Unconscious metaphors in the machine
Remember all those painfully boring 10th grade English discussions about what Lord of the Flies really meant? You probably didn’t care much for metaphor then and likely haven’t spent too much time thinking about it since.
You should have. Metaphors could really be important.
We use metaphors all the time. Our brains just handle them so well we usually don’t even notice that we are using them. Every time we read or write or speak, we are likely to encounter some kind of non-literal wording. Language is filled with constructions that require relating the properties two unrelated domains.
Let’s look at language we use to deal with relationships. Often we say things like, “I didn’t feel a connection. He was just too distant.” Or perhaps. “Forget about her; she’s too high out of your league.” The underlying spatial metaphor often passes completely unnoticed. Though if we do recognize the metaphor, it may draw up images of two people being physically distant, which is not the intended meaning.
Metaphor is so deeply ingrained in our cognition that you probably did not even notice the mind-as-a-canvas metaphor in the previous sentence. (I was not entirely aware of it until I had written it.)
When we do not recognize the non-literal language, are we really using a metaphor instead of simply expanding the definition of the phrase? Perhaps we have heard the phrase often enough that the metaphoric link is forgotten. The jury may not be completely settled on this point, but word association experiments seem to suggest that the brain is aware of the underlying metaphor (Pinker 235-279).
As a computational linguist, I have to wonder at the possibility of emulating metaphor in a computer. Would a machine – obviously a very sophisticated one – when hearing a sentences like “He’s an ass” be able to recognize the metaphor and interpret the sentence as intended: “He is not a pleasant person, much like a donkey is a generally crude animal” (or the second interpretation: “He is not a pleasant person, much like the anus is normally an unpleasant organ”).
The ability to interpret metaphor requires a great deal of contextual knowledge. No one could understand the meaning of “Stan is a Norman Borlaug” without knowing who Norman Borlaug is. (He is often called the father of the “green revolution”, and his inventions are credited with saving 250 million lives worldwide.) If I would have said, “Stan is a regular Mother Theresa.” You probably would not have had much trouble.
A system for interpreting non-literal language would certainly need to keep track of a huge amount of similar contextual information. One possibility is, as the program crawls its corpus, build a sort of map of each word or phrase. The map would store associated ideas, attributes and syntactic constructions. Using contextual information, the system could read the maps of the words used in non literally looking for relevant connotations.
This type of system might allow for simple calculations like “Does ‘he’s an ass’ mean the speaker thinks favorably of the man?” or “Is ‘the man who holds a high position in the firm’ important?” A more complex system could potentially give some illusion of understanding some non-literal uses of language.
One step closer to a computer that can tell us what Shakespeare was really talking about.
Sources
Pinker, Steven. The Stuff of Thought. New York: Penguin Books, 2007.
