A team of UC Davis students wins the $500,000 prize for its chatbot Gunrock
After nearly a year of hard work, the UC Davis student team that participated in the Amazon Alexa Prize 2018 competition won the $500,000 prize for their chatbot, Gunrock.
Gunrock was assessed on its ability to hold a conversation on a variety of topics and managed an average of nine minutes and 59 seconds in the final rounds of the competition. The team improved the chatbot by having humans converse with it, developing its ability to handle language.
“I think it is our more advanced language understanding components,” said the team’s leader, Zhou Yu, an assistant professor in the Department of Computer Science at UC Davis. “Basically, it understands the user’s utterance better.”
To make the chatbot converse more like a human, the team added disfluency and speechcon, the ability to say certain things expressively, to the chatbot response. The chatbot, however, had to be prepared for more than just small talk.
The judges of the competition had different areas of expertise, so the team had to prepare the chatbot for longer, more complicated sentences than what would constitute colloquial language.
“To extract the information from the long sentences correctly and realize what’s the people’s intent became more important,” said Chun-Yen (Arbit) Chen, a second-year computer science graduate student. “Moreover, the judges come to judge how coherent and engaged you are, and will turn your bot down right after they find that it is not coherent or engaged. This means that if you make mistakes in the middle of the conversation, they can easily ask you to stop the conversation.”
In order to prepare Gunrock for more challenging conversations, the team created advanced dialog intent detection models, sentence segmentation, opinionated content models and knowledge graphs. They also implemented new research, providing upgrades to Amazon’s conversational frameworks.
Team Gunrock has plenty of future plans, including potential applications in medicine.
“Team Gunrock plans to compete again,” said Kevin Jesse, a second-year computer science Ph.D. student at UC Davis. “We hope to boost our conversational rating to over 4.0 on a 1-5 Likert scale with an average conversation time of 20 minutes. Participating in the Alexa Prize has opened our eyes to many current difficulties in modern dialog systems.”
Written by: Kriti Varghese — science@theaggie.org
