A Meta team dedicated to fundamental AI research is working with UNESCO on an “advanced artificial intelligence” project. It draws on the principles of human reasoning to perform cognitively demanding tasks, such as translation.
Open-source AI translation models
One area of work is to better integrate underrepresented languages into AI models. This development not only promotes linguistic diversity and inclusivity in the digital world; it also helps to “design intelligent systems that can adapt to new situations,” as Meta explains.
As part of the International Decade of Indigenous Languages organized by UNESCO, the company is looking for partners who can produce around ten hours of voice recordings with transcriptions, as well as texts comprising 200 sentences or more, and sets of sentences translated into several languages.
These documents will be integrated into AI models for voice recognition and machine translation. The models will then be open sourced and made freely available to all. Participating partners will have access to technical workshops to learn how to use the models. The Government of Nunavut in Canada is also collaborating with Meta by sharing data in the Inuit languages, Inuktitut and Inuinnaqtun.
The company has also put online a translation benchmark, also open source, to help evaluate the performance of AI models. “Composed of sentences carefully crafted by linguistic experts, this benchmark aims to illustrate the diversity of human languages,” explains Meta. The tool is available at this address.
Source: Meta
0 Comments