Facebook had an AI Team Technology known as FAIR (Facebook Artificial Intelligence Research). The research team has now tested a new approach that teaches bots how to chat like humans.
Well, the research team have used a special 164,000-utterance dataset known as “Persona-Chat” to teach its Artificial Intelligence how to chat like humans.
According to the reports from The Verge, Persona Chat consists of more than 160,000 lines of dialogue which are sourced from workers found on Amazon’s “Mechanical Turk” marketplace.
For those who don’t know about Amazon’s MTurk, it’s a crowdsourcing internet marketplace which enables businesses and individuals to coordinate the use of human intelligence to perform tasks that computers can’t do.
The research team of Facebook used Amazon’s MTurk data to train neural networks used for existing chatbots, with the results then accessed by another group of Mechanical Turkers. In the test, the group of Mechanical Turkers was asked to have the conversation with persona-driven bots.
The results were somehow satisfactory, the test showed that the persona-driven bots didn’t score much on “fluency” and “consistency”. However, they were way better than normal chatbots.
The new test from Facebook is important because last year the company had to shut down one of its AI systems when they found that their chatbots have created their own language and started communication in an entirely new language which they created without human input.