Smart Reply: Automated Response Suggestion for Email

By Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg S. Corrado, László Lukács, Marina Ganea, Peter Young and 1 more. 2016

338 citations recorded by the source corpus.

In this paper we propose and investigate a novel end-to-end method for automatically generating short email responses, called Smart Reply. It generates semantically diverse suggestions that can be used as complete email responses with just one tap on mobile. The system is currently used in Inbox by Gmail and is responsible for assisting with 10% of all mobile responses. It is designed to work at very high throughput and process hundreds of millions of messages daily. The system exploits state-of-the-art, large-scale deep learning. We describe the architecture of the system as well as the challenges that we faced while building it, like response diversity and scalability. We also introduce a new method for semantic clustering of user-generated content that requires only a modest amount of explicitly labeled data.

Subfields: Artificial Intelligence, Information Systems, Information Systems and Management, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Information Systems

In Information Systems and Management

In Natural Language Processing

Citations in the corpus

Efficient Natural Language Response Suggestion for Smart Reply (2017, methodology): “…successive iterations of the Smart Reply system, showing the conversion rate and latency relative to the baseline Seq2Seq system of Kannan et al. [11]. the suggestions, see row 3 of table 3, moving the systems toward shorter and more generic responses that users were more likely to ﬁnd…”
Visual Dialog (2016, background): “While some of the earliest developed chatbots were rule-based [64], end-to-end learning based approaches are now being actively explored [9,14,26,31,53,54,61].”
A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues (2016, methodology): “Two examples are Microsoft’s Xiaolice (Markoff and Mozur 2015) and Google’s Smart Reply system (Kannan et al. 2016), which at its core uses a sequence-to-sequence model.”
Towards an Automatic Turing Test: Learning to Evaluate Dialogue Responses (2017, background): “Such models have already had a substantial impact in industry, including Google’s Smart Reply system (Kannan et al., 2016), and Microsoft’s Xiaoice chatbot (Markoff and Mozur, 2015), which has over 20 million users.”
Generative Deep Neural Networks for Dialogue: A Short Review (2016, methodology): “…goal-oriented setting [Gorin et al., 1997, Young, 2000, Singh et al., 2002], these models have already been applied to several real-world applications, with Microsoft’s system Xiaoice [Markoff and Mozur, 2015] and Google’s Smart Reply system [Kannan et al., 2016] as two prominent examples.”
Learning Semantic Textual Similarity from Conversations (2018, methodology): “Training on conversational data has been previously shown to be effective at email response prediction (Kannan et al., 2016; Henderson et al., 2017).”
Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data (2016, background): “…learning applications with great beneﬁts are enabled only through the analysis of sensitive data, such as users’ personal contacts, private photographs or correspondence, or even medical records or genetic sequences (Alipanahi et al., 2015; Kannan et al., 2016; Kononenko, 2001; Sweeney, 1997).”
Adversarial Evaluation of Dialogue Models (2017, methodology): “The generator is trained on the same data and in the same manner as the production-scale model that is deployed as part of the Smart Reply feature in Inbox by Gmail [3].”
Gmail Smart Compose: Real-Time Assisted Writing (2019, methodology): “We pre-process the data similar to that described in [28], namely”
Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models (2016, neutral): “always saying “I don’t know” in conversation models (Kannan et al., 2016).”
Key-Value Retrieval Networks for Task-Oriented Dialogue (2017, background): “Unlike other dialogue agents which only rely on prior dialogue history for generation (Kannan et al., 2016; Eric and Manning, 2017), our architecture is able to access and use database-style”
Hello, It’s GPT-2 - How Can I Help You? Towards the Use of Pretrained Language Models for Task-Oriented Dialogue Systems (2019, background): “Here, the massive datasets can be leveraged to aid task-speciﬁc applications (Kannan et al., 2016; Henderson et al., 2017, 2019b).”
Information Leakage in Embedding Models (2020, background): “learning from embeddings have been shown to be tremendously useful for many natural language processing (NLP) tasks such as paraphrasing [14], response suggestion [30], information retrieval,”
Generating More Interesting Responses in Neural Conversation Models with Distributional Constraints (2018, methodology): “As a starting point for our approach we leverage the Seq2Seq model (Sutskever et al., 2014; Bah-danau et al., 2014) which has been used as a basis for a broad range of recent work on neural conversation (Kannan et al., 2016; Li et al., 2016a; Serban et al., 2016; Shao et al., 2017).”
When Are Search Completion Suggestions Problematic? (2020, background): “Given that query suggestions are typically generated from behavioral patterns observed in past search logs [35, 47, 77], we also overview work that provides cues about how past logs can make suggestion engines more prone to surfacing problematic suggestions, and that can help us catalogue such cases”
Membership Inference Attacks on Machine Learning: A Survey (2021, methodology): “Embeddings have been successfully applied to various domains including natural language processing [90], social networks [55], movie feedback [63], and location [33].”
A Repository of Conversational Datasets (2019, neutral): “For instance, the datasets are suitable for training generative models of conversational response (Serban et al., 2016; Ritter et al., 2011; Vinyals and Le, 2015; Sordoni et al., 2015; Shang et al., 2015; Kannan et al., 2016), as well as discriminative methods of conversational response selection (L”
Scaling Graph Neural Networks with Approximate PageRank (2020, background): “Their applications occur across all media types and power many different Google products [29, 39, 40].”
Multiresolution Recurrent Neural Networks: An Application to Dialogue Response Generation (2016, background): “Another example is Google’s Smart Reply system (Kannan et al. 2016).”
Differentially Private Set Union (2020, background): “Natural language models for applications such as suggested replies for e-mails and dialog systems rely on the discovery of n -grams and sentences Hu et al. [2014], Kannan et al. [2016], Chen et al. [2019], Deb et al. [2019].”