Textio’s Learning Machine Offers Opportunities to Improve HR Writing
Remember the English composition teacher who could always find a gentle way to improve that middling passage in your writing? The editor who elevated the lightning bug to lightning? That’s a computer now, sitting with you and offering suggestions as you type.
Textio, a Seattle machine learning and natural language processing startup focused on improving job listings and recruiting e-mails for companies including Starbucks, Microsoft, and Twitter, is releasing a new feature that can identify words and phrases that are not bad, per se, but could be better.
The company’s “opportunities” feature is confined to the rather narrow—and importantly, measurable—world of writing done by recruiters and hiring managers. It only suggests improvements to a few words here and there, rather than providing a wholesale rewrite.
But it’s a fine example of human-machine collaboration and a reminder that what once seemed like the most subjective of human tasks can, with enough data and the right algorithms, be effectively handled by a computer. At least that’s what Textio is betting.
The company’s main service, in customers’ hands for about 10 months now, takes a job listing or recruiting e-mail (though people have plugged in all manner of writing) and scans it for attributes such as language, formatting, and length. Textio highlights words and phrases that attract or repel candidates for specific jobs in specific geographies, and even from specific demographic groups. Recruiters can use this information—presented as an overall document score, as well as more fine-grained suggestions—to tune their job listings before posting them. Textio claims this is helping its customers fill openings 17 percent faster.
The new feature circles words and phrases in the listing that could be better. In these cases, “there’s nothing wrong with what you’ve written,” says Textio co-founder and CEO Kieran Snyder. “It won’t send candidates running to the hills. But we know from looking at the phrase itself and the surrounding semantic and syntactic context, there’s actually a better wording we could swap right in for you.”
For example, Textio may suggest trading “critical role” for “meaningful role,” “extremely” for “deeply,” or “high level of professionalism” for “high level of integrity.”
Each individual change is “highly nuanced” and may lead to only a small uplift in the document’s performance, Snyder acknowledges. But making several small tweaks—Textio finds between two and 10 opportunities for improvement in a typical document—can add up to a job listing that performs significantly better. The company has the statistics to prove it, she says.
The suggestions come from Textio’s growing data set of more than 15 million job listings, as well as the outcomes—did the listing attract qualified applicants, did the recruiting e-mail lead to an informational interview—for a large subset of documents.
Textio is also training its system with the strongest writers among its users. Textio learns from the variations in writing put through its system, Snyder says. “If you are generally a very strong writer—and I know that because your documents always get above [a document score of] 90—then Textio might well begin looking for patterns in your writing,” she says.
That also helps Textio’s suggestions stay current in the fast-evolving, buzzword-filled world of recruiting language, particularly in technology. (Pro tip: stop using “big data” in your job listings. It’s still absolutely relevant—as Textio’s own service illustrates—but has become table stakes, and, yes, cliché.)
The Textio opportunities feature may seem simple to the user, but there’s a lot going on behind the scenes.
In order to spot these opportunities—as opposed to merely identifying good and bad language, which in itself is no small feat—Textio needs to understand the context in which phrases appear, to identify reasonable, higher-performing alternatives that can replace them. It’s more than just finding synonyms.
On the back end, Textio is generating alternatives to the source phrase—the word or words that can be improved upon—and rapidly testing them against surrounding context and the company’s data on language performance in that particular context.
Right now, that context is mostly confined to the sentence in which the source phrase appears. But the technology can also make recommendations based on a broader look at the job listing. The clearest example, Snyder says, is whether a listing has an equal employment opportunity statement. Phrases such as “inclusive workplace” that appear outside of that statement generally perform better, because “it feels less like a legal check-box and more like somebody’s authentically worded statement,” she says.
Textio puts an enormous amount of computing horsepower to work in order to return a suggestion only a few hundred milliseconds after the writer makes the last keystroke. But it’s not just raw processing power.
“It takes also real breakthroughs in how you organize the data,” says co-founder and CTO Jensen Harris. “The amount of pre-processing that you do to create these proprietary data formats that make it possible to do these quick indexes and run the variations is extremely important.”
Harris adds that the requirement of a near real-time response brings a level of difficulty to Textio’s task that a computer program playing chess or Go—and taking minutes between moves—does not face.
It’s worth stepping back to remember what Textio is bringing together to do this, and how new it is. Machine learning, natural language processing, and big data feed its predictive engine—essentially a set of tailored algorithms—which does its magic in a fraction of a second, thanks to commoditized computing power available for rent in public clouds.
“That’s something that we’re only at the precipice of right now,” Harris says. He describes Textio as part of a class of companies “born of this assumption of phenomenally huge computing power, and all of these learning technologies, and knowing how to put them together in service of a product that’s simple to use.”
What would it take to expand the contextual lens for Textio opportunities from a sentence to a paragraph or an entire document? (I wanted to know because Textio intrigues me, and, as someone who earns a living writing and editing, makes me a little bit nervous—particularly this new feature, which seems just a few lines of code away from commoditizing me, at least in part.)
Harris doesn’t address this directly—it sounds like they could do it. But he says it raises a deeper question about what writing, heavily edited by a computer, actually is. How would Textio undertake a heavy edit, while still maintaining “the author’s tone at the level of a paragraph or the level of an entire document? It’s really important to us that your writing still sounds like your writing,” Harris says. Textio right now only suggests replacements for phrases of up to about six words, in order to preserve the author’s voice.
For now, anyway, this seems like a technology that’s poised to effectively pair the best of human and machine intelligence.
Harris describes how Textio is constantly sifting through cached data about the document you’re working on, looking for possible variations. “Even when you’re sitting there in your brain, thinking about what to do, Textio is also thinking about that,” he says.
I like that image of a computer collaborator, quietly waiting as I pause to ponder the next sentence, ready to point out opportunities for improvement. We’ve come a long way from Clippy. Until it decides I’m too slow and goes ahead without me.