Facebook Create Non-English-Centric Translation System for 100 Languages

The Facebook AI team have developed an Artificial Intelligence (AI) machine translation which is capable of translating between any pair of 100 languages without the need to translate into English first, as most machine translation systems currently do, according to the Facebook AI website.

The new technology is believed to deliver a more accurate translation than previously available and was shown in tests to out-perform English-centric translations by 10 points on the 100-point BLEU scale used by academics to automatically evaluate the accuracy and quality of machine translations.

Angela Fan, one of the research assistants at Facebook AI, explained that the machine translation generated better results than previous AI translations as it removes the intermediate step of using English as a ‘bridge’ language.

She said the “milestone is a culmination of years of Facebook AI’s foundational work in machine translation” and that it would help to “bring people together, provide authoritative information on COVID-19, and keep them safe from harmful content.”

The new technology, known as M2M-100 (many-to-many), is trained on a total of 2200 language directions and includes a diverse range of languages from Arabic to Zulu.

However, according to the academic research, there are some language pairs which perform better than others using the new technology.

Pairs which have a large quantity of existing ‘mined’ data such as, Spanish-Portuguese were more successful in translation, while other languages such as, English-Belarussian learnt from parallels with ‘family’ languages (languages which share similarities) such as, Belarussian and Russian.

The machine translation is not yet available on the social media site, but they hope that it will eventually help with the 20billion translations made each day on Facebook when users click ‘Translate’ for posts on their feed.

The social media site has been slated in the past for poor translation quality across multiple languages and includes blunders which have seen a Palestinian man being arrested by Israeli police after it mistranslated his post from ‘Good Morning’ in English to ‘attack them’ in Hebrew, and in January this year (2020) Facebook publicly apologised for mistranslating Chinese leader, Xi Jinping’s name into ‘Mr Shithole’ when it was automatically translated from Burmese into English.