UofG students turn up the heat on conversational AI
A team of University of Glasgow postgraduate students is setting out to perfect the recipe for next-generation conversational assistants that perform multi-step tasks like cooking and home improvement.
The computing science students are one of just ten teams chosen by Amazon from more than 125 entries across 15 countries to compete in the first Alexa Prize Taskbot Challenge.
The team is part of the Glasgow Representation and Information Learning Lab (GRILL) and is starting work on an AI assistant to guide Amazon Alexa users through complex real-world tasks, focusing initially on cooking and home improvement.
The Glasgow team will receive a research grant of $250,000 (£176,500) to help them over the next 12 months. They will also receive four Amazon Alexa devices, free access to Amazon Web Services, and support from members of the Amazon Alexa team.
The GRILL team consists of School of Computing Science graduate students Sophie Fischer, Carlos Gemmell, Iain Mackie, Paul Owoicho, and Federico Rossetto.
The team’s leader, Carlos Gemmell, said: “We’re thrilled that our entry was successful and we’re excited to begin work on the Alexa Taskbot project.
“Over the next 12 months, we’ll develop novel multimodal deep learning algorithms grounded in tasks that people do in the real world. It will be capable of understanding what users require when performing a task and how to guide them through it by reasoning over past interactions, a structured representation of the task, and curated documents.
“The project aims to develop new conversational question answering technology specialized for complex tasks using state-of-the-art neural deep learning models. Having access to Alexa users around the world will allow us to receive real-time feedback and help us to make it an invaluable task assistant.”
Dr. Jeff Dalton, Lecturer in the School of Computing Science is the team’s faculty advisor and will help guide the project. He is the head of the GRILL Lab, which lends its name to the team’s project. He is also a former Masterchef contestant, appearing on the American version of the popular cooking program in 2011.
Dr. Dalton said: “As an amateur chef myself, as well as an artificial intelligence researcher, I’m excited to work with the team as they cook up in the lab in the coming months.
“Conversational AI assistants like Alexa have come a long way in recent years, but ambitious projects like this are vital to significantly advance their abilities. The team aims to help agents understand what users are doing and how they’re doing it by grounding conversations in the world.
“Ultimately, we hope that GRILL will be able to help people solve complex tasks in the kitchen and beyond. To adapt a recipe to the unexpected dietary preferences of a dinner party guest, explain the reason for adding cornstarch to scrambled eggs and diagnose failed sourdough bread. Although this is an ambitious vision beyond current capability, our team will make key advances towards this goal with multi-modal and visually grounded conversation capability.”
The Alexa Prize, an annual university competition dedicated to accelerating the field of conversational artificial intelligence (AI), recognizes students from around the globe who are changing the way we interact with technology. This year’s taskbot challenge is the first to incorporate multimodal (voice and vision) customer experiences.
Succeeding in the TaskBot Challenge requires significant advancements in state-of-the-art conversational AI, addressing some of the toughest problems in knowledge representation, inference, explainability, language understanding, and commonsense reasoning, while maintaining a coherent and engaging conversation. These have broad implications for conversational agents to assist with new types of real-world tasks that were previously impossible.
Throughout the next year Alexa customers’ feedback will fuel improvements to collaborative human-machine agents and accelerate progress bringing the benefits of AI into the kitchen, garage and beyond.
First published: 26 May 2021