Researchers paid people to talk to themselves so AI will sound less dumb
A team of researchers from the University of Edinburgh developed a novel method for teaching AI to respond to queries in a more conversational manner: They paid people to talk to themselves. It turns out that paying one person to pretend to be two people is a cost-effective way of training an AI to recognize natural human language. That may seem like a no-brainer, but the interesting part is that it also produces better results than current multi-speaker conversation datasets. Here’s an example from the self-dialogue database — despite the denotations this is a single study participant providing both sides…
This story continues at The Next Web
