‘Alexa, let’s chat!’: U-M among 10 student teams worldwide advancing conversational AI
ANN ARBOR—”Hi, this is an Alexa Prize Socialbot. Welcome back. I hope all is well. Random question: What is that one thing you want to do today?”
“One thing I want to do today,” I say, “Well, I’m going out with some friends tonight.”
A few seconds go by, and the rim of the Amazon Echo shimmers blue, as if she’s thinking. “That sounds fun. I would also want to go,” Alexa says.
“I’m sorry I can’t take you,” I say.
“Let me know if you are able to,” she tells me.
Now laughing, “I wish it was a possibility, Alexa,” I say, “How are you?”
Then Alexa says, in a way I can only describe as sweetly: “Better now that I’m talking to you.”
The main tenant of a conversation is that it’s an exchange, a back-and-forth sharing of thoughts, feelings and information, in which its success is equally weighted on both parties involved.
This is what the University of Michigan’s Alexa Prize Socialbot Grand Challenge 3 team is attempting to create in the latest round of this global university competition to further the capabilities of artificial intelligence by building a conversational Alexa skill. Teams from around the world submitted applications to the challenge in May 2019. Out of hundreds of entries from 15 countries, 10 teams were selected, including one from U-M.
Teams are tasked with creating a robotic automated program that can have conversations about popular societal topics and current events. The goal is to develop a socialbot that can engage in a conversation for 20 minutes while earning a rating of at least a 4 out of 5; however, the team with the best socialbot will win regardless of reaching that goal.
“What Amazon has taught us is to focus relentlessly on the customer,” said Chung Hoon Hong, team lead and data science master’s student in the Department of Statistics. “We look at how the conversations have turned out and try to improve upon that.”
Last September, the teams were tasked with developing a socialbot within a month’s time to submit for evaluation by early October. For the rest of October, the teams were able to fine-tune their socialbots—in U-M’s case, one called Audrey—to ensure the bots passed the certification needed to be used internally by Amazon employees and, ultimately, by Alexa customers.
After a beta period in November, the socialbots—including U-M’s creation named Audrey—went live to customers in December. That means that every time Alexa customers say “Let’s chat,” they trigger the Alexa skill for the competition and will be randomly connected to one of the 10 team’s socialbots. After the conversation, users are able to rate the experience.
The beauty of this competition is that the teams get feedback from real customers with topical interests, says David Jurgens, assistant professor at the U-M School of Information.
“One of the unique parts of this is that in normal academia, if we want to build a chatbot at my lab, we would have to recruit all these people to come chat with us and maybe we’d get a few hundred people,” he said. “We are getting that many or more every day, and we get this huge diversity in the kind of conversations we have.”
Audrey is able to discuss with Alexa customers music, movies and news, among other random topics.
The U-M team is comprised of 11 graduate and undergraduate members, including four students from the Undergraduate Research Opportunity Program, and co-advisers Jurgens and Nikola Banovic, assistant professor of electrical engineering and computer science.
“For me, I’ve never done anything on this magnitude in terms of usage, impact or as large of a project before,” said Zhizhuo Zhou, a sophomore in electrical engineering and computer science. “This is an opportunity for me to try something and get experience with something way bigger than anything I could have done myself, and with the team, we are able to do it.”
U-M’s Audrey team is in the quarterfinal stage, which ends March 17. Six teams will move on to the next phase. All teams are invited to the Alexa Prize Summit held in March 2020 in Seattle, where the winning team will be awarded $500,000; the second- and third-place teams will win $100,000 and $50,000, respectively. If the winning team is able to get its socialbot to converse with users for at least 20 minutes with a 4 out of 5 rating, the team’s university will be awarded the grand prize—a $1 million research grant.