LASER (Language-Agnostic Sentence Representations)

2019: Disclosure of source codes

In January, 2019 Facebook opened source codes of the LASER tool (Language-Agnostic Sentence Representations) intended for understanding of any natural speech. This step the world's largest social network made the next key contribution to accelerating the translation of the applications which are engaged in processing of natural language texts and voice messages on bigger quantity of languages.

According to the SiliconANGLE edition, the LASER system which is based on library of deep learning PyTorch is applied by Facebook to creation of a certain mathematical model which can encapsulate and understand all natural languages whatever unique they were.

Facebook opened source codes of the tool for understanding of any natural speech

Source codes of it tools were uploaded publicly a month later after an exit of a research of Facebook in which specialists told about training of uniform model of neural network which can model structure from 93 languages in 34 separate alphabets.

Finally Facebook received as the company calls it, "uniform representation" or mathematical conversion of offers in the form of vectors which includes structural similarity of all 93 languages.

This only representation then was used for training of algorithms in several tasks assuming comparison of offers between pairs of languages which the program never saw earlier. For example, the translation from Swedish into the Swahili and a method allowing to learn the languages "from scratch" was implemented.

The researcher of Facebook Holger Schwenk tells, thanks to LASER languages with "limited resources", as expected, it will be possible to study together with more popular languages that it was possible to construct models of natural languag processing for their understanding.^[1]

Notes