Alexa Prize: Amazon’s Battle to Bring Conversational AI Into Your Home | WIRED
▻https://www.wired.com/story/inside-amazon-alexa-prize
Amazon, in case you haven’t noticed, has spent the past few years pursuing voice AI with a voraciousness rivaling that of its conquest of retail. The company has more than 5,000 people working on the Alexa platform. And since just 2015, it has reportedly sold more than 20 million Echoes. One day, Amazon believes, AIs will do much more than merely control lights and playlists. They will drive cars, diagnose diseases, and permeate every niche of our lives. Voice will be the predominant interface, and conversation itself—helpful, informative, companionable, entertaining—will be the ultimate product.
Alexa does well enough setting alarms and fulfilling one-off commands, but speech is an inherently social mode of interaction. “People are expecting Alexa to talk to them just like a friend,” says Ashwin Ram, who leads Alexa’s AI research team. Taking part in human conversation—with all its infinite variability, abrupt changes in context, and flashes of connection—is widely recognized as one of the hardest problems in AI, and Amazon has charged into it headlong.
The Alexa Prize is hardly the first contest that has tried to squeeze more humanlike rapport out of the world’s chatbots. Every year for the better part of three decades, a smattering of computer scientists and hobbyists has gathered to compete for something called the Loebner Prize, in which contestants try to trick judges into believing a chatbot is human. That prize has inspired its share of controversy over the years—some AI researchers call it a publicity stunt—along with plenty of wistful, poetic ruminations on what divides humans from machines. But the Alexa Prize is different in a couple of ways. First, the point isn’t to fool anyone that Alexa is a person. Second, the scale of the competition—the sheer human, financial, and computational firepower behind it—is massive. For several months of 2017, during an early phase of the contest, anyone in the US who said “Alexa, let’s chat” to their Amazon voice device was allowed to converse with a randomly selected contest bot; they were then invited to rate the conversation they’d had from one to five stars. The bots had millions of rated interactions, making the Alexa Prize competition, by orders of magnitude, the largest chatbot showdown the world has ever seen.
THE FEVERED QUEST for conversational AI has pitted Amazon, Apple, Facebook, Google, and Microsoft in a battle for two vital resources. The first is finite: top-shelf PhDs in computer science, who, owing to their scarcity, now command starting salaries well into the six figures. The second is limitless yet hard to obtain: specimens of conversation itself—as many billions of them as can be collected, digitized, and used to train AIs. Against this backdrop, the Alexa Prize was a masterstroke for Amazon. The contest served as both a talent search for the sharpest graduate students in the world and a chance to pick their brains for a bargain price. And it provided Amazon with an opportunity to amass a conversational data trove that no other technology company has.
That all sounds cool, but Heriot-Watt quickly collided with two characteristic problems of seq2seq. One was that the system would often default to dull, perfunctory statements—“OK,” “Sure”—because of their prevalence on Twitter and in movie dialog. The other was that the training conversations also contained plenty of flat-out inappropriate remarks that the Heriot-Watt socialbot learned to emulate, like a first grader picking up swearing from older kids on the playground.
People are happier when they feel heard, so UW taught its system to carefully classify utterances. Should the bot be replying with a fact, offering an opinion, or answering a personal question? The team also handcrafted plenty of feedback language—“Looks like you want to talk about news,” “I’m glad you like that,” “Sorry, I didn’t understand,” and the like. Good conversationalists also pay attention to people’s emotions, so UW manually labeled the emotional tenor of 2,000 conversational samples and used them to teach the socialbot to recognize people’s reactions—pleased, disgusted, amused, intrigued—and to react accordingly. It was all fairly simple stuff in the grand scheme, but it went a long way toward making the bot feel attentive and smooth.