If you’ve been keeping up with localization technology, you’ve heard about the “revolution” in machine translation (MT). Some industry pundits will tell you that MT can translate anything—fast and for cheap. Others say jobs are in jeopardy because MT’s capabilities have reached parity with human translators.
Our advice? Don’t fall for the hype. Let’s look at the reality.
Yes, machine translation has improved in the last few years and become easily accessible to the masses through platforms like Google Translate. But it’s by no means 100% perfect. It’s still only one piece of the translation puzzle and often too risky to use without human involvement.
A machine may seem more cost-effective than a human, but alone, it’s not a blanket solution to all your translation needs. For example, you still may have to pay for wordsmithing and QA. And the biggest misconception of all is that MT can be used on any translation project—user manuals and marketing materials alike.
So, we’ve developed a six-step process to help you determine whether MT is right for your loc program before rushing to jump on the MT bandwagon.
Step one: Analyze content
The first step is to analyze the content’s purpose. You should be able to place each piece of content into one of four categories based on what we call the “scale of emotional weight”:
- Inform: This type of content has no emotional weight. Its purpose is to provide readers with quick information—sometimes user-generated. Examples: social posts and product reviews.
- Instruct: The content’s purpose is to provide explanation or instruction in plain, emotion-free language. Examples: product descriptions and FAQs.
- Interactive: The content contains user interface elements that lead users through an experience, like adding a product to a shopping cart. Examples: apps and e-stores.
- Inspire/influence: High-emotion marketing and sales content. Examples: websites and advertising.
The purpose of your content often maps to the way it was written: simple language versus highly branded, creative content. The more creative the content, the riskier it is to use machine translation.
For example, “inform” and “instruct” content could be prime material for MT because readers only need a general understanding of its meaning—though you still might need human editing depending on the translation quality you require. At “interactive,” human validation becomes critical to ensure users can navigate experiences correctly.
But if you want to really speak to your audience or provoke a visceral reaction, you’re better off transcreating your content or writing it directly in the target language. For your sensitive, branded content, we don’t recommend using MT at all—unless fast turnaround is more important than quality. (You could use a combination of machine translation, human validation and user acceptance testing to reach a high level of quality. But when you compare all those costs to that of straight human translation, how much are you really getting out of the machine?)
Step two: Define business goals
After sorting your content, decide where and how to apply machine translation. It’s not as simple as setting up MT for all your content in the first two categories above (inform and instruct). Your language service provider (LSP) will need you to describe specific business goals that can be fulfilled with MT.
For instance, you might want to translate only your high-volume content (like reviews or product descriptions), or only language pairs for which you have scarce resources. Perhaps you want to provide translated materials for underserved markets (or understaffed business units, like customer support), whether it’s perfect or not. Then again, if your primary goal is to reduce the strain on your call center, you may need a higher level of quality to ensure that there are no errors in the translations and that they are very easy to understand.
It’s important to define your specific business objectives for each type of content and apply the right machine translation technology—trained for a specific use and language pair—to the most appropriate situation. Even then, you don’t know if MT will work well until the pilot stage.
Step three: Conduct a pilot
Before running an MT pilot—testing how well a chosen MT engine handles translation for specific languages—you need to decide what you expect to get out of it. What does success look like in terms of translation quality, for example? Is success about speed? An LSP can help you figure out your criteria.
Then comes executing the pilot. Your LSP will choose which engines they want to try, run them on a test set of files (a sample of your content) and compare the results against what you consider a “good” human translation.
For example, you might test a bunch of English FAQ articles that have already been translated into German by humans and that meet your quality requirements. Your LSP will evaluate the outputs of various MT engines—collecting data along the way—to see which comes closest to the quality of the German translations.
Step four: Analyze the results
With the data collected in step three, your LSP can analyze which MT engines most help meet your business goals. A mix of human and machine-automated quality analysis can determine which engines, if any, meet your success criteria.
The engine that comes closest to the quality of your human-translated reference material will be the winner, but there is no one-size-fits-all solution—different MT engine outputs might “win” for different language pairs and content types; your eventual machine translation program might use several.
If none of the outputs meet your success criteria, perhaps you aren’t ready for MT yet—or as we prefer to think, MT isn’t ready for you yet because it’s not up to the level of quality you need. (And that’s OK—you have other options.)
Then again, MT may not necessarily be out of the question. You can run a few more tests—try different engines, content or languages. For example, languages with structures dissimilar to English (like Russian and Hungarian) often fail, while Spanish, Italian or French are more likely to produce better MT results.
And if your tests are a success…
Step five: Deploy your engines
…great! You’re ready to move on to deploying machine translation. That means choosing your MT engine(s), setting them up for the languages you want, building a machine translation post-editing (MTPE) program and training project managers on processes.
Now you have some decisions to make: how to integrate MT with your existing workflows and translation management systems, whether to deploy on-premise or in the cloud and how many review passes translations will need in each market, to name a few. Talk to an LSP about how to deploy MT—they can help develop the right roadmap.
Step six: Optimize, rinse, repeat
Even after choosing an MT system that meets your present needs, it may or may not evolve with your business growth or hold up long-term. It’s like buying a pair of shoes that seem to fit at first, but over time, you might notice discomfort or wear and tear and realize they’re not so perfect after all.
This is to say: despite what many companies want to believe, machine translation is not a sledgehammer. It’s a scalpel. Plus, the technology will continue to mature over time.
So, while we at RWS Moravia believe that every growing localization program should use technology to advance their goals of expanding language coverage, handling more volumes, reducing costs and improving quality, successful MT warrants continuous testing, analysis and optimization. We hope these steps can help you assess the technology’s effectiveness for your business, and feel free to come to us with questions about any stage of machine translation planning or deployment.