Have you ever wanted to automatically adjust your tone, select your preferred punctuation, sentence length and complexity of replies? In 2009, Google pulled a prank on the world and introduced "Gmail Autopilot" a feature that promised to analyse messages received within Gmail and generate contextually relevant replies. In November 2015, Google announced a similar feature named Google Smart Reply, and this time, it's definitely not a joke.
Over six years on, Smart Reply shares some of the same key features as Autopilot, promising to save the user time and thought in responding to their email messages. Smart Reply guesses which emails can be answered with a short reply and prepares a few responses for the user to reply with, based on sophisticated machine-learning and recurrent neural networks. So what does it mean for email marketing?
Think outside the box
With regards to email, machine learning has previously been applied to spam detection and email categorisation. Google researchers proposed something called sequence-to-sequence learning which is a technique for understanding and generating natural language. Like other sequence-to-sequence models, Smart Reply is built on a pair of recurrent neural networks (RNNs). One network encodes incoming email and the other predicts possible responses. Each incoming email is parsed using an encoding network, producing a thought vector that captures the gist of what is being said regardless of diction.
For example, the thought vectors for "Are you free for lunch?" and "Want to grab lunch?" would look similar. Since the inputs and outputs of email can be hundreds of words long, Google chose a specific type of RNN known as a long short-term-memory network, or LSTM. This RNN can focus on the part of the incoming email that is most useful in predicting a response, without getting sidetracked by less relevant sentences.
One problem researchers encountered when developing Smart Reply was a lack of diversity among the response options, so the team developed a machine learning system for mapping natural language responses to semantic intents. This system enabled Smart Reply to pick responses that were not only different in wording, but also in underlying meaning.
Another quirk in the original prototype was the tendency for it to reply with phrases like "I love you", "Thanks", and "Sounds good." Since Smart Reply learns from past responses, it would rely on one of them when in doubt. To correct for this, engineers had to adjust the tool to predict responses that were not just likely, but also had high affinity to the original message. Nothing’s more embarrassing than an “I love you” email to the boss.
Smart Reply or Quick Response?
Smart Reply is part of Google Inbox on both Android and iOS platforms, but the responses from Smart Reply are not available for all emails; they will only be offered to emails for which a short reply has been deemed possible. In testing Google Smart Reply, Google researchers discovered that users were reluctant to have a machine respond on emotional or intimate matters. They also found that the longer the reply, the more the user felt the reply should be in their own tone, as after about ten words, individual tone becomes noticeable.
This is why Smart Reply only offers responses of five to six words, even though it's perfectly capable of generating sentences of around a dozen words. Google Smart Reply is incredibly complex, and Google’s product management director Alex Gawley suggests the system can generate around 20,000 discrete responses that even picks up on subtle nuances, including jokes!
What’s next for Smart Reply?
Google is working on making the feature available in other languages, as well as another forthcoming improvement which has been dubbed the "exclamation-point model.” With this, Smart Reply will mimic the user's propensity for exclamation points and sprinkle them throughout replies accordingly. How exciting!
Learn more about email automation from Pure360 here.
Google Smart Reply isn’t the only tool to help users manage their inbox, but it’s set to be one of the best. In many ways, apps and technology are becoming more data-driven and predictive, evolving into our own virtual personal assistants. Smart Reply is another step in this direction and could have the biggest impact yet on inbox management. With the time and effort saved by Smart Reply, users may have more time to spend elsewhere in their inbox, perhaps browsing through promotional email messages and engaging with them.