggestions.

“What if there was a system that could automatically determine if an email was answerable with a short reply, and compose a few suitable responses that I could edit or send with just a tap?” asks Greg Corrado, Google’s Senior Research Scientist.

It’s like a second level of auto-correct that may eventually evolve into a gamified way of compiling emails by selecting a series of words to automatically generate a coherent sentence. Although, the current implementation solely provides responses in basic, short phrases, but this may evolve in time.

Inbox’s machine-learning system generates its responses by scanning your email for context, before providing a list of three suitable responses. It can capture diverse writing styles and tones by generalizing better than any rule-based system possibly could, and the more information it absorbs, the better it more efficient and accurate it becomes.

As with other sequence-to-sequence models, function is split between a pair of recurrent neural networks; one is used to encode the email white the other predicts possible responses. More specifically, the encoding network deciphers all the words in the incoming email and produces a vector which captures the gist of what’s being said without being too concerned with the diction. Meaning “are you free tonight?” is similarly labeled to the vector for “does tonight work for you?” Next, the second network creates a grammatically correct reply one word at a time.

Bear in mind that while Smart Reply may offer a heightened-level of convenience, overreliance may deteriorate one’s ability to write similar to how spellcheck as worsened our spelling. Remember the film idiocracy?

Source: Ars Technica and Google