Published March 16, 2021
While your grandma probably hasn’t ever heard the term “socialbot,” it’s possible that she’s used one — and if not, that her life could be enhanced by one.
Improving the quality and usability of socialbots, such as Amazon Echo’s Alexa or Apple’s Siri, is the mission of a UB team taking part in the Alexa Prize Socialbot Grand Challenge.
The cross-disciplinary UB team, including members from the College of Arts and Sciences and the School of Engineering and Applied Sciences, is one of nine teams selected from universities around the world. The team has undertaken the challenge of creating a socialbot “that can converse coherently and engagingly for 20 minutes with humans on a range of current events and popular topics such as entertainment, sports, politics, technology and fashion.” While this is the fourth year of the challenge, this is the first time a UB team has entered.
The multimillion-dollar competition offers all participating university teams a $250,000 research grant, Alexa-enabled devices, free Amazon web services to support their development efforts, and access to other tools, data and Alexa team support.
“We’re thrilled to be competing against prestigious teams across the globe, including some of our role models at Stanford,” says faculty adviser Rohini Srihari, professor of computer science and engineering, and adjunct professor of linguistics.
“Drawing on a unique combination of skills, we hope our charismatic, empathetic and knowledgeable socialbot provides a memorable experience. With the dream of advancing research in conversational AI, Team Proto strives to live up to its name: the first.”
Under the direction of Srihari, Team Proto’s members are Sougata Saha and Souvik Das, both PhD students in the Department of Computer Science and Engineering, and Elizabeth Soper and Erin Pacquetet, both PhD students in the Department of Linguistics.
The students were inspired to participate in the competition after taking a seminar in conversational AI with Srihari last spring in which they looked at several papers from the previous year’s Alexa Prize Socialbot Grand Challenge.
Team Proto submitted its application in August and was notified of its selection a few months later. The competition officially kicked off last November.
The team just finished the beta testing period, in which teams’ socialbots were tried out by the public. Every day, more than 7,000 people interacted with Team Proto’s prototype and provided feedback. The team uses this feedback to constantly improve its socialbot’s abilities.
“We’re really getting to witness firsthand how human psychology and natural language are correlated,” says Das, Team Proto’s co-lead. “This competition shows how other factors — such as mood, political events and time of day — can influence how people interact with the socialbot. The variety of possible interactions surprises me, and at the same time makes this competition very interesting.”
“I am thrilled by the fact that Alexa users all across the U.S. are currently interacting with a bot that we built,” adds Saha, the team’s other co-lead. “The whole idea of building a scalable AI system and serving a large user base is exciting.”
Creating an end-to-end conversational system is similar to working with heavy machinery with lots of moving parts, and multiple points of failure, Srihari explains. The work is inherently interdisciplinary: The group has a shared foundation of courses in both computer science and linguistics.
“Conversational AI falls at the intersection of natural language processing (NLP), deep learning and information retrieval,” says Das, whose PhD research explores research gaps in conversational AI. “Right now, I’m directly using the concepts I’ve learned in previous courses to make conversational systems more robust.”
The innovation of the team members’ design lies in how they leverage deep learning and neural conversation-generation techniques to generate multiple hypotheses for what the person will say next, and then experiment with an array of models to select the best-fitting hypothesis as a response.
"Current research in AI and NLP is based on deep (neural) learning models,” says Srihari. “However, due to the complexities of conversation that involve context, background knowledge and personalities, there is a need to incorporate symbolic approaches as well for more effective conversations.”
“As a linguist, it is fascinating to deconstruct what we think we know about discourse and conversations to be able to implement changes to our bot that will provide a more valuable experience to users,” says Pacquetet. “Socialbots do not produce language the same way people do, and it is forcing us to reinvent the way we think about discourse as a whole.”
“Because of the challenge, I’ve learned a lot about how humans interact with technology, and how it differs from how we interact with other humans,” says Soper. “I’ve learned that a simple solution that can be implemented quickly is sometimes better than a sophisticated solution that takes a lot of effort to build.”
The UB team recently was one of five teams to qualify for the quarterfinals. After that, the field will be further narrowed, with three teams advancing to the semifinal round. The finals take place in July, with the winner being announced in August.
The winning team will receive a prize of $500,000. The second- and third-place teams will receive prizes of $100,000 and $50,000, respectively. An additional $1 million research grant will be awarded to the winning team’s university if the team receives a composite score of 4.0 or higher (out of 5.0) and at least two-thirds of their socialbot’s conversations with interactors last for 20 minutes.
“We would like to tell stories that illustrate the potential of AI for social impact,” Srihari says. “For example, we want to show how socialbots can help a grandparent, or someone who feels lonely or isolated. We ultimately hope that they can help people better understand what’s becoming an increasingly complex, and sometimes scary, world.”