Smart Reply - Overview of the paper

When you use Google Allo or GMail on your mobile device, you have to see suggestion which the app provides when you receive a short message.

In, the cover paper - Smart Reply: Automated Response Suggestion for Email, we can observe several interesting ideas. I will write some of then in a short list below. If you want to read it all by yourself, I will strongly encourage you to do it, but at the end of this post is also a highligted version.

• The system is already in the production with responsibily for more than 10% replies on mobile phone. (with keeping in the mind that the trigger mechanism keep a majority of messages without a suggestion)
• About 25% of replies has even or less than 20 tokens (words)
• LSTM can be used generate unique reply to every input message, BUT then there will be problem with
  • Response quality
  • Likehood that reply will be chosen
  • Scalability
  • Privacy - can show names or other stuff because they will be in the vocabulary
• LSTM is used in smarter way and just select responses from pregenerated sets (good grammar)
• To make a base set of resposnes is used semi-supervised algorithm (+ Expander)
• All response messages should be part of the “semantic” cluster - from these clusters are chosen response
• What about redundant/duplicite responses - delete them
• What if all sugested responses are positive/negative - delete majority and enforce minority
• Dealing with text = preprocces it -> tokenize sentences -> tokenize words -> remove names/number etc.
• Dealing with email = preprocces it -> a quoation/signature/salutation removal

A PDF document with highlighted interesting parts is here.

Written on September 3, 2017