Google 's mail application "Inbox" will be implemented to automatically generate reply mail by machine learning

Application that automatically classifies mass mails that reach Gmail "Inbox"Automatic reply function utilizing machine learning"Smart Reply"The Gmail development team is announcing it on the official blog.

Official Gmail Blog: Computer, respond to this email: Introducing Smart Reply in Inbox by Gmail
http://gmailblog.blogspot.jp/2015/11/computer-respond-to-this-email.html

Smart reply is a function to automatically extract reply sentences by extracting e-mails that are short reply from received e-mails. As shown below, when opening the mail body, three reply example sentences are displayed at the bottom of the screen.


By choosing a sentence example and tapping it, the reply creation screen opens so you do not have to bother writing the reply sentence from scratch and you can easily reply. Of course it is also possible to add additional messages to sentences and send them.


Reply candidates are automatically generated according to the contents of the mail body, and up to three sentence examples are displayed. Inbox is a mail application that has a function to automatically generate reminders from the content of e-mails and to have functions to organize reservation e-mails of hotels and rental cars in chronological order to make it easy to understand the travel process, but Saku In order to be able to send mail quickly, "smart reply" function which automatically generates reply sentence by machine learning was developed.


The mechanism of the automatic reply function "smart reply" is explained in detail in the official Google blog below.

Research Blog: Computer, respond to this email.
http://googleresearch.blogspot.jp/2015/11/computer-respond-to-this-email.html

Smart reply said that development started, as Google staff felt it was annoying to receive large amounts of e-mails and write reply from mobile terminals every day. However, since the problem of not being able to deal with diverse e-mail contents only by making a fixed form sentence for reply occurs, we create a system that can deal with various situations by machine learning rather than just making a number of reply sentences I heard that he decided.

The smart reply employs a neural network "recurrent neural network" that can handle time series data, and it consists of two networks: a network that encodes incoming mail and a network that predicts and generates reply candidates. The former analyzes the words included in the received mail, converts the main point of the email to "thought vector", and the latter creates a reply sentence. The reply phrase example is what the system automatically generates, it seems that Google staff does not read the contents of the e-mail.


In the initial prototype of development, as a reply candidate for inquiries about the schedule, similar answers such as "How about tomorrow?" "May we meet tomorrow" "I think it is good to meet tomorrow" is generated, It seems that the reply function was sometimes not used. In order to solve this problem,Natural languageBy mapping and classifying machine learning systems, we succeeded in displaying reply sentences with different meanings as well as wording. After the official release, the accuracy of the example improves as more users select the appropriate reply candidate using smart reply.

In addition, the smart reply function, in the first week of November 2015Android versionWhenIOS versionIt is scheduled to be implemented in the Inbox application.