Breaking down an AI-for-translation middleman

Language I/O enables businesses to provide multilingual customer support.
article cover

Dianna "Mick" McDougall

· 6 min read

The word “player” can teach us a lot about the difficulties of language translation.

In English, “player” is one word with many meanings—an athlete, a gamer, a flirt, or streaming video tech, to name a few. In Spanish, there are separate words for each.

Language I/O built its business model around those kinds of intricacies. The AI translation tech startup, based in Cheyenne, Wyoming, translates customer support requests—think emails, social media messages, and live chats in 100+ languages—for corporate support teams. In other words, it enables consumers to ask questions and receive help in their own languages.

The company counts Rosetta Stone, Shutterstock, Fossil Group, and BlueJeans by Verizon among its clients and raised a $6.5 million Series A funding round in January.

“A lot of the more popular engines that folks use today were trained on well-written, structured content—content that was written by professionals and is grammatically accurate for the most part,” Heather Shoemaker, Language I/O’s CEO and cofounder, told us. She added, “When you throw conversational content into the mix, it’s just so much more challenging, because conversational content is messy.”

Translation tech, behind the scenes

Language I/O’s core premise: to act as a middleman that surveys the available AI translation tools—called neural machine translation (NMT) engines—and chooses the best one for an incoming language.

NMT engines offered by Google, Microsoft, and Amazon typically support the broadest array of languages—often more than 100, compared to a dozen or so for smaller engines, Shoemaker said. But up-and-coming engines are giving the big ones a run for their money in some markets, Shoemaker told us, DeepL, a platform based in Germany, is uniquely primed for Northern European languages and Chinese, while Systran, a France-based engine, is well suited for certain Asian language pairs.

abstract image illustrating language i/o's tech

Dianna "Mick" McDougall

“Things look pretty stagnant, from where I’m standing, looking at Google Translate,” Leon Derczynski, associate professor of computer science at the IT University of Copenhagen, told us. He added, “Probably with the larger organizations, they have the market share, it’s not necessarily a core business driver to have a translation model that works well—just that works well enough.”

“Google Translate is one of Google’s flagship products, and we’re constantly working to improve and expand its offerings," Macduff Hughes, engineering director of Google Translate, told us via email. He added, "In the past twelve months, we have retrained and updated the majority of our online and on-device translation models."

And as is the case with large language models and other language-focused AI tools, NMT engines perform best with the English language since there’s more data available. For Language I/O, most of its clients have English-speaking support agents, and in 75% of scenarios, Shoemaker said, one side of the conversation is in English.

“The sad truth of the matter is that it is better quality if one of the sides of the conversation is English, or FIGS (French, Italian, German, Spanish), and…Portuguese,” Shoemaker said. “Though Chinese is doing much better, Japanese is getting better, and we’re seeing progress with really historically problematic languages [for NMTs] like Thai and Korean. But I will be completely honest with you: If we’re talking about translating from Thai to Romanian, the NLP tools we have available to us are just not the same, at the same level, as English to French or English to Spanish.”

The language processing process

Context varies so widely between languages and areas that it’s difficult to get a local translation just right. What Language I/O does—taking off-the-shelf translation technologies and comparing what works and what doesn’t—makes sense, Derzynski said, but there will still be gaps.

Keep up with the innovative tech transforming business

Tech Brew informs business leaders about the latest innovations, automation advances, policy shifts and more to help them make smart decisions.

“The mistakes I see made by these kind of models daily are things like where the word sense isn’t completely clear,” Derczynski said. He added, “You have no guarantee that what’s being translated is what you meant to say.”

Once Language I/O’s machine learning model chooses a translation engine for a given request, the company’s AI system scans the request for acronyms, misspelled words, and company- or industry-specific terms, using clients’ personalized glossaries. (For instance, gaming companies like Wizards of the Coast have glossaries with thousands of entries with no dictionary translation, like Magic: The Gathering’s Eldrazi Devastator card.)

Language I/O’s own NLP tools also scan customer support conversations to flag potential new glossary terms—for instance, terms that aren’t in the dictionary, or words that are used much more than usual. Then it’s sent to the company’s human linguists, who determine the best translation for the term in context.

abstract image illustrating language i/o's metadata collection

Dianna “Mick” McDougall

Many of Language I/O’s clients request that they not hold onto that customer data, so instead, the company decrypts each customer support request for a few seconds to gather metadata to improve its ML models—things like which NMT engine was chosen, whether the translation received positive or negative feedback, how complex and how many words the content was, the source and target languages, how many parts of speech were included, and more.

If a client does explicitly allow Language I/O to keep data from customer interactions, then they will do so. For instance, via a partnership, one travel booking giant allowed the company to access an archive of anonymized chat transcripts once humans had gone through and scrubbed them of personal data.

Shoemaker said that though models rely on data to improve, it’s not safe for vendors to hold onto chat content long-term.

“All of this conversational content is just full of personal data, and holding onto it for training purposes isn’t safe,” she said. “There’s no way to detect every piece of personal data that may be embedded in chats and emails and say that you’ve anonymized or encrypted it. You just can’t find it all.”

Keep up with the innovative tech transforming business

Tech Brew informs business leaders about the latest innovations, automation advances, policy shifts and more to help them make smart decisions.