There’s been a small bunch of darkening bananas sitting on the counter for a week because my partner wanted to make banana bread. Convinced that it had become a passing fancy, I put them in the freezer to keep them from going moldy.
Of course, I should have predicted this would prompt the need for them to be immediately be available. “Today is the day I finally make banana bread,” she said.
I told her about the freezer, and immediately took them out, putting them on a plate outside in the sun to accelerate the thawing. This exchange ensued:
Her: What if a squirrel comes and takes them?
Me: Holding an entire bunch of bananas?
Her: That squirrel is HUGE.
Me: I don’t want to meet that squirrel.
Her: That’s actually called a raccoon.
We take for granted that language is composed of words, and words have meaning, but here’s what each of those statements actually means:
Her: What if a squirrel comes and takes them?
Me: Given the weight, size, and rigidity of these bananas, it’s unlikely an animal the size of a squirrel could carry them, so we shouldn’t worry about it.
Her: Responding as if a squirrel of the required size arrived on the scene, she exclaimed about the unusual and unexpected largesse of the rodent.
Me: I would fear for my physical safety if there were a squirrel as large as that..
Her: A raccoon might be physically confusable with a squirrel of that size.
In the first statement, she’s posing a direct question and the words have their expected meanings independent of their context – what you see is what you get as it were. But for the rest, there is a distortion. The second, phrased as a question, is not a question, and refers to the physical ability of a theoretical animal. The third refers to that non-existent animal as if it were actually there, and she experiences great surprise in response to seeing it. Shifting focus, the fourth responds to the original discussion in response to the mock surprise. Finally, the fifth uses words to show she’s clarifying, but she’s using it ironically, so she actually does not mean actually.
Ontological classification allows us to encapsulate the meaning of things, but not everything with meaning can be equally or easily usable in a semantic framework. As our technological limits are pushed, this will make speech interactions more an more usable. For virtually all of the above, the only effect is humor, and none of the statements are literally true but the first. In fact, the last 4 transform a statement about the possible size of a squirrel and change the reference to an actual squirrel, which of course does not exist.
Easy for humans, impossible for machines (today).
You wouldn’t figure out my partner and I are having a laugh about a non-existent squirrel from our words. Quite the opposite. You might though be able to detect emphasis, and thus shifts in meaning, by noting where emphasis lies, and maybe even earnest vs mocking postures based on intonation shifts, by looking at the acoustic features.
Reading through this paper from Interspeech last year, they claim they extracted whether a person was happy or annoyed and the level (to a degree) based on acoustic signals. I’m pretty excited to see big companies taking a very blended approach to how to interpret things like this. It will not immediately make speech devices and experiences more human, but it may make them more usable, by detecting error using valence and power in addition to lexical and semantic information.
Cool.