How Zendesk Achieved a Targeted, Cost-Effective Localization Strategy Through Machine Translation

Machine translation in a business setting? It’s more common today than one might think.

More and more businesses are beginning to see the advantages of machine translation in their operations, but many are still on the fence or are unsure how to go about leveraging the advantages that MT provides. The question on their mind: how can they make the best use of MT in their operations?

In a recent podcast episode of Global Ambitions, Yoko Drain of Zendesk talked about the company’s localization strategy for its support pages, centered around the use of machine translation. This strategy makes for a good case study on the advantages of using machine translation today. In particular, we’d like to bring up three lessons to take away from this conversation.

Greater flexibility in determining funding

Right off the bat, Drain touches upon one of MT’s main advantages, which is cost. Machine translation is a much more cost-effective option than human translators or even machine translation post-editing (MTPE).

Zendesk’s strategy is to use MT to translate all its support pages, and further process high-traffic pages with MTPE to improve their quality. This is a very smart way of leveraging the cost advantage of machine translation. By prioritizing MTPE funding for higher-traffic pages, Zendesk is able to work within a reasonable budget. They are able to ensure that low-traffic pages still get necessary language support, and they free up resources for human editors on the pages used by more people.

It’s true that machine translation still has its limitations in terms of quality, as compared to post-edited work or the work of human translators. But again, budget is often also a limitation of translation work that entails human intervention, whether MTPE or pure human translation. It’s very interesting how Zendesk was able to overcome the limitations of both sides through the use of web analytics.

When we think of the use cases of web analytics, it’s often in a marketing context, whether the creation of consumer profiles or geotargeting. Zendesk’s strategy is brilliant in that they use the same data to determine how they prioritize their budget allotment for translation. MT works best in conjunction with other types of translation, and this strategy has allowed Zendesk to make use of both MT and MTPE in the best way possible.

With MT and web analytics together, the end result is that Zendesk has been able to save much-needed funds while maintaining quality service where it is needed.

Potential for continuous improvement

Zendesk uses an MT engine that has been specifically trained for their purposes, which is a good practice. Not all MT engines are built the same, or perform well under a different context.

An MT engine trained on data from the manufacturing sector, for example, would be familiar with the terms used in that industry, but would not be well-suited for, say, machine translations for military and defense. And an MT engine trained on generic data tends to perform less well for any industrial purpose than one that has been specifically trained. That’s why you don’t see companies using Google Translate.

The results are clear for Zendesk. Drain mentions a case in which a localization department approached her for help with improving their translation quality. They, too, used an MT engine, but the results were far from satisfactory. Apart from other factors, such as post-editing and customer feedback, what was made clear is that the MT engine used by Drain and her team was better trained.

But these aspects shouldn’t be taken in isolation from each other. Training an MT engine is a continuous process, and the more high-quality input is provided, the better the results are. Having a good customer feedback loop means that Zendesk is able to target pages that require human intervention. Post-editors contribute to the training process by correcting errors, and the results are fed back into the system. This in turn helps make the content of machine-translated pages better.

It’s a continuous cycle of improvement on all fronts, one which Zendesk seems to have mastered.

Creation of usable, actionable content

The last point is simple, although quite easy to overlook, so it needs to be said again: Today, machine translation is capable of generating usable content.

It’s a common refrain, that’s true, but machine translation engines really have come a long way. The days when MT was a novelty with amusing errors is long gone; the quality of machine translations has gone up enough to be useful in industrial and commercial contexts today.

The key difference between then and now boils down to one thing: neural machine translation (NMT). Neural machine translation is a form of MT that uses neural networks, which can process massive amounts of translation data with relative efficiency, increasing the quality of translations exponentially. Today, all MT engines use neural machine translation.

An MT engine that is well-trained in translating for a specific sector is capable of producing information that can generally be understood. In the context of the case at hand—localizing support pages—this is precisely the point.

The language on a support page doesn’t need to have the polish and style of, say, ad copy, after all. It simply needs to be understandable enough for users to know what to do without having to open a ticket for human assistance. MT today is more than capable of doing this, compared to the past decade, or even the past five years.

Zendesk is able to create support pages in a different language that are of adequate quality for use, largely through machine translation. That in itself is a remarkable feat, and makes a good case for allaying the fear of stakeholders who are still skeptical or on the fence when it comes to using machine translation in their own business.

Key Takeaways

Zendesk’s strategy shows how far machine translation has come in providing solutions for a business setting, and the unique and original use cases to which it could be put.

Other companies should take note of how well Zendesk has been able to integrate MT into their process for developing support pages, if not as a blueprint for their own use, then at least as a positive example of machine translation’s usefulness as an industrial tool.

In particular, Zendesk’s case has shown that machine translation has great potential to synergize with other technologies and organizational systems. This synergy has resulted in savings for the company while also achieving higher quality of service through more precise targeting.

All in all, machine translation is not just here to stay, but here to provide flexible solutions for businesses looking at multilingual expansion, and there are many other possible use cases for any business with the capacity to imagine them.