Subscribe / Unsubscribe Enewsletters | Login | Register

Pencil Banner

Why email is a better virtual assistant

Mike Elgan | Nov. 10, 2015
Google is inventing something better than Siri, Cortana, Alexa or even Google Now: artificial intelligence email.

In other words, instead of stealing your job, what if A.I. helped you do your job better? Google is applying that idea with something called Smart Reply.

If you get an email that invites a concise reply, the mobile Inbox apps now suggest three possible replies for you. If you like one, just pick it and choose "Send."

This, by the way, is how human assistants work: "Bob Jones wants to move your meeting to next Tuesday. Should I accept or insist we keep the current time?"

Google implemented the Smart Reply feature last week for iOS and Android versions of its Inbox email app.

Here's how to understand Smart Reply: Think of it as a chatbot that "cheats" on the Turing test by getting a second opinion from a human, who can choose one of the three options or ignore them and write her own. Either way, these replies will convince the person on the other end of the conversation that the interaction is with a human, because it is with a human -- an A.I.-assisted human. Human judgment and self interest are the gatekeepers that choose a convincingly human reply.

Smart Reply currently has something like 20,000 possible responses and is designed to learn and become more naturally conversational over time. And it learns, by the way, without reading incoming emails. The system does keep track of the "triggers" directing how the reply options will be constructed and also which reply is chosen. By monitoring these relationships, Google can improve and optimize the responses continuously.

Responses are typically in the three-to-six-word range.

Specifically, Smart Reply uses two kinds of artificial neural network technology. One, called an encoder, understands the email you get, and the other, a decoder, builds the possible replies.

Smart Reply starts by deciding whether an email needs a reply or doesn't need one, and whether the email can be answered concisely. If so, it then zeros in on the meaning or substance of the email, rather than the syntax.

From the research blog post: "The encoding network consumes the words of the incoming email one at a time, and produces a vector (a list of numbers). This vector, which Geoff Hinton calls a 'thought vector,' captures the gist of what is being said without getting hung up on diction."

In other words, it's not an if-then rule-based system. It simulates understanding and responds accordingly.

A senior research scientist at Google named Greg Corrado told me that this "thought vector" is an abstract mathematical representation of intent. "The vector points in a direction in hundred-dimensional space, and messages with similar intents should have vectors which point in similar directions," he explains.


Previous Page  1  2  3  Next Page 

Sign up for Computerworld eNewsletters.