Artificial Intelligence
Think Like Artificial Intelligence: What'd She Say?
How do we process language and why is it so hard?
Posted September 28, 2017
Humans are born with the ability to understand language, but we take so much of that for granted. For example, if you don't speak Tagalog, watch 20 seconds of this news broadcast. Your challenge: Identify when one word ends and the next one starts. You don't need to know what they mean; just recognize the separation between them.
As it turns out, that's really hard! Other than a few English-like words, it's probably impossible. If you've learned a foreign language as an adult, you know it can be difficult to do this task when there is unfamiliar vocabulary. We speak in a continuous flow of sound and you really have to know the words being spoken to recognize the separation between words.
So imagine artificial intelligence trying to process what we say. It requires knowing a lot of words, which ones come near each other, and how they might be pronounced. Sometimes you run into phases that it really struggles with. I have this issue with Siri on my iPhone when I try to dictate the phrase "computational linguistics" (ironically, the thing Siri is doing as I dictate). And I'm not the only one; a lot of friends reported the same problem when I tweeted about it.
Siri also struggles with Southern accents—probably a lot of non-northern U.S. accents—and phrases. This task, called speech recognition, has improved dramatically since the introduction of smartphones when they could learn progressively from listening to a lot of people. Still, errors in dictation and recognition are common.
Your challenge: Think about how you process language. Listen for words that sound the same and think about how you pick which one the speaker means (e.g. "red" the color vs "read" the past tense of what you do with a book—how would you distinguish which one a person means in a sentence)? This will likely bring you back to a lot of high school grammar, and it will help you think about the challenges AI faces and some techniques that may be useful in improving its accuracy.