Smart Reply was announced on the Gmail blog today. It will be added to Google Inbox later this week, an email app specifically designed to make managing, sorting and organising mail a simpler process than it traditionally is.
The update will let Inbox use machine learning to analyse each email. It will recognise when a message requires a response, detect the subject and tone and generate a few different natural language responses all on its own. Its prepared messages will then be displayed to the user who will be able to send the finished response with just one tap.
Google’s AI “Smart Reply” system in Inbox
Google
Another Google blog post by its Research team goes into more detail about how the process works. Greg Corrado, a senior research scientist at Google, explains how Smart Reply is built on top of a neural network that looks at each word of an email in turn and predicts what a response could look like.
Corrado says the team is “actually pretty amazed at how well this works” but there were a few “unexpected quirks” during development. The first prototype system often offered three responses that were essentially the same message but with different phrasing, such as “How about tomorrow?”, “Wanna get together tomorrow?” and “I suggest we meet tomorrow.”
Another team in Google, led by Sujith Ravi, solved the problem by developing another artificial intelligence that can map natural language responses to semantic intents, allowing Smart Reply to work out how semantically similar the messages it is generating are and create new ones accordingly.
The first version of the app also had another problem: it kept sending “I love you” to the various Googlers testing it. It emerged this wasn’t actually a bug but a result of how the app was trained. It turns out “I love you” is actually a very common phrase so Inbox began to lean towards it when it needed a sure-fire way to quickly express positive emotion. The team revised the algorithms so the responses generated are always relevant to the original email.
Google’s AI “Smart Reply” system in Inbox
Google
Smart Reply’s predictions get better the more that the app is used. It can learn from previous responses to work out what the best reply will be, building on an individual user’s idiolect to tune phrases and make them seem more natural.
Inbox already takes advantage of machine learning capabilities in several ways. It can help to organise trips directly from the calendar, intelligently suggest tasks to create and automatically provide tips on how to use the app when it detects a user is confused or unsure of what to do next.
The systems in Inbox aren’t the only artificial intelligence project under development at Google. The company is committed to using machine learning to enhance every one of its products and recently revealed details of its AI “RankBrain” system that helps to decide what results are displayed for the millions of search requests made to Google each day.
