Google's Smart Reply : a summary of NLP techniques and insights used

One of the more recent additions to Google's plethora of services offered is Smart Reply.As of now, Smart Reply humbly seeks to provide short crisp responses for replying to emails on the go.Interestingly enough, a lot has gone into developing the system to a deployable ,efficient package, which has been found useful by a large fraction of the user base.

Google's smart reply system can be understood as an amalgam of 4 key components: 

1.Response Selection

2.Response Set Generation

3.Suggestion Diversity 

4.Triggering Model 

Response Selection:

Google uses a neural network with LSTM (Long Short Term Memory cells) to generate responses.

The email corpus used for training the network was extracted from Google's own mail database, after anonymization.It consists of around 238 million messages, which include 153 million messages that have no response.

Given an email e ,and set of all possible responses ,the score of a response r is defined as P(e|r)[P=Probability].

Top k responses are then taken for further processing.

Response Set Generation:

To ensure better quality of responses,i.e reduce redundant responses such as "Thanks for the update.","Thank you for the update!","Thanks for the status update!" , Google uses something called semantic intent clustering.The sentences are parsed using a dependency parser and a canonicalized representation is created.Thereafter,all responses are assigned a semantic cluster , which is a broad category of what the intent of the message is.For example, "Haha","LOL!",etc. would be categorised as funny.

To achieve this task of semantic clustering,Google uses a semi-supervised learning algorithm using scalable graph algorithms ,which could learn automatically from the data and a few human-annotated samples.

Suggestion Diversity:

The ideal behind this is to provide no two responses with the same intent to the user.The more the variety in response intents, the more utility lies for the user.This is done by intent-checking and enforcing negative and positive variations of intent, by filtering the response space suitably.The filtering mechanism identifies the intent of the mail as affirmative, missing(indirect) negatives and exclusive negatives.And picks the best suited responses from each.

Triggering Model:

The triggering model suppresses the response generator by deciding whether a response should be given to the mail or not.This decision is taken with respect to factors like whether the mails are auto-generated or not, and whether short replies are appropriate for the same or not(since it could be a sensitive letter demanding more composition).

The model was built using a feed-forward neural network which produces a score for each mail based on probability.If the score is below a threshold, it doesn't trigger the response mechanism ,and hence no response is generated!


Conclusion:

The efforts have surely paid off.Google has reported around 10% of mobile replies use Smart Reply,which is surely a good sign.Also,the system is language-agnostic and hence can be extended to other native languages as well in the future.
From this we see Google attacked each of the sub-problems individually,which were primarily in the NLP domain, in a novel way , and combined everything together to create a market-ready product.Some interesting challenges that one can think of to further this paper are :
1.How to compose longer and at the same time , legitimate mails?
2.How to take references in the real world into context and use it in the response? (In other words, grasp a proper noun say Los Angeles, as a location and not a simple token , and possibly make a better fitted response.)