Google Translate is getting more brains
Last year, a major leap in artificial intelligence quality was achieved by Google when they announced a neural machine translation (NMT) deep learning system.
It says this artificial intelligent system utilizes state-of-the-art training techniques reducing translation errors across its Google Translate service by 55%-85% on several major language pairs.
“Ten years ago, we announced the launch of Google Translate, together with the use of Phrase-Based Machine Translation (PBMT) as the key algorithm behind this service”, they said.
“Today we announce the Google Neural Machine Translation system (GNMT), which utilizes state-of-the-art training techniques to achieve the largest improvements to date for machine translation quality”
Recurrent Neural Network
A recurrent neural network is a kind of neural network that has in-state codifiication inside of it. Due to this, it is able to show a behaviour that changes with time. This is important as the behaviour of a language changes in time.
Through a recurrent neural network, the system can translate a language into another one which it has not been trained for. The system, called Google Neural Machine Translation System, works using a common language not understood by humans, specially structured to help the translation.
Traditionally used phrase-based machine translation (PBMT) system, which breaks an input sentence into individual words and phrases to be translated largely independently, is widely disregarded for it’s bad translation.
In contrast, the new unveiled Neural Machine Translation (NMT) system works on the the entire input sentence as a single unit for translation, as mentioned, passes it through a recurrent neural network.Thus giving it better context to figure out the best possible translation.
In Picture below, it is possible to see some cases were new system has better results then the phrase-based system.
Comparison between the types of translation. Source: Google
Google Neural Machine Translation (GNMT)
Most of the fast becoming popular neural machine-translation systems though only work on a single pair of languages. Google has been working on improving neural machine translation over the past few years to make it work on multiple pairs-and making it capable enough to translate between language pairs not directly trained to translate.
For example, if the system needs do translate from Portuguese into German, but is only trained to translate from Portuguese into English and from English into German, it can translate from Portuguese into German without the need for translating into English, as it builds the structure of meaning in this new language.
Taking human-generated side-by-side comparison as a metric bar, for English-to-French and English-to-German benchmarks, GNMT system translations errors are reduced by an average of 60%. The results are significantly improved compared to the previous phrase-based translation systems.
The visual illustration below shows the progression of Google-NMT as it translates a text from Chinese to English. Initially, the “Encoder” encodes the Chinese words as a list of vectors. Once the entire sentence is read, the decoder starts forming the English sentence.
This new system is already implemented in google translator in eight languages pairs: to and from English and French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish. According to google, this represent the native language of nearly one-third of world’s population and 35% of all google translator queries.
Google is also offering an API for programmers who want to use this google technology. This will help google technology to spread to other system and will make easy for programmer to access this state of art technology.
Google Translate currently translates over 140 billion words every day in 103 languages.
So human-like AI systems are just around the corner?
With human-like AI systems like GNMT, it is already being speculated that human translators will soon be put of job but that’s not the truth.
Neural translation technology will do a great job for simple translations, but unlike human translator it can still make significant errors, like dropping words, not recognizing rare or special words, mistranslating names and translating sentences away from the context of the original gist of the paragraph or sentences.
To match human capability, computers will have to be taught some basic world information as well as understanding of certain areas of translation.
There is still a plenty of work to be done for better user-friendly experience. However, GNMT is still no less than a significant milestone.