Lost in translation

If you accidentally Zoomed into a conversation between two data scientists at your company, would you understand what they’re saying?

The idea that data analysts speak a mysterious argot, one that requires translation to and from business-speak, has been with us for a long time. Searches for “analytics translator” peaked nearly a decade ago, in May 2011, then began a gradual decline.

But in 2018 it suddenly became fashionable again to talk about analytics translators. The catalyst was an article in the Harvard Business Review touting the analytics translator as the “must-have analytics role.” The article cited a forecast that there would be demand for two to four million analytics translators in the U.S. by 2026.

We’re a few years into the forecast period, and demand is trending approximately two to four million units below target. Searching LinkedIn, I can find 211 people who have put “analytics translator” in their job title and 26 open positions that use that phrase. Perhaps next year will be the one in which the analytics translator market grows by four orders of magnitude to fulfill the prophecy.

Back when I was getting a degree in linguistics, I took a seminar on how to produce translations of literature. For my project, I chose a short story by Gabriel García Márquez, “One of These Days,” in which a dentist takes corporeal revenge on a corrupt mayor.

If you’ve read works in translation but have never attempted to translate something yourself, I should warn you that it’s harder than it looks.

Your primary mission is to convey the author’s meaning; the reader doesn’t want to hear a competing authorial voice from the translator. Then you want to get the register of the language right: how the impoverished dentist, murderous mayor, and dispassionate narrator would really talk if they spoke U.S. English, along all the possibilities from pedantic to vulgar and florid to austere. It’s a plus if you can capture the sound of the original—obviously so when translating poetry, but even so with prose. And you need to keep it concise; you can’t tax your reader’s patience by hedging everything with long footnotes.

A translator has to enjoy solving puzzles.

There’s a similar challenge within a corporate analytics department. Anybody who has written job specs for an analytics team has discovered that it’s a puzzle of how to marshal three characteristics—business domain knowledge, statistical capability, and a facility with manipulating data—that rarely exist within the same person.

What exactly needs to be “translated,” anyway? In one direction, going from the business to the analysts, the goal is to take a business problem (“What’s driving our customer churn rate higher?”) and use it to generate a set of falsifiable hypotheses for the analytics team to tackle, filling in all the necessary definitions and constraints that make business sense. Then, in the other direction, the translator must take the highly nuanced answer that emerges from the analysts’ work (“Our model explains 50% of the variation in churn as being related to changes in price, cohort effects, and macroeconomic factors”) and turn it back into something the business leaders will understand, along with recommended actions.

It’s not an entirely new problem. I’m currently reading The Riddle of the Rosetta, an engrossing book by Jed Buchwald and Diane Greco Josefowicz about the early-19th-century competition between the Englishman Thomas Young and the Frenchman Jean-Francois Champollion to decipher the Rosetta Stone, a verbose rock that you or a close relative may have posed next to in the British Museum.

Consider a granodiorite slab inscribed with Egyptian hieroglyphics, Demotic text, and Greek text that appear to convey the same meaning. If you try to figure out what each individual symbol represents, you have on your hands (careful, it weighs 1,680 pounds) a predictive analytics problem: When the next slab comes along out of sample, will your model of hieroglyphics generate an answer that has a sensible, coherent meaning?

Riddle suggests that you need to combine different styles of thinking—deep, obsessive expertise along with an integrative appreciation of context—in order to perform good predictive analytics. You also need to watch out for confirmation bias, such as overcommitting to the idea that hieroglyphics are entirely representing ideas or entirely representing sounds like an alphabet (spoiler alert: it’s both).

For most people, a more common translation problem is how to survive a visit to a country where you don’t speak the local language and the locals speak only traces of yours. If you’ve ever visited Japan on hard mode—that is, packing less-than-fluent Japanese—you probably weighed four options:

• Hire a translator and pass all speech through them, back and forth. This is what the HBR article proposes for businesses: treat the foreign language of analytics as an enigma that requires formal translators, lots of them, hired into a central location to serve the enterprise. (If you can’t attract one of the 211 specimens known to LinkedIn, the authors suggest you conscript existing employees and put them through a year-long training program.) Some guidebooks to Japan, seemingly unaware of high-tech options, tell visitors that they should travel with a private tour guide/translator if possible.
• Do it yourself. At the other extreme, maybe you’re better off disintermediating the translators, mastering all three skills and distributing the knowledge throughout the company. Lines of business sometimes install data scientists within their silos, which just moves the translation problem out to the edges. My college Japanese is terrible, so for a recent trip I spent many hours using flashcards to rebuild my knowledge of the language’s three writing systems—which was a long way around the barn just to be able to point out the words for “ramen” and “karaoke” on signs.
• Discover and use relevant technology. Perhaps your company could use a tool that lets business users type or dictate in plain language what they want to know, and have artificial intelligence do the heavy lifting to produce answers. In Japan we used the always-astonishing Google Translate app on our phones to take turns talking with very patient locals, which worked well enough for us to buy the right kind of Fauchon tea in a department store.
• Do a little bit of everything. You could combine elements of each: on-demand access to an expert human (carrying the business card of a bilingual concierge at the hotel to call when all else fails), deliberate training (practicing the few key phrases needed for everyday use and emergencies), and tools (downloading the translation app but recognizing its limits, e.g., if there’s no internet access in the fifth subterranean level of the department store).

The same options apply to a corporate analytics effort. You should hire one supremely capable translator in the role where it matters most: the top analytics leader in the company. You should supply, at a level that matches demand, data and analytics literacy training for the lines of business. You should take a hard look at the current adoption of business intelligence tools to understand the obstacles to empowering true self-service. But, above all, you should invest in a range of approaches to see what works for your company.

The lovely thing about both foreign travel and corporate analytics—in contrast with Young’s and Champollion’s struggles to communicate with ancient interlocutors—is a tight feedback loop. If you get the words wrong, people give you an annoyed look, then you just try it again in different ways until everyone gets what they want. And that’s the tea.

This article first appeared on the Braff & Co. blog at braff.co/advice. Subscribe for automatic updates.