Machine Translation Is More Honest in Customer Reviews, Says GetYourGuide

E-commerce platform GetYourGuide offers packaged tourist experiences in 14 languages. We spoke to the company’s Head of Content Projects and Innovation, Anne-Cécile Dousson-Lhéritier, who says they sell over 30,000 tours and activities, “and we are growing! Each of these products has unique content that describes the experience. We, therefore, deal with a high volume of translations — probably around 1–2 million words per language per year.”

While they do not manually translate customer reviews, all other content requires professional translation. GetYourGuide’s job as content curators for their suppliers entails editing content to fit the site’s tone; and then translating it.

As for customer reviews, Dousson-Lhéritier points out how “machine translation gives, in an odd way, a more honest feeling. Having a raw review in the native language of the customer, which is not curated by rewriting or translation, has more value in terms of customer experience.”

All About Context

As an e-commerce platform, GetYourGuide needs to continuously deliver accurate content in all languages at all times for the 30,000 activities it has available. Their mission is simple: to make sure customers enjoy their experience at their destination. So translation quality is connected to customer satisfaction; that is, how much they buy and how satisfied they are after the purchase.

“For customer reviews, machine translation gives, in an odd way, a more honest feeling” — Anne-Cécile Dousson-Lhéritier, Head of Content Projects and Innovation, GetYourGuide

“We, of course, care about the linguistic side of things, but the scorecard that matters most is the satisfaction of our customer,” Dousson-Lhéritier says.

According to GetYourGuide’s Head of Content, the main challenge of working with high visibility marketing content has to do with the fluency of the translated text, consistency for the entire inventory, and accuracy despite frequent changes to small parts of content.

However, they do work with different levels of quality, Dousson-Lhéritier says. She recalls the 2016 Paris flooding when they had to display the banner in English on a different locale for a few hours.

“It was more important to get the message out quickly rather than wait for a perfect translation and leave our customers uninformed. On the other hand, if we send out a newsletter, we expect perfect accuracy, language, and fluency — this is content we cannot modify on the fly,” she says.

Teams at GetYourGuide are set up by tasks or specialties rather than by language, according to Dousson-Lhéritier. To ensure efficient communication and a smooth transition from task to task, they rely on a solid workflow.

She explains, “We can manage and adapt for content at scale. We have one process for the entire chain — from source to translation — and everyone knows their role and how their actions impact the rest of the team (i.e., how source content editing impacts the translation into a particular locale). The feedback loop inside the content team is fast and we can iterate fast as well.”

“If we send out a newsletter, we expect perfect accuracy, language, and fluency”

Dousson-Lhéritier points out that being able to control the source content is a big advantage that avoids back and forth and frustration on both sides.

She adds that they constantly train their translators and editors as quality issues occur, typically, due to a lack of understanding of the context. Dousson-Lhéritier says training the team and giving them more insights into the context of the content they produce increases overall quality.

Asked what happens when translation quality falls, Dousson-Lhéritier replies, “We feel it at the heart of our business — customer reactions.” She says bad quality leads to a lower conversion rate and a lower NPS (net promoter score).

“On the operational side, it affects efficiency and processes. If a piece of content is not edited or not translated to the quality level we expect, all the different steps that happen afterward will be delayed and become more time-consuming. It will, therefore, create some resource shortage,” she says.

“Bad translation quality leads to a lower conversion rate and a lower NPS (net promoter score)”

Major Step Forward

Dousson-Lhéritier, who has been closely following developments in machine translation, is convinced neural and auto-adaptive MTs are “a major step forward” for the industry. She says, “They provide a great long-term solution for enhanced productivity and improvement of baseline quality. I see also a lot of competition in this area, which is always good for the clients: more innovation to come!”

GetYourGuide has been an early-adopter in the use of adaptive machine translation by using Lilt and “the in-house team also works a lot with direct vendors for most of our translations,” she says.

As for the rest of the tech stack, they use a custom-made CMS, “created by GYG for GYG,” which allows them to publish content as it is produced. Dousson-Lhéritier says, “For now, all the quality checking process is rather manual.

We do not have a centralized automated system for this part yet. We have a lot of opportunities ahead, to improve and optimize our efficiency, tooling, and output in terms of quality and quantity.”