A View from Christopher Mims
How to Fix Awful Smart-Phone Autocorrection
Are you ready to give up gigabytes of phone storage for a language model of sufficient power?
Certain smart phones are demonstrably awful at autocorrection. Considering that in 2009, 1.5 trillion text messages were sent in the U.S. alone, this is more than just an annoyance, it’s a menace to productivity.
After Spark Capital VC fund principal Andrew Parker’s iPhone auto-corrected Harvard to garbage, Carnegie Mellon doctoral student in machine learning Brendan O’Connor speculated that one way to fix autocorrection on an iPhone would be to add a language model to iOS that would be gigantic enough to know the difference. As in, a language model that could take up gigabytes of the phone’s storage.
Language models try to figure out, based on what you’ve already written, what word you’ll write next. They’re far larger than a normal dictionary – they have to contain the zillions of possible combinations of words that might appear in typical usage.
O’connor guessed that the iPhone thinks that “Harvard” equals “garbage” because the distance between every character in Harvard is close to every character in ‘garbage,’ when laid out on a QWERTY keyboard.
Google’s autocorrect, which is quite a bit better than Apple’s, seems to work by watching what users change a word to, and then feeding that information into a leaning algorithm, opines one of O’connor’s commenters. Another pointed out that the “real problem” with Apple’s keyboard is that it’s not personalized – it doesn’t learn from the words that a user prefers.
The bottom line is that autocorrect doesn’t have to be awful. But perhaps Steve Jobs doesn’t send many texts, so it’s one of those areas of development at Apple that are (in)famously neglected because they don’t affect El Jobso directly.
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today