A post on the Google Research Blog comes to reveal. Engineers will propose a new feature to GMail users. The Google email should offer a function of automatic response emails. The researchers declined to Gmail “Deep Learning” already algorithms used by Google voice search or auto-generation of thumbnails of Youtube. Based on the analysis of the message, the mobile application by GMail Inbox will propose 3 possible answers. With a click, the user will send the response which seemed the most appropriate.
Google reads your email… and answers. Isn’t life wonderful?
Deep Learning applied to messaging
The approach is called the Sequence to Sequence Learning. In the same way that Google translation algorithms analyze the meaning of a sentence in a language to generate its translation in a language, the algorithm for Gmail will analyze the message in order to generate a response sentence.
If the principle of operation is simple, generating consistent and relevant responses to the user is otherwise more complex that simply paraphrase the message of departure. Cannot create simple rules to create responses automatically both scenarios are endless. To provide the 3 most plausible answers, Ilya Sutskever, Oriol Vinyals and Quoc v. The, three researchers at the origin of this innovation have appealed to the artificial intelligence, neural networks to create responses based on the original message. Difficulty encountered by researchers, the three proposed responses must truly represent 3 different possibilities and not three different formulations of the same idea. The example presented in the post seems to demonstrate that the algorithm can be relevant even if the researchers note mentions a funny anecdote. Their algorithm was propensity to offer the reply “I love you” a little too frequently. Therefore, researchers have added their models to make slightly more… formal responses. The function should be activated this week on enforcement “by Gmail Inbox” later this week.
Translation : Bing Translator
Source : “Computer, respond to this email.“, Google Research Blog, November 3, 2015