The Robots Are Coming – But Not for Me
“What do you do for a living?” Small talk is so predictable – it’s as if we have a script for it programmed into us, and the question I opened with is on the first page. When people learn that I’m a translator, the focus of the conversation commonly shifts to machine translation and whether I’m concerned about the future of my career. Machine translation, or MT, refers to programs that translate from one language to another (Google Translate is a well-known example).
The short answer: No, machine translation doesn’t pose a threat to my field.
The long answer: No, machine translation doesn’t pose a threat to me – to my specific business entity. While today’s MT output is immeasurably better than that of its early days, it’s still a far cry from translations that are fully accurate and employ appropriate style. And until literal artificial intelligence exists (it does not as of this writing, even though the media would have the masses believing so), qualified human translators will always deliver a better result than MT. Translators who find MT output on the whole to be flawless and fit for purpose are – brace for brusqueness – not qualified for the job.
But what could a [qualified] human possibly do that MT can’t? It’s just words, right?! No – translation is so much deeper than “just words!” Only a human has the capacity to perform two core talents that are indispensable for a proper translation:
Comprehend what is meant by the source text (what’s being translated)
Convey the source text in another language such that the readers of the target text (the finished product in another language) will understand the intended meaning and tone
The human brain brings together a whole world of experiences to accomplish these tasks. MT eats the words you feed it, runs it through an algorithm where all the other words it’s eaten are swimming around, and spits out what its formula says are the equivalent phrases. It can neither comprehend nor convey. It just eats and spits.
All that eating and spitting, with no thought behind it, leads to a handful of stereotyped errors that only a human translator can identify and rectify. I cover seven of the main types of these errors below, followed by a description of several more aspects of translation that MT just can’t contend with.
Error category 1: Localization
Localization entails accounting for the culture in the country where the target language is spoken. Although it may be surprising, this category even applies to domains that are traditionally thought of as less creative in nature. In other words, even technical texts need a human translator in order to preclude localization errors.
| Machine translation ✘ |
|---|
| Our COO is our gentle silverback who keeps our operations running smoothly. |
| Human translation ✔ |
| Our COO is our gentle and powerful leader who keeps our operations running smoothly. |
The problem: Silverback could be understood as a racist term in U.S. culture.
The takeaway: In addition to trusting your text to a human translator, remember that it’s critical to work with a translator who is a native speaker in the specific target culture (the term in question is acceptable in other English-speaking regions, such as the UK).
| Machine translation ✘ |
|---|
| 18x95x2500mm (note: MT simply used the source text as the “translation”) |
| Human translation ✔ |
| 98 7/16 in. × 11/16 in. × 3 3/4 in. (L × D × H) |
The problem: MT doesn’t know anything about localization of numbers. What units are used in the country of the target language? Do you want your text to include multiple units of measure, such as °F and °C? Should converted numbers include tenths or be rounded to whole numbers?
The takeaway: Your translator can help you figure out how to handle number conversions. The text this example came from was for trims and moldings. The client first tried MT for his product descriptions and then contacted a translator after seeing that the measurements were an issue. Which leads to another takeaway...
Bonus takeaway: Here’s what the first translator delivered: 45/64 3 47/64 98 27/64 inches. Setting aside that there are no symbols between the numbers, these values are just plain batty. No homeowner is going to try to figure out if a trim height of 3 47/64” is what they need – they’re going to click on the next listing in the hopes it doesn’t make their head spin. Again, hire a translator based in the country you’re marketing to – and one with experience in the relevant industry (see The SME Translator: Why Industry Experience Is Critical for more on this topic).
| Machine translation ✘ |
|---|
| German has a lot of words for money: Monets, beeps, mice, moss, gravel, toads, coal, gravel... |
| Human translation ✔ |
| We have a ton of words for money: bucks, dough, cheddar, lettuce, moolah, green, bacon, cash... |
The problem: There are several problems with the MT output of this text about saving and spending money wisely: 1) “German [having] a lot of words for money” has zero relevance to the English-speaking reader. 2) The words the MT program spit out are not synonymous with money in English. 3) Two words from the original list translate to gravel – did you catch that duplication?
The takeaway: MT cannot comprehend culture.
Error category 2: Context
Context is everything: A well-traveled phrase, and perhaps more pertinent to translation than any other…context! Take any word from the previous table as an example. Is a sentence with the word green referring to money? A can of paint? A place to practice your golf swing?
| Machine translation ✘ |
|---|
| Insert an endoscope camera through the gap. |
| Human translation ✔ |
| Insert a borescope into the gap. |
The problem: The source language for this example uses one word for endoscope and borescope. A borescope is a tube with a camera on the end used for inspecting tight spaces in machinery. An endoscope is similar in concept, but it goes in different types of spaces. I’ll give you a hint: This device is used in the medical field.
The takeaway: Once again, this example illustrates the importance of not blindly accepting MT output as well as working with a translator with credentials in the applicable industry.
