The 5 Risks of Raw MT for Software Localization
November 6, 2019 by Sonja
What risks lie in wait for software companies who use unedited machine output – or “raw MT” – in their translations? Milengo investigates some of the typical pitfalls associated with this evolving technology.
Offering the ideal combination of low-cost translations and short throughput times, many software companies seem to consider machine translation (MT) a cure-all solution with negligible risks and side effects. But using machine translations without any human verification step can have far-reaching consequences for companies, negating any cost savings you gained from choosing this cheaper option. In this blog post, we’ve highlighted five problems you could face if you use this technology and show you how to use machine translation in a way that’s both responsible and effective.
5 problems that software companies who use raw MT may face
Terminology in your company’s content is translated haphazardly
Modern MT rests on the pillars of artificial intelligence and neural networks. While this enables it to produce impressively fluent translations, it still lacks the basic comprehension skills a human being would have.
Let’s take a look at one very simple, everyday example commonly found in company e-mails and letters: the German form of address “Sehr geehrte Damen und Herren.” A human translator would instantly recognize that the correct translation here is “To whom it may concern,” whereas a machine would revert to “Ladies and gentlemen” – making you sound much more like a ringmaster at the circus than a consummate professional. While the human translator is able to understand this information from the context or situation, this is where machine translation hits a brick wall.
But mistranslations of set terms and phrases don’t just result in amusement or bemusement – they can also land your company in hot water.
Machines are sometimes unable to consistently translate terminology that has several meanings – for example, “Gewährleistung” and “Garantie” and “Warranty” and “Guarantee” are often commonplace in simple boilerplate texts such as product instructions. Just as most native speakers don’t understand the difference between these terms, a machine is likely to translate them inconsistently, giving rise to tricky legal situations.
An expert linguist will be able to spot even the most minute errors and resolve them – thus saving your company a whole lot of trouble.
Your customers are misled or misinformed
While certain texts – like online documentation for complicated financial software – don’t have to be literary masterpieces, they do have to be correct in terms of content and phrased in a way that’s easy to understand. Mistranslations can detract from the customer experience, and in the worst case scenario, they can result in improper product use and possible liability issues for your organization.
Even the most advanced MT engines can do annoying things like leave out words. For example, imagine what could happen if a sentence in the instructions for a set of plug-in LED string lights was translated as “Connect the light to the line power” – accidentally omitting “NEVER.” In this case, the consequences could be fatal.
While modern translation engines may now be able to produce texts that are highly readable, they can also warp the original meaning. The reader is imbued with a false sense of trust in the text, as studies have shown that translations that “sound good” but are factually inaccurate tend to be better received by users than clumsily phrased translations that are actually correct. These deceptively good translations can result in customers becoming dissatisfied with your company – and if worse comes to worst, they may even start eyeing up your competitors.
Your company develops a bad reputation among your customers
While we’ve seen that MT is becoming increasingly intelligent, there’s one very specific type of intelligence that it can’t bring to the table: a human translator’s understanding of social norms and cultural differences, or in other words, what is and isn’t appropriate to say.
In this sense, it’s often helpful to view human translators as a testing ground for your company’s content. For example, information that you may think is acceptable to include in a whitepaper may in fact be considered taboo or tasteless in the target audience’s culture. The translator is able to flag this material to the client and suggest an alternative way of presenting the issue or rewriting it in a more “digestible” manner.
Always remember that unintentionally offensive, inappropriate, or culturally unacceptable content can paint companies in a rather unflattering light, and thus alienate even the most loyal customers.
Your brand loses its recognition value
Let’s pretend that you run a German app called “Schneeflocke Kids” that sells outdoor gear for children, and you’re looking to expand into the US market.
When localizing your app, you’ll need to bear in mind that fixed product names made up of real words won’t be spared by the MT engine. In fact, the likelihood is that it will recognize the word and translate it. If your company consists of an invented name and another word, then it will most likely translate it inconsistently.
Your software’s name could be a very common word in the target language, making it near-impossible to find online, confusing your costumers, and weakening your brand identity. Put simply - the danger of incorrectly translated product names should not be underestimated.
In the worst case scenario, your brand name could be translated into a word with drastically different connotations to the cool company name you had initially selected. If we take the example from above, “Schneeflocke Kids” would be translated as “Snowflake Kids” – turning up a lot of think-pieces on helicopter parenting and political correctness “gone mad” in a Google search instead of cool ski pants for ten-year-olds.
Your company’s public image is undermined
While machine translation has come on leaps and bounds, the fundamental errors that arose when it first came to prominence continue to plague raw MT output. Let’s take one age-old problem from the world of translation – words with multiple meanings.
Imagine you work for a German company that offers e-learning courses to third parties, and you’ve just sent a new module on tender procedures to be translated using MT. The sentence “In diesem Fall ist der Ausschuss des Erstangebots rechtswidrig” comes back translated as “In this case, the committee of the initial offer is unlawful” – when what you really want is something like “In this case, excluding the initial offer is unlawful.”
Here, the word for “exclusion” (Ausschluss) and “committee” (Ausschuss) differ by a single letter in German, and only a human translator would be able to notice this typo and correct it in the translation. A “committee” that hasn’t been mentioned anywhere else in the course suddenly pops up out of the blue, and if the third party client doesn’t pick up on the error and send a complaint, then the end user will be left somewhat perplexed.
Think about all the viral images you’ve seen of “hilarious translation fails” online – you don’t want your company to become the butt of your audience’s jokes. If bewildered customers doubt your company’s professionalism, then it’s sure to reflect poorly on your company’s public image.
Manage your MT wisely: keep humans in the loop
It should be clear by now that using machine translation without any human intervention can have some pretty dire consequences. In spite of all this, MT has captured the attention of more companies than ever before for cost-related reasons, as human translations and even conventional post-editing processes have simply become too expensive for large volumes of content. Text types like documentation, online help, or knowledge bases have since become typical application cases for MT.
Milengo has developed a new localization solution called “Managed MT” to help software companies avert this problem. Translations are produced using cutting-edge AI technology, before being sent to a human subject matter expert to correct any serious errors and polish the final text. All of our client’s previous translations are maintained in secure neural networks to continuously improve the quality of your future translations. An innovative quality assurance step specifically targets the weak points of machine translation that we highlighted above, from inconsistent terminology and incorrect product names, to serious mistranslations, right through to potentially offensive or culturally inappropriate content.
We’ve been intensively developing this solution and the initial figures look promising – our solution allows customers to save up to 80% in translation costs compared to conventional translation workflows.
We believe that this solution is the future of our industry. Software companies will finally be able to free up their localization budgets – and because all machine translated texts undergo a further in-depth post-editing process, you can rest easy knowing that your translations will meet your quality standards.