Google Neural Machine Translation
Google Translate is one of the most widely used tools offered by the tech giant. It has more than 500 million users, who use it to translate among 103 languages, spanning 140 billion words every day. That said, anyone who has ever used Google Translate knows that the translations lack accuracy and sometimes the translations are simply hilarious. In simple terms, Google Translate helps one understand the gist of the foreign language text and the example below emphasizes this point.
However, it is no secret that Google has been trying to incorporate Machine Learning and Artificial Intelligence techniques into all its tools. In September this year, Google unveiled the Google Neural Machine Translation (GNMT) System, the company’s most public demonstration of its experiments with Machine Learning. GNMT uses Neural-Machine Translation instead of the conventional phrase-based and word-based translation mechanisms for 16 language pairs. Phrase-based and word-based translation machines work by translating a word or phrase from one language to another. In this approach, the quality of translations is low as the translator does not take the context of the sentence into account.
In a complex approach, a neural-machine considers the whole sentence as an entity while taking care of the words and phrases that the sentence contains. To make this possible, the translator is first trained by showing it millions of examples of translations for every language pair. One major pitfall of traditional translation systems is the difficulty in translating rare or uncommon words from one language to another. GNMT tries to surpass this roadblock by teaching the neural network to break down uncommon words into smaller parts using principles of etymology and then translating these words into other languages.
Another important feature of the GNMT framework is Zero-Shot Translation. Zero-Shot translation refers to the ability to translate between language pairs that the system has never seen before. The Google Research Blog explains it by using a simple example. To enable GNMT to translate between English-Korean and English-Japanese language pairs, it is trained using millions of examples of accurate translations as demonstrated in the animation. This means that the “translation knowledge” can be shared between other language pairs which use the same system.
The knowledge shared prompts an obvious question – Is it possible to use the same system for a language pair it has never seen before? or Can we translate between Korean and Japanese even though the system has not been explicitly trained for this pair? As iterated by the above example, the answer is yes. GNMT is able to produce reasonable translations for language pairs that the system has never seen in training. As shown by the animation, during training, the framework is trained by showing it many examples of translations between English-Japanese and English-Korean pairs. Yet, the system is able to generate reasonable translations for the Japanese-Korean pair as well .This is possible because all the language pairs use the same neural network, depicted by the box in the animation.
Further research into the working mechanism of GNMT suggests that the translator uses the logic and meaning of phrases as well as words to encode them in an ‘interlingua’ that cannot be read or understood by us. A simplification of this observation is that GNMT has created an internal language which it uses for translations. In phrase-based and word-based systems, the translator remembers the translation of many different words and phrases. Contrary to this, GNMT uses a common language to remember words on the basis of their meaning with the help of which the translator can translate between language pairs that it has never seen before.
GNMT is a giant leap forward because it achieves near-human translations for about 16 language pairs including Chinese, French, and Spanish. By now, it is fairly evident that GNMT is a remarkable breakthrough in the field of Artificial Intelligence and this system is here to stay. Further research into the working mechanism of this framework and the Interlingua that GNMT uses will open the doors to many useful innovations in the near future. There are many more advances from a technical standpoint that one can read more about in this page describing GNMT.