Top Five Reasons Why Human Translators Out-perform Neural Machine Translation
by Mark J Williams
Technology has made our lives easier, with access to almost anything with just a few clicks on a laptop or mobile device. In the world of document translation, the same is true – Neural Machine Translation (NMT) is a technological advancement that has improved even from a few years ago. If you are not familiar with the term, Neural Machine Translation refers to machine translation that uses an artificial network to predict the likelihood of a sequence of words. NMT has a large database for French, Italian, German, and Spanish (often referred in the language industry as FIGS) and is a good solution for translating non-technical material such as basic user manuals.
But does NMT out-perform human translation? In my view, NMT is a very useful tool, but human translation is still the best option for accurate translation. Below are my top five reasons:
- In the legal field with litigation and / or arbitration there can be a significant amount of content requiring review. If the case in question has a cross-border component, and one of the parties is from an English-speaking country, then any non-English content requires translating. Working alongside eDiscovery solutions, NMT can quickly and effectively provide a gist of the content from which a lawyer can identify possible incriminating or supportive content. This content must then be sent to a human translator, for an absolutely accurate translation which captures colloquialisms and subtlety of meaning. While using the NMT tool improves speed and lowers cost, a human translator is vital for fluency and consistency, and to capture the true meaning.
- Neural Machine Translation is not a good solution for translating technical information, such as in the biomedical field. Because NMT has short-term memory loss, it cannot translate big blocks of text, so if the same sentence is in a block of text three times, it may be translated by the NMT three different ways. In the biomedical field, this could alter the diagnosis, therefore using a NMT tool would be far too risky.
- But even for short sentences such product descriptors, Neural Machine Translation may not be the best option. A large e-commerce retailer recently used NMT for all of its product descriptions in Sweden and the translations were so poor, they were actually offensive, causing an outcry on social media. The company had to have their products translated again by a human translator.
- The intricacies of language differences by region are the most important elements in providing error-free translation. For example, a car in Spain is a "coche," whereas in Latin America "coche" is a baby stroller - and sometimes word-for-word translations miss these nuances. It’s these subtleties that can completely change the meaning of a translation. For example, I recently read a news story about interpreters at the Summer Olympics. A few interpreters were having a conversation in the same language, but because their dialects were different, they had trouble understanding each other - so they switched to a completely different language mid-conversation! The importance of these intricacies is typically not a consideration when businesses or individuals need translation services but are essential elements for accuracy and fluency.
- Another important factor where NMT can’t outperform human translation is grammar. Other than short sentences, typically the sentence structure produced by NMT is poor. One glaring problem is that NMT mixes up pronouns. With the use of pronouns being the most recent cultural hot topic, this is another important factor – how will this cause confusion for certain personalized document translation?
It’s important to note that NMT isn’t replacing language service providers – in fact, it’s actually generating more content for translation. Many companies are going into their archives and having their records translated, using a combination of NMT and human translation as mentioned earlier. As technology continues to evolve, it will be interesting to see how NMT will further develop, and how this will affect the language access industry and those who use their services.
Working from London, England, Mark Williams is the CEO of New Language Capital, a language service provider (LSP) that includes ASLI Interpreting Solutions and TB Alliance. Mark has a Masters’ Degree in Asia Pacific Studies from the University of Leeds and has an extensive background, including leading LSPs and content management companies to exponential growth since the early 2000’s, providing strategic advice and industry insight. Mark is connected by technology to his global team.