Building a Production Model for Retrieval-Based Chatbots

Swanson, Kyle; Yu, Lili; Fox, Christopher; Wohlwend, Jeremy; Leí, Tao

doi:10.18653/v1/w19-4104

Cited by 6 publications

(6 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One such technique is called selfattention, which is essentially a mechanism that allows us to examine which features of the data a neural network is paying "attention" to most. 57,62 Mathematically, self-attention is akin to a dot product that yields a set of weights for each feature, which are then interpreted as "attention" scores.…”

Section: Discussionmentioning

confidence: 99%

Deep learning for automated classification and characterization of amorphous materials

et al. 2020

Self Cite

View full text Add to dashboard Cite

It is difficult to quantify structure-property relationships and to identify structural features of complex materials. The characterization of amorphous materials is especially challenging because their lack of long-range order makes it difficult to define structural metrics. In this work, we apply deep learning algorithms to accurately classify amorphous materials and characterize their structural features. Specifically, we show that convolutional neural networks and message passing neural networks can classify two-dimensional liquids and liquid-cooled glasses from molecular dynamics simulations with greater than 0.98 AUC, with no a priori assumptions about local particle relationships, even when the liquids and glasses are prepared at the same inherent structure energy. Furthermore, we demonstrate that message passing neural networks surpass convolutional neural networks in this context in both accuracy and interpretability. We extract a clear interpretation of how message passing neural networks evaluate liquid and glass structures by using a self-attention mechanism. Using this interpretation, we derive three novel structural metrics that accurately characterize glass formation. The methods presented here provide us with a procedure to identify important structural features in materials that could be missed by standard techniques and give us a unique insight into how these neural networks process data. arXiv:1909.04648v1 [cond-mat.soft]

show abstract

Section: Discussionmentioning

confidence: 99%

Deep learning for automated classification and characterization of amorphous materials

et al. 2020

Self Cite

View full text Add to dashboard Cite

show abstract

“…While RS has been modeled as a sequence to sequence model (Kannan et al, 2016), it more commonly appears as an information retrieval (IR) system by ranking responses from a fixed set (Henderson et al, 2017(Henderson et al, , 2019Ying et al, 2021;Swanson et al, 2019;Zhou et al, 2016Zhou et al, , 2018 due to better control over quality and relevance for practical systems. We briefly describe two retrieval architectures from prior literature which serves as the baselines for our multilingual RS model.…”

Section: Background and Preliminariesmentioning

confidence: 99%

A Conditional Generative Matching Model for Multi-lingual Reply Suggestion

Deb¹,

Zheng²,

Shokouhi³

et al. 2021

Findings of the Association for Computational Linguistics: EMNLP 2021

View full text Add to dashboard Cite

We study the problem of multilingual automated reply suggestions (RS) model serving many languages simultaneously. Multilingual models are often challenged by model capacity and severe data distribution skew across languages. While prior works largely focus on monolingual models, we propose Conditional Generative Matching models (CGM), optimized within a Variational Autoencoder framework to address challenges arising from multilingual RS. CGM does so with expressive message conditional priors, mixture densities to enhance multi-lingual data representation, latent alignment for language discrimination, and effective variational optimization techniques for training multi-lingual RS. The enhancements result in performance that exceed competitive baselines in relevance (ROUGE score) by more than 10% on average, and 16% for low resource languages. CGM also shows remarkable improvements in diversity (80%) illustrating its expressiveness in representation of multi-lingual data.

show abstract

“…Chatbots or social bots have gone beyond chit-chat, can be further categorized as generative methods and retrieval-based methods. These methods are applied to goal-oriented dialogues as well, aiming to directly select or generate a dialogue response given an input (Gandhe and Traum, 2010;Swanson et al, 2019;Henderson et al, 2019).…”

Section: Related Workmentioning

confidence: 99%

“…For evaluation, goal-oriented dialogue systems can be evaluated to measure task-success and dialogue efficiency . Retrieval-based chatbots often report performance on Next Utterance Classification, to test if a next utterance can be correctly selected given the chat context Henderson et al, 2019;Swanson et al, 2019). Conversational QA systems, on the other hand, are evaluated based on the correctness of their answers and the naturalness of the conversations (Reddy et al, 2019;.…”

Section: Related Workmentioning

confidence: 99%

Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue

2021

View full text Add to dashboard Cite

We are glad to pen the first few words for the proceedings of SIGDIAL 2021, the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue. The SIGDIAL conference is a premier publication venue for research in discourse and dialogue. This year, the conference is organized as a hybrid event with both in-person and virtual participation on July 29-31, 2021, right before ACL-IJCNLP 2021. The 2021 Young Researchers' Roundtable on Spoken Dialog Systems (YRRSDS 2021) is also held as a satellite event. The SIGDIAL 2021 program features three keynote talks, 6 paper presentation sessions, 1 demo session, and 2 special sessions, entitled "Summarization of Dialogues and Multi-Party Meetings", and "Safety for E2E Conversational AI".COVID has changed the way we work, but it doesn't hamper our research progress. We received 142 submissions this year, comprising 88 long papers, 49 short papers, and 5 demo descriptions. We had 12 Senior Program Committee (SPC) members who were each responsible for 11-12 papers, leading the discussion process and also contributing meta-reviews. Each submission was assigned to an SPC member and received at least three reviews. Decisions carefully considered the original reviews, meta-reviews, and discussions among reviewers facilitated by the SPCs. We are immensely grateful to the members of the Program Committee and Senior Program Committee for efforts in providing excellent, thoughtful reviews of the large number of submissions. Their contributions have been essential to selecting the accepted papers and providing a high-quality technical program for the conference. We have aimed to develop a broad, varied program spanning the many positively rated papers identified by the review process. We accepted 59 papers in total: 40 long papers (45%), 15 short papers (31%), and 4 demo descriptions, for an overall acceptance rate of 41.5%, in line with prior years. One keynote will highlight each of the three days of the conference. In organizing this hybrid inperson/virtual conference, we have tried to maintain as much of the spirit of a fully online conference as possible. Recordings for all papers and demos have been made available several days before the start of the conference, for participants to watch asynchronously. Long and short papers are organized into sessions taking into consideration the presenters' time zones. Regular papers sessions span 8-11 papers, each presented as a two-minute pre-recorded talk followed by five minutes of live Q&A. For demos, we organized four parallel zoom rooms to allow participants to interact with and observe live interactions with the systems. The topics represent the breadth of research in discourse and dialogue. A conference of this size requires the energy, guidance, and contributions of many parties, and we would like to take this opportunity to thank and acknowledge them all. We thank our three keynote speakers, Julia Hirschberg (Columbia University), Raymond J. Mooney (University of Texas at Austin), and Jason Weston (Facebook AI & NYU), for...

show abstract

Building a Production Model for Retrieval-Based Chatbots

Cited by 6 publications

References 23 publications

Deep learning for automated classification and characterization of amorphous materials

Deep learning for automated classification and characterization of amorphous materials

A Conditional Generative Matching Model for Multi-lingual Reply Suggestion

Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue

Contact Info

Product

Resources

About