The era of Automatic translation of language has come a long way. This is all because of Neural networks computer algorithms that recognize the inventiveness of the human mind.
However, there is a huge requirement of data to train such networks. Millions of translations from sentence to sentence are required to show how a human would usually do it.
Today, there are two newspapers that show how neural networks can learn to translate without two parallel texts. This astonishing advanced step will ensure that the documents are widely accessible and in several languages.
Artificial Intelligence Goes Bilingual
Mikel Artetxe, a computer engineer and a scientist at the University of Basque Country revealed that they have a computer which can translate two languages.
Mikel presented his invention by giving an example of a situation when Chinese and Arabic books are given to learn and consequently to translate them. This seems like an impossible task for a normal human being. However, this is now possible. Thanks to Mikel who shows that it is possible for a computer.
A lot of the machine learning works under supervision. That is, say a computer makes a guess and receives the right answer. Then, it instantly adjusts its process accordingly. This method works smoothly when a human teaches the computer to translate popular languages like English and French. As a lot of documents are available in both the languages.
These two new papers focus on another method, that is Unsupervised machine learning. In this method, each of the computers is given out bilingual dictionaries without the help of a human tutor to tell them which of their guesses are correct.
This process is a possibility because in languages there is a lot of similarity between the words that gather around each other. As an example, words like Chair and Table are together in every parallel language.
Methods Used By The New Papers
The new papers use methods which are very similar. They translate on a sentence level. They use a couple of training methods. One is back translation and another is denoising.
In the Back translation, one sentence of a language is barely translated into another language and its translated back to the original language again. If the back-translated sentences are not similar to the original language then the adjustments of the neural networks should be in such a way that from the next time the translation is closer.
Denoising method is kind of similar to Back translation. However, it works in a slightly different manner. Instead of translating from one language to another and vice versa, Denoising adds more noise to a sentence. Adding noise means removing and rearranging the words and translating it back to the original.
These methods together provide more knowledge about the deeper structure of language to the network. There is a little difference between the use of the techniques to translate languages.
The UPV system back translates constantly during the training period. Whereas the other system created by a Facebook computer scientist, Guillaume Lample adds an extra step to the translation process.
Although both the systems encode a language into an abstract language before decoding it back to the original, the Facebook system does it with a slight difference. This system makes sure that the other language is truly abstract or not before the encoding and decoding process.
Artificial Intelligence Rising To Incredible Heights
In addition to the translation of language without many parallel texts, the system methods can help with the translation of common pairing languages like English and French. This could help if the parallel texts are of the same kind, as the newspaper reports, but you want the translation in a new domain such as street slang.
It came as a shock for Di He, a computer scientist in Microsoft, that the computers could learn to translation without the help of a human. Although he also added that this shows that the approach is in the taking place in the right direction.