Why is autocorrect still so bad?

We’ve all been through that. You want to enter a simple sentence like “What would you like for lunch today?” and it comes across as, “What do you want clean for the start?” Autocorrect errors are so commonplace, and have been for so long, that we hardly notice them anymore unless they’re unintentionally hilarious.

Why is this? It’s 15 years of the iPhone — the device that pioneered and popularized touch-only typing — and autocorrect has been with us in one form or another since the ’90s, when Word automatically corrected accidental Caps Lock keys or common misspellings.

After decades and billion of devices sold, not to mention the meteoric rise of machine learning and AI, autocorrect feels just as silly as ever. In a way, it feels like it’s even gone back by making nonsensical substitutions when a simple letter swap would yield the correct word. Is autocorrect really hard? Or is it not even trying to work as it should? Is it no longer a priority?

The March of the Nines

I first heard about the concept called “March of the Nine” about 20 years ago (although I don’t know where the term came from). I’ve researched and written about the latest voice dictation software. That was back when computer users had to buy software like Dragon Dictate to communicate with their machines.

Dictation software that is 90 percent accurate may sound good, but it’s worthless. If you have to correct one word out of 10, you’re not really going to save much time. Even 99 percent accuracy is really not good enough. It gets interesting at 99.9 percent… if you can dictate 1,000 words onto your computer and only need to correct one of them, you’ve got a tremendous time saver (not to mention an incredible accessibility tool).

But 99 percent accurate isn’t just 9 percent better than 90 percent. It’s actually 1,000 percent better — a 10x improvement — because the error rate goes from one error in 10 words to one error in 100 words.

For every “nine” you stack on the accuracy of an automated process, you make it to appear only marginally better for humans, but you need to improve tenfold to get there. In other words, 99.9999 percent don’t feel much better than 99.999 percent for a user, but it’s still ten times harder for the computer.

Is autocorrect stuck in a March of the Nines? Does it secretly make huge leaps that seem infinitesimal to us? I do not think so. Autocorrect’s error rate is still quite high, while available computing power (especially for machine learning tasks) is a hundred times higher than it was a decade ago. I think it’s time to look elsewhere.

Natural language processing that isn’t

Whether you’re talking about voice assistants like Siri or Alexa, voice dictation or autocorrect, tech companies like to say they use “natural language processing.”

but true Natural language processing remains unattainable for any of these consumer systems. What we are left with is a machine learning-based statistical analysis of parts of speech, which is almost entirely absent semantic meaning.

Think about this: “Go to the corner store and get me a stick of butter. Make sure it’s unsalted.”

If I asked anyone what “it” referred to, everyone would immediately know I was referring to the butter, although grammatically “it” could just as easily refer to the store. But who has ever heard of an unsalted business? If we change that second sentence to “Check if it’s open today,” we know that “it” refers to the store.

This is pretty trivial for humans, but computers are abominable because language systems are built without understanding what words actually mean, only what kinds of words they are and how they are written.

All of these speech-based systems (voice assistants, dictation, autocorrect) rely on a large number of poorly paid contractors to take samples of speech or sentences of text and meticulously mark them up: noun, verb, adjective, adverb, swear words, proper nouns, etc. The computer speech system may know that the misspelled word “soup” should be when you typed “try this soup I just made” because it should be a noun and has most of the same letters as the non-word “you” typed my accident. But it doesn’t know what Soup actually is. Neither do any of the other words in the sentence: taste, did, simple…

I think that’s the real reason autocorrect continues to be so bad. It doesn’t matter how sophisticated your machine learning is or how large its training set is if you don’t even have a surface knowledge of what words mean.

I.D.G

Google will automatically predict full sentences for you in Gmail, but even that is just very sophisticated static analysis. It uses machine learning to determine which phrases most often follow the words you just used when replying to an email with a specific distribution of keywords and phrases. It still doesn’t know what it is means.

To use my original example, autocorrect suggested “what do you want for clean boot” because it doesn’t to know that’s a nonsensical sentence. If my iPhone knew what any of those words actually meant, and not just its grammatical role, it would be easy for autocorrect to just make suggestions, which, you know, are possible human speech. (Of course, that it’s also a mishmash of impossible grammar just goes to show how bad autocorrect continues to be.)

Autocorrect doesn’t seem to be a priority anymore

The fact is, autocorrect isn’t the priority it once was. When was the last time Apple touted a massive leap in autocorrect accuracy when marketing iOS?

In the early days of smartphones, when we were all getting used to typing with big thumbs on tiny touchscreens, the ability to fix our fat-fingered mistakes was a big selling point. It was a core feature that indicated a device’s sleek, easy-to-use software.

Autocorrect, for all its flaws, is old and boring. We’ve lived with its weaknesses for so long that the market doesn’t really see it as a hallmark of usability. We’ve moved on to other issues, like fancy camera features and notifications. I’m sure there are smart, hardworking engineers at Apple and Google working on autocorrect, but it’s probably getting a fraction of the resources given to the team responsible for taking marginally better photos, because slightly better photos can sell phones and slightly better autocorrect can’t.

It’s going to take an absolutely massive leap in AI modeling and performance before our phones have any sense of the semantic meaning of words. But even now, surely a lot more could be done to filter out nonsensical sentences and garbage autocorrect suggestions that generate meaningless chatter.

I would just like to see any improvement at all. Anything to get autocorrect out of the rut it cleaned up for launch.