DeepL Panel Explores the Best Use Cases for Automated Translation Workflows

Deepl Panel Explores the Best Use Cases for Automated Translation Workflows

Business 101: Every company should understand their customers’ needs before offering solutions. 

This is especially true in a field like localization, in which customers — often laypeople — may feel overwhelmed by the hype surrounding the next big thing, or may become convinced that today’s headline holds the answer to their problems. 

Language AI powerhouse DeepL, inspired by a desire to hear from customers and better understand their market, surveyed more than 400 global marketers on a range of topics related to machine translation (MT), including MT adoption, MT initiatives, and pain points such as SEO

DeepL’s Customer Success Operations Director Lynn Nguyen shared the results of this survey at SlatorCon Remote March 2024. She was joined by Dierk Runne, Senior Manager of Localization and Systems at HubSpot, whose company was featured as one of two case studies in the survey (the other being Phrase).

According to Nguyen, 98% of respondents said they were using MT as part of a localization workflow, and 96% saw a positive ROI related to localization projects, making localization and translation a revenue driver for marketing departments. 

In addition to marketing, Nguyen and her team at DeepL see PR and corporate communications using similar localization functions. Customer support is also ripe for localization, as websites often default to English, while users typically prefer to submit reports and questions in their native (non-English) language.

Marketers are naturally focused on customers, “but there’s a lot of internal stuff […] that we could be using machine translation for,” Nguyen said, adding that MT is ideal for content users need for translation in real time. 

Of course, certain highly regulated industries, such as law and pharmaceuticals, require human involvement. In those cases, Nguyen said, the use of MT is intended to make life easier for human translators, rather than to replace translators. 

Why Not DIY?

Most of the companies surveyed rely on third-party solutions; for example, HubSpot does not build its own MT or large language models (LLMs). But as the technology continues to improve, and tools are increasingly open-sourced, some companies may be tempted to experiment with building their own. 

“You could probably DIY something, but it’s not going to give you the same result,” Nguyen explained. “You’re also then going to be tying yourself to a huge amount of technical debt” in terms of maintaining the product. 

Unlike many of the prospects considering the DIY approach, DeepL has its own product and engineering teams focused on improving user experience; a research team dedicated to building models; and significant investments in servers, infrastructure, and supercomputing power.

Runne agreed that experimenting would likely not pay off for non-experts, noting, “There might come a time when that’s more feasible. But at this point, I would much rather let the experts handle it.”

Even so, HubSpot has plenty of content to keep its in-house localization team of about 30, plus its freelancers, quite busy. The HubSpot platform is very content-driven, primarily with marketing material but also in other areas, such as education, services, and sales enablement. 

The volume of content, plus frequent brand “refreshes,” requires the company to update its terminology and ensure consistency across languages. 

DeepL’s translation capabilities are integrated in a range of tools, including Salesforce and ZenDesk, complementing HubSpot’s development of automation systems to identify and prep content that requires an updated translation.

Runne demonstrated how Mova, one of HubSpot’s internally developed tools, can create a blog post, publish the blog post on HubSpot software, and, with a few clicks, translate the text. This procedure can also be fully automated. 

Localization Priorities

HubSpot uses a tiered approach to determine the localization procedures needed for specific content. 

Specific performance metrics can indicate whether a human should review MT output. For example, high bounce rates could indicate that users are abandoning poorly translated pages. A long time spent on a certain page might point to engaging content, or to translations that take twice as long to read. 

It remains to be seen how multilingual content generation, still in its early days, may soon be integrated into suites of localization tools. Nguyen told SlatorCon attendees that DeepL customers are already asking for multilingual GenAI; DeepL Write is currently available in two English locales and in German. 

Runne believes that, as end-users become more familiar with machine-translated and machine-generated content, transparency will be key to widespread adoption and acceptance. (HubSpot typically provides disclaimers for its own machine-translated content.)

“We’ve all seen those news articles about a lawyer […] using ChatGPT to write an affidavit or something, and it falling apart completely,” Runne explained. “So being upfront and honest, […] I think this is going to be the way to go.”