Error category 3: Source text errors
If I translate 10 orders in a month, I’ll probably find an error in the source text of about 6 – 8 of them. Sometimes the error is as simplistic as a typo. Other times, it’s related to meaning, such as a procedure step that instructs a user to turn a machine on when it needs to be turned off. Unlike MT, I can catch these mistakes because I’m consciously following along with the content.
| Machine translation ✘ |
|---|
| We have cards for birthday, anniversary, mourning wedding and birth but also neutral versions. |
| Human translation ✔ |
| We have cards for birthdays, anniversaries, loss, weddings, and birth as well as blank versions. |
The problem: Although some may consider their wedding day to be one of mourning, a human can clearly deduce that this wasn’t the intended meaning.
The takeaway: Human translators ensure your translated text doesn’t have hilarious blunders such as these, and they’ll let you know about the error in the source text.
| Machine translation ✘ |
|---|
| A drop of blood is taken from a patient's finger via capillary juice until the capillary is completely filled. |
| Human translation ✔ |
| A drop of blood is collected from a patient's finger via capillary action until the capillary tube is completely filled. |
The problem: 1) Capillary juice?! Absolutely not. MT just didn’t “realize” there was a typo. 2) This example is a twofer: The second emboldened correction is an issue of context, as the source language uses one word for both capillary and capillary tube. The device being patented was not some magical device that could simultaneously draw and inject blood into your capillaries.
The takeaway: I think I can summarize it by now: specialist human translators.
Error category 4: Meaning
MT gives you a translation result in seconds. But the results are not always accurate, and if the meaning is wrong, what does speed matter?
| Machine translation ✘ |
|---|
| In the case of machines that are usually designed with a crosshead in the process gas area, there are internal leaks. |
| Human translation ✔ |
| Internal leakage is a concern in the gas processing sector, where machine design usually incorporates a crosshead. |
The problem: Look how convincing that first sentence is. Although it’s a bit awkward in terms of style, it seems quite plausible as an accurate translation – too bad the meaning has been butchered.
The takeaway: Running a text through MT and having a human translator review it is a subset of translation known as PEMT: post-editing of machine translation. It can be a perfectly acceptable workflow for certain texts and with the right professional, but the second step – review by a human – can’t be dispensed with in any case.
| Machine translation ✘ |
|---|
| MT program 1: Food loss that is screaming all over the world. |
| MT program 2: Food loss is being called for around the world. |
| Human translation ✔ |
| The issue of food waste is a dire one all around the world. |
The problem: After one MT program spat out an outlandish result, I checked another well-known program to see if it could do any better – evidently not.
The takeaway: Just keep in mind that if you trust your text to MT alone, the errors won’t always be so plain to see.
| Machine translation ✘ |
|---|
| This standard is stipulated with a policy of not taking responsibility. |
| Human translation ✔ |
| This standard is designed with the intent of not placing excessive burden on utility providers. |
The problem: The way the source text was written, the subject of the sentence was unclear and had to be inferred by the reader. MT is incapable of knowing this, so it simply spit out something grammatically acceptable – and hilarious.
The takeaway: This sentence was buried within the lengthy text of a company’s terms and conditions. Some companies will run a text through MT and have an employee who took a couple years of English in college give it a once-over. The result: Examples like this slip through.
Error category 5: Connotation
Denotation is the plain and literal meaning of a word. MT usually chooses words that are denotatively correct. Connotation is the “hidden,” somewhat emotional meaning behind a word. As you’ll see in the examples below, MT doesn’t have a great handle on the latter.
| Machine translation ✘ |
|---|
| Fire & Flames - your next career |
| Human translation ✔ |
| A spark and a flame – your next career |
The problem: Fire and flames?! That’s not how I want my next career to end up! MT can see in its big repository of text that this is a phrase used in English, but it can’t pick up on the connotative meaning.
The takeaway: This was the header for the job vacancies page of a company that produces fire monitoring systems. Hence, I wanted to keep the “theme” of the source text. But unlike MT, I used a phrase that has a connotation of prosperity rather than damnation – the human translator can navigate the nuances of language for cases like this exactly.
| Machine translation ✘ |
|---|
| Our company guidelines, the so-called "Corporate Principles", serves as the basis for our corporate culture. |
| Human translation ✔ |
| Our company guidelines, known collectively as the Corporate Principles, set the foundation for our corporate culture. |
The problem: So-called carries a connotation of sarcasm. This MT output implies that whoever wrote this text thinks the company’s Corporate Principles are a joke!
The takeaway: This type of error, and indeed specifically the use of so-called, is commonly made by both MT as well as translators who are not native English speakers. Translation is done into your native language for precisely this reason. Never hire a translator who isn’t a native speaker of the target language.
Bonus takeaway: Corporate Principles was in quotes in the source text, so the MT program adopted this punctuation in the output. Professional translators know the rules of orthography, which define how a language is written (i.e., spelling and punctuation). MT doesn’t, which is why it was happy to retain the quotes. They don’t belong here in the target text and read as scare quotes, which make the phrase look doubly sarcastic.
