Efficient Natural Language Response Suggestion for Smart Reply

By Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, R. Kurzweil. 2017

514 citations recorded by the source corpus.

This paper presents a computationally efficient machine-learned method for natural language response suggestion. Feed-forward neural networks using n-gram embedding features encode messages into vectors which are optimized to give message-response pairs a high dot-product value. An optimized search finds response suggestions. The method is evaluated in a large-scale commercial e-mail application, Inbox by Gmail. Compared to a sequence-to-sequence approach, the new system achieves the same quality at a small fraction of the computational requirements and latency.

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

Dense Passage Retrieval for Open-Domain Question Answering (2020, background): “…pairs to effectively generate a large number of examples, such as the in-batch random negative pairs (Yih et al., 2011; Henderson et al., 2017), or when there exist a big collection of positive and negative pairs (Huang et al., 2013), dense representation can outperform sparse representation.”
Learning Dense Representations for Entity Retrieval (2019, methodology): “These pairs constitute only positive examples, so we use in-batch random negatives (Henderson et al., 2017; Gillick et al., 2018): for each mention-entity pair in a training batch, the other entities in the batch are used as negatives.”
Learning Dense Representations of Phrases at Scale (2020, methodology): “We then adopt negative sampling strategies such as in-batch negatives (Henderson et al., 2017; Karpukhin et al., 2020), to better discriminate the phrases at a larger scale.”
Pretrained Transformers for Text Ranking: BERT and Beyond (2020, neutral): “The jury’s still out.”
CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP (2021, methodology): “In analogy to these human cognitive processes, we suggest to use modern Hopﬁeld networks to amplify co-occurrences and covariance structures of the original data.”
Universal Sentence Encoder (2018, methodology): “, 2015) for the unsupervised learning from arbitrary running text; a conversational input-response task for the inclusion of parsed conversational data (Henderson et al., 2017); and classification tasks for training on supervised data.”
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).”
Universal Sentence Encoder for English (2018, methodology): “See Yang et al. (2018) and Henderson et al. (2017) for additional architectural details of models similar to those presented here.”
SimCSE: Simple Contrastive Learning of Sentence Embeddings (2021, methodology): “In NLP, a similar contrastive learning objective has been explored in different contexts (Henderson et al., 2017; Gillick et al., 2019; Karpukhin et al., 2020).”
Pre-training Tasks for Embedding-based Large-scale Retrieval (2020, methodology): “The model is learned with multiple tasks including the unsupervised Skip-Thought task (Kiros et al., 2015), the supervised conversation input-response task (Henderson et al., 2017), and the supervised sentence classiﬁcation SNLI task (Bowman et al., 2015).”
Learning To Retrieve Prompts for In-Context Learning (2021, background): “An advantage of this approach is that for batch size B the effective batch size is of order B 2 , with the in-batch negatives trick (Henderson et al., 2017).”
End-to-End Retrieval in Continuous Space (2018, neutral): “This was suggested by Henderson et al. (2017).”
Multi-Stage Document Ranking with BERT (2019, background): “of a simple metric such as cosine similarity (Henderson et al., 2017; Zamani et al., 2018; Ji et al., 2019).”
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).”
Explanations for CommonsenseQA: New Dataset and Models (2021, methodology): “We use multiple negatives ranking (MNR) (Henderson et al., 2017) as the loss, which is negative log-softmax over similarity of zqac and zp∗ .”
RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering (2020, methodology): “Several recent studies (Karpukhin et al., 2020; Luan et al., 2020; Chang et al., 2020; Henderson et al., 2017) tried to address the ﬁrst challenge by designing complicated sampling mechanism to generate hard negatives.”
Investigating Societal Biases in a Poetry Composition System (2020, methodology): “Instead, we follow Henderson et al. (2017) and sample K responses to estimate P (y|x) as Papprox(y|x) = P (x, y)∑K k=1 P (x, yk) .”
Effective Parallel Corpus Mining using Bilingual Sentence Embeddings (2018, methodology): “Not only is the dual-encoder architecture is more efﬁcient (Henderson et al., 2017), it also allows us to directly train toward extracting parallel sentences from a collection of candidates.”