Neural Conversational AI Workshop
What’s left to TEACH (Trustworthy, Enhanced, Adaptable, Capable and Human-centric) chatbots?
Invited Speakers (Sorted in alphabetical order by last name)
Emily Dinan
Emily Dinan is a Research Engineer at DeepMind in New York. Her research interests include natural language generation and conversational AI, as well as safety and responsibility in these fields. Recently she has focused on methods for integrating generative language models with expert systems, and was a co-creator of Cicero, an agent that plays the game of Diplomacy at a human level. Prior to joining DeepMind, she worked as a Research Engineer at FAIR (Meta AI). She received her master's degree in Mathematics from the University of Washington.
Building a dialogue agent for Diplomacy
Last November we announced Cicero, the first AI agent capable of playing the board game Diplomacy at a human level. In this talk, we'll focus on the language-related aspects of this work, and in particular, how we built a dialogue agent capable of negotiating, coordinating, and strategizing with other humans through natural language in this complex seven-player game.
Pascale Fung
Pascale Fung is a Chair Professor at the Department of Electronic & Computer Engineering at The Hong Kong University of Science & Technology (HKUST), and a visiting professor at the Central Academy of Fine Arts in Beijing. She is an elected Fellow of the Association for the Advancement of Artificial Intelligence (AAAI) for her "significant contributions to the field of conversational AI and to the development of ethical AI principles and algorithms", an elected Fellow of the Association for Computational Linguistics (ACL) for her “significant contributions towards statistical NLP, comparable corpora, and building intelligent systems that can understand and empathize with humans”. She is an Fellow of the Institute of Electrical and Electronic Engineers (IEEE) for her “contributions to human-machine interactions” and an elected Fellow of the International Speech Communication Association for “fundamental contributions to the interdisciplinary area of spoken language human-machine interactions”. She is the Director of HKUST Centre for AI Research (CAiRE), an interdisciplinary research centre on top of all four schools at HKUST. She is an expert on the Global Future Council, a think tank for the World Economic Forum. She represents HKUST on Partnership on AI to Benefit People and Society. She is on the Board of Governors of the IEEE Signal Processing Society. She is a member of the IEEE Working Group to develop an IEEE standard - Recommended Practice for Organizational Governance of Artificial Intelligence. Her research team has won several best and outstanding paper awardsat ACL, ACL and NeurIPS workshops.
Safer Generative ConvAI
Generative models for Conversational AI are less than a decade old, but they hold great promise for human-machine interactions. Machine responses based on generative models can seem quite fluent and human-like, empathetic and funny, knowledgeable and professional. However, behind the confident voice of generative ConvAI systems, they can also be hallucinating misinformation, giving biased and harmful views, and are still not "safe" enough for many real life applications. The expressive power of generative ConvAI models and their undesirable behaviors are two sides of the same coin. How can we harness the fluency, diversity, engagingness of generative ConvAI models while mitigating the downside? In this talk, I will present some of our recent work in making generative ConvAI safer via mitigating hallucinations, misinformation, and toxicity.
Arvind R Neelakantan
Arvind Neelakantan is a Research Lead and Manager at OpenAI working on deep learning research for real-world applications. He got his PhD from UMass Amherst where he was also a Google PhD Fellow. His work has received best paper awards at NeurIPS and at Automated Knowledge Base Construction workshop.
Embeddings and Retrieval Augmented Generation
Embeddings are numerical representations of concepts converted to number sequences, which make it easy for computers to understand the relationships between those concepts. This talk will first focus on our work on embeddings that are useful to perform natural language and code tasks like semantic search, clustering, topic modeling, and classification. Then we will dive deeper into the application of embeddings for retrieval augmented generation with language models.
João Sedoc
João Sedoc is an Assistant Professor of Information Systems in the Department of Technology, Operations and Statistics at New York University Stern School of Business. He is also affiliated with the Center for Datascience ML^2 Lab at NYU. His research areas are at the intersection of machine learning and natural language processing. His interests include conversational agents, hierarchical models, deep learning, and time series analysis. Before joining NYU Stern, he worked as an Assistant Research Professor in the Department of Computer Science at Johns Hopkins University. He received his PhD in Computer and Information Science from the University of Pennsylvania.
New Frontiers in the Evaluation of Conversational Agents
The rapid advances in large language models brought about disruptive innovations in the field of conversational agents. However, recent advances also present new challenges in evaluating the quality of such systems, as well as the underlying models and methods. As conversational agents increasingly match and or even surpass human performance in dimensions like 'coherence,' we must shift our focus to the qualities of conversational agents that are fundamental to human-like conversation (e.g., empathy and emotion).
In this talk, I will focus on how we can integrate psychological metrics for evaluating conversational agents along dimensions such as emotion, empathy, and user traits. I will also introduce our Item Response Theory (IRT) framework, an innovative approach for evaluating the quality of agents across various dimensions. Finally, I will discuss future directions of conversational agent evaluation.
