A new feature in Google’s Inbox by Gmail will let users reply to emails faster by simply selecting a response from choices served up to them by the app based on the contents of each email message.
The new Smart Reply feature, announced at the Web Summit event in Dublin, Ireland, Tuesday, will become available later this week on Google Play and Apple’s App Store.
The feature harnesses technology from Google’s extensive work around machine learning to recognize emails that require a quick response and generate a natural language response on the fly, Google Software Engineer Bálint Miklós wrote in a blog Tuesday.
Smart Reply will suggest up to three replies for emails that can benefit from a quick response. For instance, the tool might suggest a quick “I’m working on it” response to an email requesting the status of a particular request.
Or for emails that require a little more analysis, the tool will give users a jump-start so they can respond right away, Miklós said. For instance, in response to an email inquiring about an individual’s vacation plans, Smart Reply might suggest the response: “I’m working on them” and then let the user fill in the rest of the response.
The Smart Reply system uses machine learning to analyze the contents of an email and guess an appropriate response to it using natural language processing. The technology is based on what Google describes as a pair of recurrent neural networks. One of the networks encodes incoming email while the other predicts possible responses.
The network that does the encoding captures each word of an incoming email and attaches a list of numbers to it in a process designed to teach the computer how to understand the meaning of a word in the context in which it is used.
The approach leverages a new artificial intelligence concept called “thought vectors” that is used to teach computers to get at the gist of what is being said, Google Senior Research Scientist Greg Corrado said in his description of the technology.
“For example, the vector for ‘Are you free tomorrow?’ should be similar to the vector for ‘Does tomorrow work for you?'” Corrado noted.
The other network uses the thought vector as a starting point to generate a natural language response in word-by-word fashion almost like it is being typed out, he said. “Amazingly, the detailed operation of each network is entirely learned, just by training the model to predict likely responses,” Corrado wrote.
In developing the Smart Reply capability, Google had to ensure it met with email privacy requirements, he said. As a result, the company had to find a way to ensure that its machine learning would consistently work on data that humans themselves could not read. “[It] is a little like trying to solve a puzzle while blindfolded—but a challenge makes it more interesting!” Corrado noted.
Miklós described Smart Reply as an extension of the work Google has been doing to make Inbox more intuitive and user-friendly. As examples, he pointed to features such as Reminders and Trip Bundles for organizing travel-related information in one place.