Google has introduced a ‘Smart Reply’ feature in Gmail that auto predicts responses by analysing the context in emails. The system is built on a pair of recurrent neural networks, one used to encode incoming emails, and the other to predict possible responses.
The feature will be made available this week on the Inbox app on both Google Play and the App Store in English. It will provide three responses to each email, of which the user can select one and forward it as it or by modifying it. The response examples given by Google are pretty neat, however we will see how well it works for real world usage when it launches this week. Examples included replies like “I don’t, sorry”, “I’ll send it to you”, and “I will have to look for it”, for an email that asked for the documentation of a new software.
How it works: Inbox uses machine learning to recognize emails that need responses and to generate the natural language responses on the fly. The responses users choose (or don’t) help improve future suggestions. It does not use ‘rules’ to determine what to reply, rather uses a recurrent neural network to consume each word in an email and create a vector (list of numbers). This vector fed through the second neural network which synthesizes a reply one word at a time.
Privacy issues: Google mentions that the network’s operations are entirely learned, and that it was trained to predict possible responses. The system is entirely automated and no humans get to read the mails. It works similarly to how Google displays ads within Gmail (it uses keywords in the email to determine user interest) and features like suggesting reminders work.
MediaNama’s Take: Artificial neural networks have been getting more elaborate and powerful. From Facebook’s algorithms that influence news feed results based on our own actions on the platform, to IBM’s Watson which went from beating humans at Jeopardy to an open-for-all-fields data analytics machine, we will only see such networks get better. Google itself started with simply providing ads based on words used in the email to now understanding the context behind these emails and automatically formulating replies.
However, as cool as this feature is, the thought that a bot that can understand the context of private emails is scary. Granted this network only composes replies, but with some tweaking it could say easily identify ideological affiliations etc. In such cases, privacy being assured by automation rings a little hollow. On a lighter note, maybe this will happen (Source: Randall Munroe xkcd):