The American social media conglomerate corporation, Facebook has introduced a new AI-based MMT model that is capable of translating 100*100 languages without any dependence on English-centric data. The new single multilingual model known as M2M-100 is trained on a total of 2200 language directions and has attained a 10 BLEU point progress when compared to English-centric multilingual models.
Forming a large volume of quality parallel sentences for translation directions without involving English was the major struggle of developing an MMT model. What’s more, the required volume of data for translation grows quadratically with the adding up of languages. However, Facebook took this as a challenge and created 7.5 billion sentences over 100 hundred languages.
The firm combined complementary data mining resources like ccAligned, ccMatrix as well as LASER. With this novel model, it came up with a new LASER 2.0. The LASER 2.0 encloses improved fast Text language identification that will further enhance the quality of mining and includes open-sourced training as well as evaluation scripts. Facebook prioritized the most translation requests to race with this intense, high computational data. In addition to these, it also prioritized mining directions with the highest quality and the largest quantity of data.
At first, Facebook used novel mining strategies to craft 7.5 billion translation data with an original and accurate many-to-many data set for 100 languages. Later, several scaling techniques were used to bring this number to 15 billion parameters. Also, the model was able to detain information from correlated languages and reflected a more varied script of languages along with morphology. This will improve the quality of translations for billions of people daily. Additionally, they also came up with a new bridge mining strategy to group languages based on classification, geography, and cultural similarities. Also, this model is the first to use Fairscale, the new PyTorch library. These enhanced the results on zero-shot settings and were significantly better than English-centric models.
This AI-based language models like M2M-100 will assist researchers to apply their best effort and skill towards coming up with a single universal language model that can be deployed across diverse tasks. Further, It will move forward the industry to create a single model that supports all languages, keep translations up-to-date, and finally, assist the people.
