Google's machine learning Inbox can now reply to your emails

Machine learning can now reply to your emails.

A new feature in Google's Inbox app can recognise the content of emails and tailor responses using natural language, without a human being having to do a thing.

Machine learning is used to scan emails and understand if they need replying to or not, before creating three response options. An email asking about vacation plans, for example, could be replied to with "No plans yet", "I just sent them to you" or "I'm working on them".

The feature, dubbed Smart Reply, is only available in Google's Inbox app for Android and iOS. It has been designed for emails that can be answered with a short reply such as "I'll send it to you" or 'I don't, sorry'.

Google said the system would enable users to reply to emails in just two taps -- one tap to open it, one tap to select a response and send.
The responses a user chooses, or doesn't choose, will also improve future suggestions. According to Google software engineer Bálint Miklós, who first came up with the idea, the system used to suggest "I love you" as a suitable response to workplace emails.

This was as a result of what Smart Reply had learned from the emails it had scanned. As "Thanks", "Sounds good" and "I love you" were common responses, it gambled on them when it was unsure of what to say. The adorable anomaly has since been removed.
Smart Reply is built on a pair of recurrent neural networks, one that encodes incoming emails and one that comes up with possible responses. Each word is captured in turn to create a list of numbers, known as a thought vector, that gives the machine learning system the gist of what is being said. From this, the second network builds a grammatically correct response one word at a time.
"Amazingly, the detailed operation of each network is entirely learned, just by training the model to predict likely responses," explained Greg Corrado, a senior research scientist at Google.

 

Not only does Google's machine learning system have to understand the complex and varied way people communicate, it also has to digest emails that can be hundreds of words long.
To do this, Smart Reply focuses only on the most important sentences and ignores everything else. Google used a variant of a "long short-term memory" network to help the machine learning system hone in on what parts of incoming emails would be most useful for predicting a response.
Another challenge for Google's engineers was working blind. For privacy reasons, and to make a machine learning email system truly useful, no human being has access to the data. That means everything is worked out by a computer.
"This means researchers have to get machine learning to work on a data set that they themselves cannot read, which is a little like trying to solve a puzzle while blindfolded," Corrado explained.

Smart Reply will be introduced later this week as an update for Google Inbox for Android and iOS.

Share on Google Plus

About Unknown

This is a short description in the author block about the author. You edit it by entering text in the "Biographical Info" field in the user admin panel.