Why Automated Translation Platforms Cannot Fully Replace Humans
With automated translation platforms changing the way information is spread and allowing for global interaction, there is no doubt that they play an integral role in developing cross-country communications, especially as they continue to improve. Even so, can we count on a future in which automated translation platforms completely replace human translators and linguists?
Only recently did online translation tools evolve to acceptable levels of competency.
Within the last few months, free online translators have improved tenfold. In November of 2016, Google revealed a new version of Translate that employs a translation engine called Google Neural Machine Translation (GNMT). This system translates complete sentences using an artificial neural network. It links digital “neurons” in several layers, each one feeding its output to the next layer—a method loosely modeled after the human brain.
Neural translation systems are first “trained” by huge volumes of human-translated text. The system then takes each word and uses the surrounding context to turn it into an abstract digital representation. Next, it tries to find the closest matching representation in the target language, based on what it “learned” before. This neural translation system handles long sentences much better than previous versions.
In contrast, older systems commonly used a piece-by-piece method (“phrase-based”). This method would translate phrases separately and then piece them together to produce incomprehensible, nonsensical translations that did not tap into the context behind each word.
However, the number of languages able to be translated using the new system is limited.
The new Google Translate began by translating eight European languages to and from English and then expanded to Chinese, Arabic, Hebrew, Russian, and Vietnamese. Later, Google extended neural translation to include nine Indian languages. Other translation platforms, such as Microsoft, also have a neural system for several difficult languages. However, the list of languages that can be translated using this new method is far from extensive. It will be a while before neural translation is able to translate back and forth between multiple languages at ease.
Even the most advanced translation systems often produce incoherent, nonsensical translations.
There are still weaknesses to be addressed with the neural translation system. For example, it has no way of knowing which words are proper nouns and therefore should not be translated piece-by-piece. At the same time, neural translation systems still produce odd transliterations for many phrases they’re not familiar with.
Sometimes, neural translation systems produce translations that are mysterious and illogical. For example, users recently found that typing variations of the Latin text placeholder “Lorem Ipsum” into Google Translate yielded random English phrases. In all lowercase, “lorem ipsum” produced “China.” Capitalized correctly (“Lorem Ipsum”) became “NATO.” In all lowercase “lorem lorem” was “China’s Internet,” and so on. Internet users were perplexed by these cryptic messages, and soon after, Google had to correct these erroneous translations.
Therefore, neural translation systems aren’t ready to replace human translators any time soon.
Literature requires a far too robust understanding of the author’s intensions and culture for machines to suffice. For critical translations that are technical, financial, or legal, even the smallest errors can have catastrophic consequences. In any case, a human will need to run through the final product to vet and revise the output of automated translators.
Additional Resources:
https://motherboard.vice.com/en_us/article/the-search-for-meaning-in-a-cryptic-google-translation
