A group of researchers at the UMass Amherst College of Information and Computer Sciences (CICS) led by Assistant Professor Hamed Zamani and doctoral student Helia Hashemi is one of 10 teams worldwide selected to compete in the 2021 Alexa Prize TaskBot Challenge. Each team receives $250,000 in funding to compete in a conversational artificial intelligence challenge to develop agents that assist customers in completing cooking and do-it-yourself projects that require multiple steps and decisions. The winning team will receive a $500,000 prize.
Zamani, who serves as associate director of the Center for Intelligent Information Retrieval (CIIR) at CICS, and Hashemi will work with a group of students at the CIIR advised by Zamani, along with students advised by Distinguished Professor Emeritus Bruce Croft and Assistant Professor Mohit Iyyer.
“We are excited to take on the TaskBot Challenge,” Zamani says. “As a world-leading research group on conversational search, we have recently developed models for mixed-initiative conversational search and invested in pushing the boundaries of multi-modal conversational search systems. This challenge is a unique opportunity for us to expand our development and evaluate our systems within the Alexa platform by interacting with real users.”
The team’s focus will be on the information retrieval and natural language processing aspects of the challenge. According to Zamani, they will work to create an information-need understanding pipeline that can identify correct recipes or instructions, as well as a pipeline for walking users all the way through a selected project. To support these pipelines, the researchers will work to develop an interactive retrieval model that can handle various types of complex queries, design a novel multimodal result presentation model that highlights the unique characteristics of each retrieved item, and train models that can create mixed initiative interactions with the users.
More information about the TaskBot Challenge can be found at the Alexa Prize FAQ page.