On the whole, Germany’s tech scene trails other countries’. In a country where politicians are still debating the merits of “digitalization” and venture capital is considered too risky, startups have generally had a harder time finding funding and getting off the ground.
But two German startups are global leaders in using artificial intelligence for language, competing with the tech giants — and even working with them.
You can think of Berlin startup Acrolinx’s product like the next level of spell check. The product offers suggestions as people write to eliminate passive verbs, shorten sentences and phrases that are too long, and ensure the document has the proper tone for the company.
While you’re writing, the Acrolinx tool quickly identifies what could be improved, such as eliminating the mind-numbing passive constructions of bureaucratic language, breaking out subordinate phrases, or removing superfluous words.
Acrolinx can be adapted to suit the needs of an individual corporation and its style rules, so that the thousands of documents produced by the company have a coherent style and tone. It can be formal or informal — a bank is likely to sound more proper than a sporting goods company. It can be updated if a company feels its language has grown old-fashioned.
Customers include a who’s who of digital innovators — Google, Facebook and Microsoft — as well as multinational corporations including Boeing and Nestlé. Acrolinx was spun off from the German Research Center for Artificial Intelligence back in 2002. GENUI Partners of Hamburg acquired Acrolinx for $60 million in 2017.
Acrolinx has mastered six languages: along with German, English, Japanese, Chinese, French and Swedish. Three-fourths of its business comes out of the US, so the company sees no need to add further languages at this time.
Nor is Acrolinx eager to move into translation. The program’s style improvements are already a big help in translation, as the example of SAP shows. The global software firm produces documents in German and English, but then must adapt them into 42 additional languages, employing a network of 100 translation agencies for the task.
“When the quality of the original text improves, then translation is much simpler,” says Ulrich Callmeier, chief technology officer for Acrolinx. In addition, it saves SAP time and money.
Another startup that is specializing in translation itself is Cologne-based DeepL, which grew out of the linguistic search engine Linguee. The search engine looked for translations of a word or phrase in various contexts, providing users with more nuanced options than most dictionaries. Linguee’s data from 10 billion search queries became the basis for the translation program, which was launched in 2017.
The US technology site TechCrunch found DeepL’s translations better than those from Google Translate or Bing. The German company has “trained” its algorithms with especially good translations in ways it keeps proprietary. It uses a deep learning technology based on neural networks to get its results.
Nor is DeepL looking over its shoulder to see if the US competitors are catching up. “We are confident we can keep our lead,” says technology chief Jaroslaw Kutylowski.
DeepL just got a fresh injection of capital from Benchmark, a US venture capital firm that was an early-stage investor in eBay, Dropbox, Twitter, Uber, Snapchat and Instagram. Benchmark took a 13.6 percent stake in the company but did not disclose the funding amount. DeepL is unusual in that it is already profitable, having generated a profit of €1.2 million ($1.4 million) in 2017, per regulatory filings.
The fresh capital from Benchmark will enable DeepL to double its current staff of 30 over the next year and ramp up research and development. It also wants to expand globally much more rapidly. The basic translation service is free, and paid upgrades for professional service are available.
Translators often use DeepL to make their work easier, getting a basic translation they can then correct and improve. But DeepL and its competitors are not about to put translators and interpreters out of work any time soon.
The programs are often very good, and “DeepL is in the meantime often better than a human translator,” attests Andrea Bernard in a recent paper for the German Association of Translators and Interpreters, DVÜD. But the quality could mislead customers to blindly trust the machine translations and publish their documents without any further review, Bernard warned. The programs still make mistakes that may be obvious only to native speakers or professionals.
Oliver Voss is a reporter for Tagesspiegel, a sister publication of Handelsblatt, where this story originally appeared. Darrell Delamaide and Grace Dobush adapted it into English for Handelsblatt Today. To contact the author: [email protected]