Applying neural machine networks to machine translation

Machine translation is about to turn a corner, with a new emerging technology to support its advancement.

Development on machine translation began in the 1950s with the first computers referencing a database containing linguistic rules.  A heavy workload was required to ‘programme’ dictionaries and apply grammatical rules while the output was frequently disappointing.

Machine translation really started to develop a decade ago with the creation of translation engines using a highly statistical approach.  The dictionnaries and linguistic rules were replaced by complex mathematical algorithms which could select from millions of phrases across several languages.  The machine learning model further refined automated translation through the creation of automated translation tools which were very finely tuned to accurately translate particular types of text or documents because the engines could be ‘fed and trained’ with specific corpus.

Logically, the continued improvement in machine learning output has focused on creating a ‘hybrid’ approach by including linguistic rules together with the statistical algorithms.  Current machine translation uses algorithms which work by ‘word’, matching the previous and following words in a phrase to select the best possible translation in this context while the use of linguistic rules allows the correction of grammatical and other errors. This allows the machine translation to produce good results though it remains very visible that the text has been machine translated due to the presence of certain grammatical errors because of the ‘word by word’ approach, even if the context is taken into consideration.

Neural machine networks open up the possibility of approaching machine translation in a different way. These networks can translate by entire phrase where possible as opposed to word by word.  This approach can be more effective for Asian and languages such as German where the words can be in a very different order to English or French.  The entire phrase is translated which makes it much more coherent and therefore more ‘human’ and no longer reads like a machine translation.

While neural machine networks are compared to the human brain, there are certain keys differences: they don’t highlight what they don’t know!  So, for example when a phrase includes a word that the machine is not familiar with, a hybrid machine translation engine will leave the word aside and assume that it might be a noun, therefore indicating to the post editor that there is a problem with the phrase.  A neural machine translation engine however, might ignore the word in order to translate the entire phrase.  In this case, the post editor’s task will be to to pay attention to the meaning of the phrase, to ensure the translated phrase has the same meaning as the original one whereas with the hybrid engines, the job was more about the correction of sentence structure and grammatical mistakes.

Technology is evolving rapidly and artificial intelligence will dramatically change how translation is performed in the future. Like automated cars, you will still need an experienced set of hands behind the wheel but automated machine translation, when specialised in a particular sector such as the finance sector, allows financial institutions to communicate faster and more efficiently with their clients and as well as meeting the increasing transparency requirements of the regulators, markets and investors.


Enjoy this blog? Please spread the word :)