The Google technologies that simulate be you writing emails

Google Technologies

Manage the email is often heavy and cumbersome task: many of the messages we receive and urgent need quick answers, but when their number is very high that process takes time.Or she wore, because Google has announced a new feature that will make from Google Inbox can generate predetermined automatic responses to various messages.

The feature Smart Reply that starts these days and are limited for the moment at support is a demonstration of the progress the company is making in machine learning (or machine learning) and combined with the use of natural language that can generate Google servers to simulate an amazing thing that really seem our automatic responses.

Google Technologies
Credit: Google Image

Deep Neural Networks into action

In the Google Research Blog we found detailed article in which Greg Corrado responsible for this development in particular, explained how he had managed to implement this feature in Google Inbox. The basis of everything is in machine learning and specifically in deep learning algorithms.

Such systems have been utilized in other Google services extensively. It occurs for example in the field of voice search Google, which uses the Deep Neural Networks (DNNS) as the technology on which these models are based. Improvements on Gaussian mixture models (GMM) were evident, and allowed to improve the speed and accuracy of speech recognition significantly.

Another application of these DNNS has been made ​​to obtain something as seemingly random as thumbnails YouTube videos . Behind this process are really detailed studies on how to choose the image that will finally make users click on that video because that model is striking.

How it works Smart Reply

The basis of this feature is the use of so-called sequence-to-sequence learning, a system that allows automatic conversational synthesis and which among other things served so that Google will create a nice experiment: A chatbot that debating the meaning of life with a human, and he did remarkably well.

The application of this system of learning for generating answers to all kinds of mailings was a challenge, but managed to solve the problem with the use of recurrent neural networks. One “encodes” incoming mail, the other generates possible answers.

The encoder post works studying the words of a message and generating a vector that allows “understand” the machine is being said and what tone. Not only that: it is able to identify two differently formulated phrases mean the same thing. Thus, this part of the system knows “Are you free tomorrow?” and “How about if we meet tomorrow?” They have the same goal.

The second system, the “decoder” takes the vector and synthesizes a correct answer word for word. To avoid responses of tens and hundreds of words in Google used a variant of a neural network of the type “Long Short-Term Memory” (LSTM) that allows the system to focus on the part of the mail that is most useful for predicting later response, downplaying less relevant before and after sentences.

The privacy flag

Throughout this system it had a clear potentially controversial component: privacy. The proper functioning of the system or machine learning machine learning is based precisely on training with millions of emails.

However, these emails are never read by a human, and as Corrado said, “That means that researchers had to run the learning of a set of data that could not read what it was like trying to solve a puzzle blinkers“.

Still the prototype worked after correction of some fine initial failures. For example, in the early stages of developing a continuously suggested by the system response was “I love you”. The system explained Corrado, “did just what he had trained him to do, generate probable answers, and it turns out answers like” thank you “,” it sounds good “or” I love you “are extremely common, so the system can support them as safe bet if he was not sure in other cases. “

Normalized to solve the likelihood of a response to past candidate contrasting answers, which made ​​the system less “loving”, but as Corrado says, much more useful. If you want, you can test the results using Inbox setup instructions for Android and iOS. We’ll see if your acquaintances detected that response is automatic.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.