Gmail Smart Compose: Real-Time Assisted Writing

By M. Chen, Benjamin Lee, G. Bansal, Yuan Cao, Shuyuan Zhang, Justin Lu, Jackie Tsay, Yinan Wang, Andrew M. Dai, Z. Chen and 2 more. 2019

248 citations recorded by the source corpus.

In this paper, we present Smart Compose, a novel system for generating interactive, real-time suggestions in Gmail that assists users in writing mails by reducing repetitive typing. In the design and deployment of such a large-scale and complicated system, we faced several challenges including model selection, performance evaluation, serving and other practical issues. At the core of Smart Compose is a large-scale neural language model. We leveraged state-of-the-art machine learning techniques for language model training which enabled high-quality suggestion prediction, and constructed novel serving infrastructure for high-throughput and real-time inference. Experimental results show the effectiveness of our proposed system design and deployment approach. This system is currently being served in Gmail.

Subfields: Artificial Intelligence, 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 Natural Language Processing

Citations in the corpus

RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models (2020, methodology): “To systematically evaluate and compare the generations from language models, we create REALTOXICITYPROMPTS as a testbed for toxicity in conditional language generation that mirrors real world applications (e.g., autocomplete systems; Chen et al., 2019; King, 2019).”
Extracting Training Data from Large Language Models (2020, methodology): “Our attacks directly apply to any language model, including those trained on sensitive and non-public data [8, 14].”
Red Teaming Language Models with Language Models (2022, background): “Data leakage compromises user privacy when the LM (e.g., GMail autocomplete; Chen et al., 2019) learns from confidential data (e.g., emails with Social Security Numbers; Carlini et al., 2019; Henderson et al., 2018).”
The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks (2018, methodology): “We are grateful to Martín Abadi, Ian Goodfellow, Ilya Mironov, Ananth Raghunathan, Kunal Talwar, and David Wagner for helpful discussion and to Gagan Bansal and the Gmail Smart Compose team for their expertise.”
Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LMs (2022, background): “…Jaques, 2019), e.g., voice assistants, and tutoring systems; or where AI helps enhance communication between multiple users, e.g., email autocom-plete (Chen et al., 2019), AI-assisted counseling (Kearns et al., 2020; Allen, 2020; Sharma et al., 2021), or facilitated discussion (Rosé et al., 2014).”
Differentially Private Learning Needs Better Features (or Much More Data) (2020, methodology): “Machine learning (ML) models have been successfully applied to the analysis of sensitive user data such as medical images (Lundervold & Lundervold, 2019), text messages (Chen et al., 2019) or social media posts (Wu et al., 2016).”
Machine Unlearning (2019, background): “This data is often sensitive in nature and could include information like medical records [1] or personal emails [2].”
IntelliCode compose: code generation using transformer (2020, methodology): “IntelliCode Compose serves as a universal programming language modeling tool, effectively generating syntactically correct code in multiple programming languages, capable of completing an entire line of code in a couple of key strokes, with a user experience inspired by Gmail Smart Compose [15].”
What Does it Mean for a Language Model to Preserve Privacy? (2022, neutral): “For example, LMs are being used in call centers, medical applications, mobile phones and personal computers and home assistants (such as Apple Siri, Amazon Alexa, Google Assistant, Microsoft Cortana, etc), email and message auto-complete services, document translation and search, writing companions ”
A Study of Face Obfuscation in ImageNet (2021, positive): “The system is implemented in MMDetection (Chen et al., 2019a).”
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).”
Training and Evaluating a Jupyter Notebook Data Science Assistant (2022, background): “These agents can be as simple as next-word or phrase prediction like in GMail Smart Compose (Chen et al. 2019) or as complex as conversational agents based on language models like GPT-3 (Brown et al.”
Automatic Document Sketching: Generating Drafts from Analogous Texts (2021, background): “Most end-user applications for assisting user writing, however, are conﬁned to sentence-level generation (Chen et al., 2019; Kannan et al., 2016; Alikaniotis and Raheja, 2019; Prabhumoye et al., 2019; Faltings et al., 2021).”
FRUIT: Faithfully Reflecting Updated Information in Text (2021, background): “However, 041 existing work has mainly focused on correcting 042 grammar (Wang et al., 2020), reducing repetitive 043 typing (Chen et al., 2019), and following rhetori- 044 cal directives (Sun et al., 2021), whereas the prob- 045 lem of producing edits grounded in external knowl- 046 edge has…”
Wiki-40B: Multilingual Language Model Dataset (2020, background): “They have been deployed in a wide variety of practical applications such as machine translation, automatic speech recognition, spelling correction and writing assistants (Kannan et al., 2016; Henderson et al., 2017; Chen et al., 2019).”