III: Small: Toward User-Centric Task-Oriented Dialog Systems via Zero-shot Adaptation and Personalization

III: Small: Toward User-Centric Task-Oriented Dialog Systems via Zero-shot Adaptation and Personalization Overview Task-oriented dialog systems help users accomplish tasks through expressive and accessible natural language interactions. Examples abound, from the most basic tasks such as ordering a pizza or reserv- ing a table at the restaurant to more involved ones like customer support and automatic disease diagnosis. Recent advancements in language modeling have eliminated the need for expensive, domain-speci?c hand- crafting. Nonetheless, state-of-the-art task-oriented dialog systems (a) require labeled training data for each new task or domain; (b) can not accommodate user requests beyond structured databases; and (c) lack the ability to offer personalized interactions. These limitations hinder adoption and restrict the potential for optimal user experiences. In this proposal, our goal is to investigate the next generation of task-oriented dialog systems that can: (i) seamlessly adapt to new unseen intents or domains without requiring labeled training data for each new domain; (ii) provide grounded responses to users’ requests beyond the limited database ?elds; and (iii) tailor their conversational ?ows according to the users’ personality traits. Keywords: natural language processing; task-oriented dialog systems; natural language interfaces. Intellectual Merit We plan to accomplish this goal via three research thrusts. In each, we aspire to extend the state of the art to enable truly utilitarian and human-to-human-like interactions in task-oriented dialog systems. In our ?rst thrust, we aim to build mixed-initiative dialogue models. These models can learn generic task- completion skills from seen domains and generalize to unseen domains by comprehending task semantics on the ?y. Next, we seek to equip models with the ability to interactively utilize a range of external tools such as relevant enterprise systems. This enables them to ful?ll “off-script” user requests – different from originally planned scenarios. Finally, we focus on personalizing dialog models to contextualize individual users’ needs (explicit and implicit) and provide personalized assistance. Methodologically, this proposal advances the areas of large-scale language modeling, sequential decision- making, and their intersection. Notably, we seek to enhance the adaptation of task-agnostic pre-trained large language models to complex tasks requiring mixed initiatives and multiple interactions by incorpo- rating interactive feedback and performing long-term planning. Moreover, we aim to extend task-speci?c sequential decision-making systems to generalize better to unseen tasks by leveraging the robust represen- tations of human language and world knowledge from pre-trained large language models. The contributions outlined in this proposal will establish a solid foundation for developing robust and reliable natural language interfaces, a long-standing goal of AI research. By enabling the most intuitive and natural way of communication, this research has the potential to revolutionize how individuals engage with computing systems for performing a wide range of tasks, leading to improved productivity, increased accessibility, and enhanced user experiences. Broader Impacts Industry: Task-oriented dialog systems have signi?cant implications in a broad range of areas including customer service, healthcare, and education. This proposal intends to establish a collaboration with Home Depot to develop a dialog system that facilitates users to ful?ll their home improvement intents. We are also collaborating with PreventScripts to develop a conversational AI companion that offers scienti?cally grounded information to individuals seeking weight loss guidance. Education: This project will have a signi?cant educational impact at all levels – from elementary to high school, and from undergraduate to graduate studies. Speci?cally, we will teach periodic seminars for Hispanic students at the University of Kentucky in collaboration with the Society of Hispanic Professional Engineers. Moreover, we will organize an informational workshop series named “Bits & Bots” for the stu- dents of Cardinal Valley Elementary School in Lexington, Kentucky. This school serves a population that is 73% Hispanic. Furthermore, we will engage with students of the Math, Science, and Technology Center program at Paul Laurence Dunbar High School in Lexington, Kentucky. In particular, we will organize a series of interactive workshops, called “AI Code Camp”, to introduce students to the fundamental princi- ples of AI and to demonstrate its potential using a combination of hands-on experiences and collaborative exploratory activities. We also intend to integrate the ?ndings of our research into undergraduate and graduate Machine Learning and Natural Language Processing courses, taught by the PIs.