In this special guest feature, Sirena Rubinoff, Content Manager at Morningside Translations, discusses how Neural Machine Translation (NMT), although still relatively new, is quickly transforming into a robust platform for translations. Hardware advances will further contribute to neural networks overall, and NMT in particular. Sirena Rubinoff is the Content Manager at Morningside Translations. She earned her B.A. and Master’s Degree from the Medill School of Journalism at Northwestern. After completing her graduate degree, Sirena won an international fellowship as a Rotary Cultural Ambassador to Jerusalem. Sirena covers topics related to software and website localization, global business solutions, and the translation industry as a whole.
Years of research and development backed by consistent funding has allowed machine learning and artificial intelligence to make its mark on machine translations. These technologies are starting to be applied to new industries and purposes, such as website translations. Over the past year, several important developments and initiatives announced by Google, Facebook, and Microsoft regarding their use of artificial intelligence (AI) have been made:
- On April 28, 2016, Google’s US patent application “Neural Machine Translation Systems With Rare Word Processing” was published. Google claimed exclusive rights to an NMT system made specific by how the system is implemented and then, two weeks later released the code for its language parser free to the public.
- Facebook, which supplies two billion translations each day, published results in May 2017 on their approach for neural machine translation (NMT) using convolutional neural networks, said to be up to nine times faster than sequential reading methods, beating Google’s system.
- In November of 2016, Microsoft announced they were using neural networks to power all speech translation, a further development after their February news that they were leveraging AI and deep learning for their Translator Hub.
What is Neural Machine Translation?
NMT is a new approach in which a single, large neural network is trained, maximizing translation performance. Neural networks are a deep learning technology that simulates the interconnected billions of neuron cells in the brain that help us learn, recognize patterns, solve problems and make decisions.
Traditional NMT methods read a sentence word by word, and remember what the sentence meant up to that point. Facebook’s convolutional method allows NMT to be applied to multiple parts of a sentence at the same time.
The convolutional method is inspired by an animal’s visual cortex. Units respond to stimuli in the receptive field. Each of these fields slightly overlap and are variations of multilayer perceptrons designed to use minimal amounts of pre-processing. The computer is then trained to give meaning to parts of a sentence and process the sentence with each of the ‘fields’.
Through processing the sentence part simultaneously with these overlapping fields, the computer gets a notion of what every part of the sentence means. A different neural network turns this representation of meaning back into another language.
The traditional, fully-connected AI neural network consists of inputs, hidden processing units, and the output. All the units are interconnected in layers on each side and assigned a ‘weight’ to indicate the strength of the connection.
Inputs work from left to right, and are trained by comparing the results of the output with the desired results. Then, they use the difference between these two to modify the connection weights between the network’s units.
NMT systems are more adaptive and complex than trained models. Once the initial hurdle of setup and training is complete, a translation neural network has contextual awareness and deep learning capabilities that it can build on to understand the place in which a conversation is taking place and the vernacular of the environment. Therefore, it leads to an overall higher quality translation
The Future of NMT
Although still relatively new, NMT is quickly transforming into a robust platform for translations. Hardware advances will further contribute to neural networks overall, and NMT in particular. Google is already cooperating with start-up Nervana to design and build a custom ASIC processor for neural networks and machine learning applications that will enable up to 10 times faster processing than the typically used graphic processor units (GPU) for neural network computations.
As NMT continues to develop, human translators will have more powerful tools to help in their mission to reach and maintain the highest levels of quality for the vast quantities of translated content. Even in translating for ecommerce success, NMTs can make a large impact as their reach extends into new and exciting territory. The scope of the platform’s influence is massive, with the potential to affect any industry in which internationalization is prevalent (meaning, almost all of them).
Sign up for the free insideBIGDATA newsletter.
Useful article for neural machine translation.
But we can’t trust that machine translation will be 100% accurate.
A good introductory article. However, there are still many challenges involved in getting a neural machine translation system doing useful work in the translation office. I have one production system but it still relies on automatic pre- and post-processing to achieve a result that’s acceptable to the end user. The rare & unknown words issue isn’t completely solved by Byte Pair Encoding.