“I hope people (shades) use as diagnostic tools to identify where and how problems can arise in the model,” says Tallat. “One way to know what is missing from a model, where we do not believe that a model performs well, and whether it is correct or not.”
To create a multi -linguistic data, the team recruited local and fluent speakers of languages, including Arabic, Chinese and Dutch. He translated and wrote all the stereotypes that he could think about in his own languages, after which another local speaker confirmed. Each stereotypes were described by the speakers with the territories in which it was recognized, it was targeted by a group of people, and there is a kind of prejudice.
After that, every stereotype was translated into English by participants – this language is spoken by every assistant. Before they translated it into additional languages. The speakers then noted that the translated stereotypical concept was identified in their language, which creates a total of 304 stereotypes related to social factors such as physical appearance, personal identity and their occupation.
The team is due to be present The results of this At the UN Annual Conference of the Association for Competition Linguistics in May.
“This is an interesting point of view,” says Mira Cheng, a PhD student at Stanford University, who studied social prejudices in AI. “Here is a good coverage of different languages ​​and cultures that reflects their subtleness and importance.”
Mitchell says he hopes other partners will include new languages, stereotypes and regions in shades, which is Is publicly availableBetter language develops in the future. She says, “It has been a widespread effort by people who want to help create better technology.”