Smart Reply: Automated Response Suggestion for Email Authors - - PowerPoint PPT Presentation

Smart Reply: Automated Response Suggestion for Email

Authors Anjuli Kannan* Karol Kurach* Sujith Ravi* Tobias Kaufmann* Andrew Tomkins Balint Miklos Greg Corrado László Lukács Marina Ganea Peter Young Vivek Ramavajjala

* Google Employees

Presenter Andrew Janick

Why Smart Reply?

• Email continues to be a primary medium for

communication for billions of users across the world

• It is increasingly challenging for users to process and

respond to incoming messages especially on mobile devices

• Roughly 25% of email replies have 20 or less tokens (words

and punctuation) The Authors’ Question: Is it possible to assist users in composing short replies to email messages when appropriate that they can use in one tap on a mobile device?

2

What Is Smart Reply?

• Abstract
• Generates semantically diverse suggestions that can be

used as complete email responses with just one tap on mobile

• Currently used in Inbox by Gmail
• Responsible for assisting with 10% of all mobile responses
• Designed to work at very high throughput and process

hundreds of millions of messages daily

• System exploits state-of-the-art, large-scale deep learning

3

The Smart Reply Flow

4

The Smart Reply Flow: Preprocess Email

5

• Language detection - The language of the message is

identified and non-English messages are discarded.

• Tokenization - Subject and message body are broken into

words and punctuation marks.

• Sentence segmentation - Sentences boundaries are

identified in the message body.

• Normalization - Infrequent words and entities like personal

names, URLs, email addresses, phone numbers etc. are replaced by special tokens.

• Quotation removal - Quoted original messages and

forwarded messages are removed.

• Salutation/close removal - Salutations like Hi John and

closes such as Best regards, Mary are removed.

The Smart Reply Flow: Trigger Response?

6

• Two requirements:
• 1. Needs to be fast – 100s of millions of message means decision

needs to happen in milliseconds

• 2. Figure out cases where a response is not expected to save

resources on the calls to the LSTM inference and to not provide useless responses

• Uses a feedforward network - the first and simplest type of artificial

neural network devised. In this network, the information moves in only

• ne direction, forward, from the input nodes, through the hidden

nodes (if any) and to the output nodes. There are no cycles or loops in the network.

• Probability score – the feedforward network produces a probability

score that is compared to a threshold. Above threshold = Yes/Below threshold = No

• Trigger response says Yes – only 11% of emails get Smart Reply

suggestions

The Smart Reply Flow - LSTM

7

• Long short-term memory (LSTM)

1. A recurrent neural network (RNN) architecture (an artificial neural network) 2. Has loops and can reprocess errors and data to form cause and effect relationships 3. Has persistent data so is learning as time passes

• Response tokens to the original message get scores – the

most likely response tokens are fed back in and rescored until the best responses are found.

• Response set is generated – a few million unique sentences

The Smart Reply Flow - LSTM

8

• Challenges:

1. need quality responses (no bad grammar, spelling, or mechanics) 2. Need responses with a large amount of utility – specific to original message 3. Model needs to scale due to number of messages

• Responses are clustered – broken down into clusters based
• n intent. Thank you versus sorry versus lol – all seeds in

different clusters.

• 100 clusters and on average 3 to 5 labeled seed

examples per cluster.

The Smart Reply Flow – Diversity selection

9

• Remove redundant responses – Iterate over the clusters until

the resulting list contains only the highest-scored representation of each intent in order

• Enforce negative and positive responses – LSTM has a strong

tendency to produce positive responses (probably due to the fact that most email conversations are positive). A second LSTM pass is performed limited to negative responses.

The Smart Reply Flow – Can you join tomorrow’s meeting?

10

Conclusion

Smart reply – an end-to-end system that generates short, complete email response using a deep LSTM model that accurately predicts valid responses. Challenges met:

• 1. Responses are of high quality
• 2. Enforcing diversity has given the responses high

utility

• 3. Efficient searching has made the solution scalable

Best metric of success: 10% of mobile replies in Inbox are now composed with assistance from the Smart Reply system

11