Error category 6: Interpretation
Remember all that stuff about comprehending and conveying? Because MT inherently lacks these capabilities, it simply miscommunicates the message.
| Machine translation ✘ |
|---|
| Our equipment is also important in the laboratory - whether in university clinics or full-service homes. |
| Human translation ✔ |
| Our equipment is also important in the lab – whether at a teaching hospital or in a residence receiving skilled nursing services. |
The problem: University clinic isn’t an official term per se. But while there are cases in writing and translating when you have to fabricate a phrase, the image this one conjures up doesn’t even convey what is meant. Full-service homes is utter nonsense, and without a human to decipher the intended meaning, the sentence loses nearly all value.
The takeaway: Human translator, industry specialist, familiarity with the culture of the target language…same ol’ song and dance!
Error category 7: Style
Medical and technical are my specialty domains, and it is often proclaimed that style isn’t a component of the latter in particular. I reject this! Style is by nature a component of any text, even if the goal isn’t to be attention-getting or to evoke emotion. Besides, every domain has marketing materials (websites, press releases, etc.), and these must be handled by a human being with specialist knowledge and writing skills to pack the most punch.
| Machine translation ✘ |
|---|
| Lawn care gets even "neater" when your riding mower is equipped with our mechanical devices. |
| Human translation ✔ |
| Lawn care gets an extra dose of “care” when your riding mower is equipped with our mechanical devices. |
The problem: The source text incorporated a pun very similar to what you see in my translation. MT almost always neglects puns, as it doesn’t have the capability to detect or understand the play on words.
The takeaway: Human translators can detect humor and emotion in the source text and can often render these with the same “flair” as the source text. That means your text will be as appealing to buyers in the source and target language countries alike.
| Machine translation ✘ |
|---|
| In these areas, our products prove to be high-quality, "all-water" helpers on board. |
| Human translation ✔ |
| In these applications, our high-quality products take to the challenge like a duck takes to water. |
The problem: The source text incorporated an idiom that literally means washed with all waters and is commonly translated as to know every trick in the book. What the client was communicating: Their product is perfect for use in boating equipment, which is a specialized application area for the very common device being described. MT ditched the idiom altogether and used a translation that was equal parts peculiar and confusing. Conversely, I found an idiom pertinent to the subject matter to retain the style of the source text.
The takeaway: Just like in the previous example, a human translator can try to transfer the engaging nature of such phrases to the target text (or at least write something sensical, if translating a reference, rhyme, etc. would be like smashing the round peg into the square hole).
| Machine translation ✘ |
|---|
| Dependable equipment makes the difference between operational efficiency and failure or safety and danger in emergency situations. |
| Human translation ✔ |
| Dependable equipment makes the difference between safety and danger in an emergency and between operational efficiency and failure. |
The problem: What on earth is this MT output communicating… Disregarding the multiple possible interpretations of what the MT program spit out, even if the reader landed on the correct meaning, the phrasing is just plain unwieldy.
The takeaway: Style is a necessary consideration even in the driest of texts. Even if no puns, no jokes, no anything that would cause a smile to crack is involved, part of a translator’s job is to write in a style that’s unambiguous and easy to understand on first read (unless such is the intention of the text, but that’s a-whole-nother topic).
I spent hours paring down my list of bad-MT-output examples to form the expansive list above. This is noteworthy for two reasons:
It means I had to let go of several fantastically bad eggs just to whittle this posting down to a digestible length (and that’s a debatable evaluation).
I even left out certain error categories that I found less critical, such as MT’s difficulty with abbreviations (it occasionally “guesses” them incorrectly) and transliterations (words spelled out phonetically because they don’t exist in the source language).
Plus, my examples neglect several aspects of translation that MT can’t manage because they relate to the full piece of text you’re having translated, or to all of your business’ texts combined:
Tone: For a grasp of this by way of example, consider which of the statements below suits your website:
a. Please browse our selection and contact us if you have questions.
b. Hey there! Take a look around and just drop us a line if you need us!
Consistency: Professional translators will reference your existing texts to make sure terminology and phrasing are consistent. This way, your customers can enjoy a seamless experience across your entire brand. Your translator will also make sure terminology is consistent throughout the individual text you’ve hired them to translate – MT uses surprisingly inconsistent terminology within a single text.
Formatting: Do you want each item in your bulleted lists to end in a period? Are Document Headers Written in Title Case, or are Document headers written in sentence case? What is your company’s standard date format for the target language: MMDDYYYY, DDMMYYYY, or Month DD, YYYY?
Oh, and I haven’t even touched on the issue of data security. Ready to run your patent draft through DeepL? Then you’d better be ready for the repercussions of effectively releasing your patent online.
Because MT lacks the capacity to comprehend and to convey, all the while drawing on a lifetime of experiences and not just a warehouse of words, it simply never will replace the human brain. So no, I’m not worried about job security. Especially because half the customers I “lose” to MT will eventually catch wind of the problems in their target texts – they’ll be contacting me to help them gain their image and customers back sooner rather than later.
Updated: September 3, 2023