Meta announced the open-sourcing of their newest creation, a single artificial intelligence (AI) powered tool that can translate more than 200 languages. This takes Meta deeper into its mission of interconnecting the world using some of the world's breakthrough technologies in the modern day.
Meta Announces a New AI Model That Can Translate 200 Languages
Early this year, Meta announced an AI research project on universal speech translation. A few languages such as English, Spanish, Mandarin, Italian, and German are already supported by several machine translation web tools. However, Meta is both ambitious and highly committed to offering a fairer experience for native speakers of less commonly used languages, even those without a standard writing system. To date, the AI model works with 200 different languages with a potential for further growth as Meta is open-sourcing this.
In their blog post, Meta has identified three main challenges for machine translation research: data scarcity, modeling challenges, and evaluation & improvement of results. The company believes that through their two projects, No Language Left Behind and the Universal Speech Translator. These projects make the internet way more inclusive space for native speakers all around the world. Collectively, initiatives like these are expected to help produce real-time and reliable translations in hundreds of languages, ranging from Asturian to Luganda to Urdu.
According to Angela Fan, a research scientist at Meta, the purpose of the project is to pursue an inclusive translation technology. Currently, they are testing the AI to help editors at Wikipedia translate their articles to other languages. Soon enough, the model will be deployed on Meta's ecosystem.
Testing is Still Ongoing, But Deployment on Meta can be Expected Soon
MT experts claim that the AI model's translations are not expected to be perfect. What makes it a cut above other translation tools is its interest in languages other tools do not usually care for. This is how they address the said challenges in data scarcity. About 100 new languages are involved, including low-resource or languages with less than 1 million available translations.
In order to improve and evaluate translation results, Meta tested with sentence pairs for each language included in the AI model. Sentences were translated from English to the respective languages by native speakers who are also professional translators. Machine translation and human-made sentences were compared, and numerical scores were assigned to measure overlaps. The goal is for the model to learn and eventually produce human-quality translations.
To ensure responsible and culturally sensitive translation, Meta claims to do case studies with speakers of over 20 languages. The company is not only studying the technicalities of language-to-language translation but also what people consider important in translating their native languages. Cultural nuances, biases, and potentially harmful language are not easily considered in simple transliteration, and information may not be passed on successfully without these considerations. Meta knows this well, saying their long-term goals in universal translation do not need just simple expertise on AI but the cooperation of numerous experts, researchers, and individuals from around the world.
Of course, much of the benefit goes back to the company itself. A successful universal multilingual translator tool serves Meta's future innovations from a more inclusive VR experience to vast AR features to simply understanding more and more about its users across Facebook, Instagram, and WhatsApp.
Related Article: Meta Sues China-Linked Firm Over Facebook, Instagram Data Scraping
