Neural Machine Translation – the newest technology explained

Neural Machine Translation is the latest development in Machine Translation – what does this technology mean for the future of the translation industry and what are the best implementation practices?


With the growing importance of global communications and international business relations, new and more efficient translation methods have been developed. The latest technological advancements have brought more opportunities for automated language translation. These innovative methods work towards eliminating the language barriers businesses oftentimes face when communicating with foreign partners. Machine Translation (MT), developed in the 1950s, has revolutionised the translation industry, sparking a debate around “What’s next for human translation services?”. Are we looking at a future where translation becomes completely automated? Can the human cognition and cultural capital be completely replaced by lines of code and a sequence model for predicting the meaning of text?



Most recently, a new approach to MT has emerged- Neural Machine Translation (NMT), which uses Deep Learning to translate text. The new technology not only uses a fraction of the memory needed by its predecessor, Statistical Machine Translation (SMT), but it is also modelled around the neural frameworks of the human brain. By employing both deep learning and representation learning, NMT allows for contextually precise and fast translation of whole sentences, reducing translation errors by 60%. Neural Machine Translation has already been adopted by the best MT service providers, including Microsoft, Google and Yandex. These technological giants have employed the NMT technology in order to improve the functionality of their translation services when in offline mode. The goal is to move away from the phrase-based offline translations and provide more accurate interpretations in more languages.



Microsoft Research has implemented the NMT technology for the autonomous translation of news articles from Chinese to English. What’s innovative about this approach is that it uses tens of millions of parallel sentences from the news domain as training data. While still a great success, this method proves that machine translation systems require a huge amount of data and resources to be trained in recognising different languages. Microsoft have tackled this problem by using a Semi-Supervised Universal Neural Machine Translation approach, which can produce high-quality translation of a low-resource language. The ultimate goal is to utilise NMT for supporting low-resource languages, as well as spoken dialects.

An important issue which can be resolved with Neural Machine Translation is the interpretation of low-resource languages. Facebook, which requires over 4.5 billion translations a day, has come up with a way to tackle this problem using NMT. The company has developed a fully unsupervised system established on Phrase-based and Neural Machine Translation, to deal with low-resource languages lacking parallel corpora for training machine translation. Neural Machine Translation eliminated the need for having both the source and the target data of the two languages. Therefore, translations can be executed with less data on hand and at a lower cost.


It is likely to see the increased implementation of NMT within various translation apps and services. The technology is able to translate text from grammatically complex languages, by learning and utilising their specifics. Unlike Statistical Machine Translation, Neural Machine Translation is considered more effective in handling word ordering, morphology and syntax.

The technology comes with the promise of cost-effective translation for under-resourced languages, which makes it beneficial for all businesses, regardless of their location or spoken language.



Despite the great opportunities for automation Neural Machine Translation offers, the technology has not reached the point where it can completely replace human translation. While a good way to interpret documentation and various texts in a cost-effective way, MT’s margin of error is still too great to be trusted with external translations. Currently, the best practice for translation companies looking to streamline their operations is to use Neural Machine Translation as an addition to their experts’ operations, allowing human translators to focus on the more complex, high-value projects.  Only time will tell whether the advanced technologies will make translation completely autonomous from human input. Until then, it is best not to leave your important correspondence and international business communications in the hands of online translation apps. When it comes to your Asian language needs, allow 1-StopAsia’s to be your one-stop source for translation, where our language experts will help you maintain seamless and efficient communication no matter the industry and country you operate in.

Author: 1-StopAsia

SHARE THIS PAGEShare on FacebookTweet about this on TwitterShare on LinkedInShare on Google+

Sorry, the comment form is closed at this time.