Can Neural Machine Translation Compete with Human Translation? (Part 2)
When it comes to translation, humans still outsmart machines. Learn how MotionPoint technology helps humans become even better website translators.
Editor’s Note: This is Part 2 in a 2-part series. Read Part 1 here.
As we reported on Wednesday, recent breakthroughs in a special technology called Neural Machine Translation (NMT) have generated lots of buzz within the translation industry, and beyond.
It’s also left some growing businesses wondering about the viability of machine translation—particularly when using it as a website translation solution to serve global customers online, in their preferred languages.
There’s lots to love about expanding into international markets, online. For most industries, such locales represent major marketing and revenue growth opportunities. And offering websites in these markets’ preferred languages—particularly to their growing middle-class consumers—is a very important key to success.
Sixty percent of global consumers who speak English as a second language rarely make purchases on English-language websites.
Indeed, about 60% of global consumers who can speak English as a second (or third) language rarely or never make purchases on English-language websites. Despite their fluency, they still prefer shopping in their native tongue.
With language playing such a mission-critical role in global digital marketing and commerce, the topic of machine translation is a noteworthy one. Do recent innovations in NMT-and extensive consumer use of conventional machine translation products-represent a newfound reason to use machine translation on global websites?
No. As the anecdotes we shared in Part 1 of our series revealed, machine translation remains notoriously unreliable and risky, especially for companies that place particular importance on word choice, wordplay and creative copywriting.
Further, any industry with high-sensitivity content-such as family finances, health care, or regulatory compliance-are particularly at risk of machine-translation missteps. Brands with distinct "voices" are also ill-served by machine translation.
Eric Frank, Global Account Director for MotionPoint’s sales team, explains further.
"Take one of our fast-fashion retail clients; it has a very strong brand voice," Eric says. "It's all about style, it's funky, it's for a young demographic. To connect with those consumers, you must literally speak the language and also be relevant, credible and cool.
Brands with distinct “voices” are ill-served by machine translation.
"But with machine translation, there's no personality, no spark," he continues. "When you read it, it sounds weird. Flat. Even if you don't know something is 'off' about it, you know there's something off about it."
Why Human Translation is Superior
We also spoke to Juan V. Ayala Millán, a MotionPoint Site Redesign Manager, about this topic. Juan and other MotionPoint linguists believe that until a computer can fully mimic the human mind, the results of machine translation will always underwhelm.
"We believe understanding the mechanics of language is nothing less than understanding the mechanics of human thought," Juan explains, "one that's been demonstrated to be dictated by an individual's mother tongue. A machine-created translation will only be as good as our way of replicating the complexities of the human brain."
This is a common position among linguists … and even among most evangelists for machine translation. It's widely known that machine translation often suffers from a lack of word-choice consistency. It omits words or mistranslates content outright. Other technical limitations abound. The technology simply isn't "there" yet, and everyone in the know knows it.
Machines can play chess, but they can’t write books or poems, which are actions that share many mechanisms with translators.
"In the past, we've seen machines play against-and even defeat-professional chess players and perform other amazing feats," Juan says. "But so far, we don't have machines writing books or poems, which are actions that share many mechanisms with translators. Even the most enthusiastic defenders of machine translation admit that it's nowhere as accurate and authentic as translations crafted by a human being."
Minimising Cost, Maximising Impact
There's one edge machine translation often has over human translation, and it has nothing to do with quality or readability. It's cost. We assert that the risks, time and expense associated with vetting machine-translated content far outweigh its benefits … but there's no denying that machine-translated content is often far cheaper (or free) to deploy on global websites.
Then again, you get what you pay for. Great, brand-worthy human translations are an investment. But MotionPoint goes the extra mile to minimise translation costs by developing innovative technologies and services.
For instance, our Segment Optimisation technology uses translations in new ways to reduce translation costs by 20% or more. And our exclusive Multi-Country Sequencer technology cleverly leverages translations to reduce the cost of entering new markets by 75% or more.
MotionPoint's Eric Frank explains another approach. Companies can craft shorter product descriptions (or other content) for their global-market websites. This can preserve brand voice, while proactively reducing costs (thanks to a reduced word count). Some MotionPoint customers have seen success with this approach, he says.
"We conducted a special A/B test with one of our retail customers to explore this," Eric says. "We examined the conversion rates for products that had translated descriptions that were about 75 words long, with products that had translated descriptions of about 25 words. The conversion rate wasn't affected. It was clear we, and our customers, could be smart about how to collaboratively reduce costs."
It's uncommon, but MotionPoint has recommended machine translation for some of its customers. These businesses often operate colossal e-commerce sites featuring millions of words of content. To radically reduce translation costs, we'll sometimes recommend machine translation for product titles, for instance.
MotionPoint goes the extra mile to minimise translation costs by developing innovative technologies and services.
“But even then,” Juan points out, “we’ll insist on a human to handle brand-related content, or more wordy and ‘trendy’ product descriptions that may require vast comprehension of the nuances in both the origin and target language.”
To further reduce costs, MotionPoint offers its customers two levels of human website translation. Our Enterprise offering features a multi-step translation, proofreading and revision process. Our Professional offering features less editorial oversight, and costs less. Both provide terrific translations.
Another respect in which human translators thoroughly outflank machine translation solutions is in the ability to apply critical thought—and use the best-possible translated words—on the fly, in production environments.
A perfect example of this is when translators tackle word growth. This common phenomenon, which often occurs when translating content into a Romance language such as Spanish, results in translations with more words or characters than the original English text. (Spanish content can be up to 30% longer than its English counterpart.)
This additional content can negatively impact the layout of web pages, or break page templates—templates that were originally designed for English content.
Human translators using MotionPoint’s platform can easily see how these translations might impact a live site, and adjust them in real-time to avoid problems. Machine translation can’t do that. These MotionPoint technologies provide human translators with contextual integrity, or a full, holistic view of a webpage’s content.
Human translators using MotionPoint’s platform can see how word growth impacts a live site, and can adjust their selections.
This visibility also greatly improves translations, Eric Frank says. With MotionPoint's approach, "translators can see the on-page imagery. They can see the paragraphs of text that appear before and after what they're currently translating. They're able to view the whole gist of the page," he says.
"Without that all-important context, you're simply translating string after string of text, and hoping it'll come together in the end. It usually doesn't, which nearly always results in expensive revisions."
We certainly live in exciting times, and the amazing technological breakthroughs seen in Neural Machine Translation are undeniably worthy of the buzz they’ve received in recent months.
However, all forms of machine translation still can't hold their own against a well-trained human translator. Humans can deduce creative intent, adjust translations for space constraints, and deliver brand-worthy translations in ways that machines may never be able to achieve.
As MotionPoint manager and linguist Juan V. Ayala Millán puts it, the simplistic conception of translation as an action of “replacing words” instead of “transporting ideas” is, in part, what makes some people believe we’re quite close of making machine translation as good as human translation.
“But because of the intrinsic and intricate nature of the language-thought relationship,” Juan says, “the day we can ensure a machine can translate as good as a human will be the day that we have completely understood and been able to mimic the whole of the human brain.”
And that certainly isn’t now—and won’t be for some time.