AI isn’t the end-all-be-all when it comes to conversational experiences like Amazon’s Alexa. In fact, manual solutions to problems are sometimes superior to automated, AI-driven fixes. That’s according to Amazon Alexa AI director of research science Janet Slifka, who spoke during a session today at VentureBeat’s Transform 2020 conference.
“If someone calls customer service and says something to the effect, ‘Alexa doesn’t understand me,’ in many cases, we can be faster to the customer with a manual fix,” Slifka said, describing her team’s work in triage. “In some cases, it’s almost required when you need words to enter the lexicon. You’re probably not going to wait until you rebuilt and deploy a statistical model when a new word gains prominence, like ‘Brexit.’”
This dedication to consistency at the expense of the bleeding edge is reflected in Alexa Answers, a decidedly simplistic approach to expanding Alexa’s base of knowledge. Alexa Answers lets any customer in the U.S. and Germany submit responses to unanswered questions. And while the platform isn’t perfect, Amazon says it allows Alexa to answer questions faster than any automated backend could — especially for breaking news.
“When you start with the customer, you look at all of your tech options. You may not choose the latest and greatest that caught your eye,” Slifka said. “You might allow that the newer things that you’re doing are messier and more challenging. For that reason, you might have more manual processes in there until you really hone things.”
That said, Amazon hopes to apply more autonomous, AI-driven techniques to problems in the future. Scientists at Alexa AI are pursuing semi-supervised and unsupervised learning, in which AI systems learn to make predictions without ingesting gobs of annotated data. One manifestation is Alexa Conversations, an AI-based service that enables developers to create Alexa apps that seamlessly interconnect with other apps. Another is Alexa’s recently launched smart home routine suggestions. If customers ask something like “Alexa, turn on the Sofa Lights” but the lights they’re trying to turn on are actually named “Living Room Lights,” Alexa might helpfully suggest “Did you mean Living Room Lights?”
“The idea is to remove the manual fix because the statistical model is the stronger option for the customer,” Slifka said.