Marilyn Walker
Marilyn Walker is a Professor of Computer Science and Engineering at UC Santa Cruz and a fellow of the Association for Computational Linguistics (ACL), in recognition of her fundamental contributions to statistical methods for dialog optimization, to centering theory, and to expressive generation for dialog. Her current research includes work on computational models of dialogue interaction and conversational agents, evaluation methods for dialogue systems, analysis of affect, sarcasm and other social phenomena in social media dialogue, and methods for training the dialogue manager and the natural language generator for dialogue systems. Before coming to Santa Cruz , Marilyn was a Professor of Computer Science at University of Sheffield. From 1996 to 2003, she was a Principal Member of Research Staff in the Speech and Information Processing Lab at AT&T Bell Labs and AT&T Research. While at AT&T, Marilyn worked on the AT&T Communicator project, where she developed one of the first architectures for spoken dialogue systems that used statistical methods for dialogue management and generation. Marilyn has published more than 300 papers, has an H-index of 75, and has 10 U.S. patents. She received her M.S. in Computer Science from Stanford in 1988 and her Ph.D. in Computer and Information Science from University of Pennsylvania (1993).
Neuro-Symbolic Dialogue Management using Prompt-Based Transfer Learning for Dialogue Act Controlled Open-Domain NLG
In order to create interesting and engaging conversational interactions with users, open domain SocialBots need to interact using a range of dialogue acts (DAs).
For example, a SocialBot should be able to ask factual and opinion questions, inform the user of facts and express opinions, agree and disagree with the user, provide appraisals and acknowledgements, make recommendations or suggestions, and confirm what the user said. For many applications it is also necessary to ground these DAs in knowledge of some kind, either structured or unstructured. In the past, such dialogue-act controlled response generation was typically trained from a large paired corpus that maps from a domain-specific meaning representation that specifies the desired DA and associated attributes, to one or more reference utterances. However recent advances in pretrained language models offer new possibilities for semantically controlled NLG. Here we show that we can achieve near perfect DA and semantic attribute control using Prompt-Based Transfer learning (PBL). We apply an overgenerate and rank method to compare eight few-shot prompt styles that include a novel method of generating from textual pseudo-references using a textual style transfer approach, a second novel approach that provides definitions of DAs in the prompts, inspired by previous work on schema-guided NLG, and a baseline of simply linearizing the MR. To our knowledge, this is the first work on NLG for dialogue that automatically evaluates and ranks outputs using DA accuracy. We then show that we can use PBL to successfully transfer these conversational DAs from WikiData triples in one domain, namely Video Games, to Wikidata triples in three other domains, namely Music, Movies and TV, providing a universal dialogue policy that can be used across all 4 domains in Athena, UCSC's Alexa Prize SocialBot.
Jason Weston
Jason Weston is a research scientist at Meta AI, USA and a Visiting Research Professor at NYU. He earned his PhD in machine learning at Royal Holloway, University of London and at AT&T Research in Red Bank, NJ (advisors: Alex Gammerman, Volodya Vovk and Vladimir Vapnik) in 2000. From 2000 to 2001, he was a researcher at Biowulf technologies. From 2002 to 2003 he was a research scientist at the Max Planck Institute for Biological Cybernetics, Tuebingen, Germany. From 2003 to 2009 he was a research staff member at NEC Labs America, Princeton. From 2009 to 2014 he was a research scientist at Google, NY. His interests lie in statistical machine learning, with a focus on reasoning, memory, perception, interaction and communication. Jason has published over 100 papers, including best paper awards at ICML and ECML, and a Test of Time Award for his work "A Unified Architecture for Natural Language Processing: Deep Neural Networks with Multitask Learning", ICML 2008 (with Ronan Collobert). He was part of the YouTube team that won a National Academy of Television Arts & Sciences Emmy Award for Technology and Engineering for Personalized Recommendation Engines for Video Discovery. He was listed as the 16th most influential machine learning scholar at AMiner and one of the top 50 authors in Computer Science in Science.
Improving Open Language Models by Learning from Organic Interactions
We discuss techniques that can be used to learn how to improve AIs (dialogue models) by interacting with organic users ``in the wild''. Training models with organic data is challenging because such interactions include both high quality conversations and feedback, as well as adversarial and toxic behavior. We thus study techniques that enable learning from helpful teachers while avoiding learning from people who are trying to trick the model into unhelpful or toxic responses. We present BlenderBot 3x, an update on the conversational model BlenderBot 3, trained on 6M such interactions from participating users of the system, which we also publicly release. BlenderBot 3x is both preferred in conversation to BlenderBot 3, and is shown to produce safer responses in challenging situations. We then discuss how we believe continued use of these techniques -- and improved variants -- can lead to further gains.
Zhou Yu
Zhou Yu is an Associate Professor at Columbia University Computer Science Department. She obtained her Ph.D. from Carnegie Mellon University in 2017. Dr. Yu has built various dialog systems with major practical impacts, such as a job interview training system, a depression screening system, an Alexa social chatbot, and a second language learning system. Her research interest includes dialog systems, language understanding and generation, vision and language, human-computer interaction, and social robots. Dr. Yu's work earned a 2019 ACL best paper award nomination. She was recognized in the Forbes 2018 30 under 30 in Science and won the 2018 Amazon Alexa Prize. Zhou also is the CEO of Articulate AI Inc, a startup that provides online second-language communication training using AI.
LLMs with long-term memory and better factuality
Seamlessly communicating with machines has always been the ultimate goal of artificial intelligence. This talk addresses the two key milestones towards general intelligence: how to effectively track infinite long history and improve the generation content's factuality. Specifically, we will talk about a stateful transformer architecture that can achieve effective memory read and write. We will also address how to use reinforcement learning with human feedback to improve generation content's faithfulness.