Google researchers announced the launch of Smart Reply in November 2015. Smart Reply is a new feature for Gmail that suggests three responses based on emails you get.
What’s the benefit? Well, as TechCrunch points out, one benefit, perhaps the biggest benefit of all, is that it reduces the need to key things on your smartphone, which is a great improvement to trying to key a response of several words with your thumbs when you are standing on a crowded train trying to respond to emails.
How does it work? It is not rule-based. Instead, it uses deep neural networks, plus something called sequence-to-sequence learning, which until now has been used mainly for machine translation. Essentially, it is determining predictive responses to triggers in the incoming email. One key problem is knowing what to ignore in the incoming email, because it is not relevant to the response.
The system “learns” in the sense that if you never choose one response that is offered, it no longer offers that response as one of your three choices.
One benefit of the new service is that it respects the importance of privacy – as the Google researchers point out, we can’t have humans reading your email, so you will be reassured that this is done entirely automatically. The researchers describe it as “trying to solve a puzzle while blindfolded” – but this is true of algorithm-based machine learning generally. The surprising thing is that the system works so well, at least for emails with a simple message.
Limitations? Well, it’s not difficult to think of responses that provide an appropriate response for 90% of all emails; the single word “Great!” would suffice for many. In fact, the researchers state that early versions of the system tended to offer responses such as “Thanks” and “sounds good”, so much so that the Google engineers admit that the system “would lean on them as a safe bet if it was unsure”. More alarming, apparently the system had a tendency to often produce a response “I love you”, which might cause embarrassment as a response to many emails. Perhaps it’s best that they dropped that suggestion.