Ambiguity Rocks

What’s the big deal with language anyway? Why is it so hard for computers to process it? After all, once we learn a language as a child, we don’t really think too much about how it works (except maybe in English class).

The most difficult aspect of language to process is ambiguity—when one word, phrase, or sentence can have multiple meanings. This happens frequently with words—as evidenced by spelling errors (e.g., they’re, their, and there) and homographs (e.g., bass—do you hear it or eat it?)

However, these ambiguous meanings become even more pronounced as words are strung together into phrases and sentences. Humans usually have a good filter for only hearing the one that makes sense in the given context. Computer have some difficulty with this, though AI research is making headway. Consider the follow example, in which a computer program might have difficulty matching a short sentence with a visual scene:

I saw her duck.

A computer might return three possibilities for this sentence; do any of them match your first impression as you read the sentence?

A B C

Probably, you pictured something like B, maybe like A, and almost assuredly not C (unless you are a sadistic duck-hater). However, a computer program would need to be trained to prefer one of these over the others in any given context.

How do AI systems utilize probability and statistics to behave “intelligently”?

Humans are able to resolve ambiguities and missing information almost without conscious thought through intuition. Computers, as we know, however, must have explicitly defined instructions for processing data. Even if the computer were to choose randomly among possible choices, this random selection routine would have to be specified in the program’s code.

Computers can mimic this intuition by statistically analyzing which of the possible choices are most likely given the context. So, autocorrect might notice that you are typing Should we incl... and autofills the rest ude. Other words may be relevant (e.g., incline), but include is more likely to occur in that particular phrase. If instead the text read Head up the incl..., autocorrect may have made a different choice.

Of course, the calculated most likely choice (based on context and prior history) isn’t always the best choice: Eh, humans make mistakes, too. Ambiguity is a difficult problem.

Are we there yet?

So, can we talk to our computers now? Are they poised to be great conversationalists? What about Siri?

Common misconception: We are close to the Star Trek computer-type linguistic interface.

https://www.youtube.com/embed/MA1hD3XRlh0

Not really. On a surface level, what Siri (integrated with GPS positioning) does is much like the Star Trek computer. However, Siri’s functionality is limited in scope—to the phone and the functions it provides. It cannot play chess with you, or understand and process compound sentences, like “Message my mom that I got home safely AND set my alarm for 3 AM.”