Have you ever plugged a sentence into a machine translation software and received a result that was so bad it blew your mind? The Bad Translator app can generate a perfect example of what can happen to a sentence when you run it through too many translations. I tried it out by inputting:
“I'd rather not deal with poorly translated sentences.”
Six translations later (from English to Bulgarian and back to English and into Korean, etc.), the translator spat out:
“I’d rather act with the translation of a sentence.”
Granted, there is no such thing as a perfect translation, and there’s a good reason for that. By its very nature, translation is a subjective activity. Language is fluid and open for interpretation. But we can take measures to get as many translations as accurate as possible.
There are various factors informing the quality of any given translation. One, of course, is skill in the language. Another is personal bias, which you can find almost everywhere, from theological opinions carrying over to Biblical translations to modifying news reports in a way that slants them differently than the source report.
Many biases and mistranslations are relatively harmless. But high-stakes agencies of the government can’t afford to misjudge their translations. If, for example, the government receives information about an emergency abroad, and there’s a mistranslation, relief efforts could be miscoordinated, and lives put at stake.
How Do We Assure Quality Translations?
Quality assurance is systemized in various layers of government. The FCC, which is in charge of all of the communication spectrum in the country, conducts ongoing tests to ensure that there’s no overlap between radio waves, TV stations, Wi-Fi, local broadcast and other frequencies. Based on those tests, they draw conclusions and put out new standards for each spectrum. In the same vein, the USDA has standards for food handling, the FAA for airplane safety and NIST for standards in the fields of science and technology.
Yet there are no cross-agency standards in translation, and there is no system of quality assurance for overall accurate translation. One reason is that translation is less objective than broadband or bacterial growth. Because language contains so many possibilities for interpretation, it is difficult to adopt black-and-white standards for the acceptance or rejection of content. There are, however, opportunities to use technology in novel ways to accomplish a different kind of quality assurance.
Standards for Translation
In order to create standards for translation, machine translation (MT) and humans must work together. Machine translation (MT) still doesn’t work well as a standalone technology. In order to generate accurate overall translations, human translators must be involved. Yet human translators have their own biases. You can’t yet trust a machine to generate the most accurate result. So how do you build real quality controls into your translation system?
In my experience, it helps to have several key components in place:
1. Measure translation quality. If you can’t measure a thing, chances are, you won’t have the foundation to improve it. In my work with the federal government, I’ve found that measurement is limited to the number of words translated each month. It would help greatly to know what’s trending, which people focused on what content, or what the volume of translations was in each area of focus. Such measurements help facilitate a better understanding of the process of their translations—and help users know what to improve.
2. Build a language model. A language model is anything that represents the properties of a language. Language models are ubiquitous, from Microsoft Word’s grammar checker to autocomplete on your smartphone. Building a language model for translation could involve identifying translators who consistently generate quality results, throwing their work into a big data bucket, and counting how many times a certain word or phrase translates into another. That way, as translators start to type, the translation platform can insert predictive text completion—autocomplete—for translations, offering suggestions based on the most likely next work or phrase to appear.
3. Use the cloud. Most ingrained translation management systems have no way to centralize data and run translation as a big data problem, because they’re not on the cloud. Without the cloud, it becomes difficult to measure and build language models, because there’s simply not enough data for the system to decipher quality results. If the government could focus on cloud-based translations, quality assurance and integrating standards would become a much easier process. Siloes of translation measurement could also be shared more easily in the cloud. If one agency has a solid set of standards for translating terminology, they could use the cloud to share it with another agency.
Once these pieces are in place, technology can be adapted to enable translators to, in aggregate, produce a higher-quality result. Bias and lack of skills will never be completely eliminated, but with the right technology, we can certainly control for them.