When we talk about AI, things like driverless cars, lightning-speed medical diagnoses, and smart infrastructure tend to dominate the conversation. That makes sense. On any given day, you might find us talking about those same scenarios around our water coolers too. But, we’d be naive to think that future use-cases like these are simple inevitabilities. For AI to deliver on its potential, we’ll have to peel back the curtain and scrutinize how these models are made.
It’s no secret that the vast majority (90%) of data floating through the digital realm is unstructured. As a result, it’s critical that AI models are properly trained to make sense of that ever-growing pile of text, images, audio and video.
That’s why precisely annotated data is to AI models as high-quality ingredients are to a fine meal. With strong datasets as a base, AI “chefs” can confidently focus on their craft. Without it, they’re trying to make French Onion Soup with no butter and a bag of rotten onions. Things can only end badly.
While we’ve had thousands of years to perfect the art of cultivating produce and harvesting grains, we’re not so far along when it comes to AI training data. Of course, everybody knows that better ingredients make for better products, but as an industry, we’re still in the early years. Right now, that means we’re constantly finding out all the minute seemingly inconsequential details that can cause a training dataset to “spoil.”
AI and ML scientists know this all too well. Right now, most of them spend more than half their time retroactively “scrubbing” tainted training data, trying to salvage what they can.
Take text sentiment annotation, for example. The goal is deceptively simple: Does this sentence express positivity, negativity, or neutrality? However, when you consider the domain-specific, ever-in-flux slang that dominate subcultures across social channels, you start to understand all the ways that can go wrong.
To illustrate the point, let’s consider the following two sentences. “What a screamer!” and “What a howler!” On the surface, those are two sentences with the same structure and meaning. Agree? Good. But now, let’s pretend we’re tweeting about the World Cup Final. In soccer lingo, a “screamer” connotes an epic goal, while a “howler” indicates a boneheaded mistake. Those two sentences we agreed were effectively the same now have completely opposing meanings and correlated sentiments.
That seemingly small variation would make a world of difference for, say, a Sports Marketing firm deciding when to put a jersey on sale, or unfortunately but more crucially, where a police force might need to deploy extra protective measures after a big match.
Not only do data researchers need to be cognizant of specific social contexts, they also need to pay close attention to the biases that arise out of common social contexts. In the realm of computer vision, particularly facial recognition technology, we’ve seen how harmful poorly considered datasets can be in perpetuating inequity by excluding people from access to new technologies.
Truth is that, while data annotation may not garner the same buzz as the sci-fi future use-cases we all know so well, if we don’t really scrutinize and refine our processes for cultivating precision datasets, we’re going to see a lot of firms trying to serve full tasting menus with empty pantries.
That’s why at DefinedCrowd, we’re always pushing for new ways to scrutinize data collection and annotation processes and anticipating the edge cases that other firms tend to let slide through the cracks. Check out our use-cases to learn more about how we’re able to guarantee clients high-quality data at speed and scale. To see what high-quality data can do for you, request a trial or email us at firstname.lastname@example.org.