Since 2017

Strictly Spanish Blog

Neural Machine Translations: Are They Really Replacements for Human Translators?

In previous blog articles, we have discussed the detriments of over-reliance on statistical machine translations, such as Google Translate, exploring how their phrase recognition systems often lead to mistranslations and altered messages. However, a new advancement in machine translation, neural machine translation, has gained prominence, research, and major advocates in increasing numbers. After comparing the two models and mentioning advocates for neural machine translation, we propose why human translators still remain viable in the wake of this new translation technology.

Differences between Neural and Statistical Machine Translations; Interested Parties

With statistical machine translations, data is drawn from both the source language and the target language and is used to guess or predict what translation would be appropriate to use. This guess can be word, phrase, or syntax based, but the element of uncertainty is still present. While the machine will make its best judgement, it could very well be the wrong one in a certain context of the translation. 

With neural machine translations (NMT), data is gathered and processed through a system that resembles the human brain’s neural network, consisting of several layers that a translation must go through before the final translation. NMT also exhibits deep learning, meaning it will teach itself linguistic rules and patterns through usage, creating a translation memory to better predict the translation.  Businesses conducting research into this fields development are Facebook, Amazon, and Google, all three with a global presence that must be addressed on an efficient scale.

Human Translators’ Viability

However, human translators still hold a valuable place in the translation industry through their assurance of quality and assurance of knowledge. While both statistical and neural machine translations provide instantaneous results, many companies still implement a human translator for post-editing work, knowing that quality may still lack that a human element can imbue into a project. Furthermore, both statistical and neural machine translations make predictions, guesses as to what would be the best translation based on the context, while human translators know the best contextual translation. They can analyze not only singular sentences or phrases, but take into account the entire passage, ensuring coherency across the entire document. This knowledge guarantees a polished product, one that businesses can easily distribute internally or externally with the assurance of quality and accuracy.  

Sara Leonhartsberger