Gmail’s Inbox is ‘smart response’ to email




Google’s Gmail Inbox app will offer response options’ Smart Reply’ function. Smart Reply suggest three answers that are based on the emails that the user gets. For emails that require only a quick reply, ‘invents’ Inbox three different fast answers.

Initially, the function will only be available in English. Users who have their language set to English, the update “later this week” can expect as an update, writes the Gmail team on his blog. The system is “smarter” the more it is used. If all goes well, the answers are ‘sfw’ or suitable for the workplace.

To arrive at the responses, Google has set up a neural network which serves as the basis for the Smart Reply function. In a lengthy blog post on Google Research describes researcher Greg Corrado how the investigation in order to achieve the automatic answer function was performed. The sensible answer system is built on two recurrent neural networks to encrypt incoming mail and to predict possible answers. The coding network ‘reads’ the words of the incoming e-mail, word for word. It makes it a vector or list of songs. Which vector should capture the essence of what is being said tackle without remaining stabbing on language or vocabulary. As an example Corrado indicates that “Are you free tomorrow?” should be comparable to the vector as’ Does tomorrow work for you? “. The second network starts from that thought vector and takes advantage of two grammatically correct answers, again word for word. The amazing thing is, according to Corrado that “the entire operation of the network is fully learned, just by training the model to predict possible answers.

lstm network

One of the biggest challenges is that an e-mail often hundreds of words long. In addition, a special type of neural networks around the corner, called a “long short-term-memory network or lstm network. These networks will remember information for a long time, something recurrent neural networks, or RNN’s do, but in practice, the worse estate information link which lies further apart. Lstm’s because it can better know, these networks make up a meaningful response from the relevant sentences, without being distracted by intermediate information.

In the first prototype of the system was some strange reactions and other weird quirks. As the generation of candidate responses resulted in three similar answers very close to one another, such as “tomorrow we will come together, ” we will meet tomorrow ‘and’ what about tomorrow? ‘. Then, a system for bringing natural language in map added, whereby responses were more diverse. But the system did more crazy things, like standard ‘I love you’ to propose an answer, something that is not very useful as an answer in most cases.

Corrado obviously raises the security of the system, that privacy is guaranteed and there are no real people along. Something problems directly represents for researchers because they work with data sets that they can not read. Something like “solving a puzzle while you’re blindfolded.


In: Technology & Gadgets Asked By: [15446 Red Star Level]

Answer this Question

You must be Logged In to post an Answer.

Not a member yet? Sign Up Now »