According to Google software engineer Bálint Miklós, Smart Reply is a two-part ordeal: It scans the content of the e-mail and, based on what it learns, generates three possible responses. It’s a machine learning workflow Miklós calls “sequence-to-sequence learning” — one flexible enough to understand the meaning or intent of an e-mail without getting stuck on unfamiliar words but powerful enough to spit out replies that make grammatical and contextual sense. The end result is a system which works far better than “brittle, rule-based systems ever could,” said Miklós.
But the first iteration of Smart Reply wasn’t perfect. It had a tendency to suggest replies that, while worded differently, were identical in meaning. And more amusingly, it frequently offered “I love you” as a possible response (the reason, Miklós said, is because “I love you” is an unsurprisingly common answer). A few tweaks to the system’s understanding of semantics later, though, and Smart Reply was “a less lovely, but far more useful, email assistant,” Miklós wrote.
Useful enough for prime time, at least. Miklós said that Smart Reply will roll out on Inbox for Android and iOS later this week.
Smart Reply isn’t Google’s first application of machine learning to e-mail’s more onerous tasks — in July, it rolled out an “artificial neural network” for the purpose of better identifying spam and phishing scams. Google’s artificially intelligent email may be a far cry from an inbox that clears itself, but here’s hoping that’s a dream realized — within reason, of course — sooner rather than later.
