People people left left.

The other day I mentioned the wonderfully twisted little sentence in the title of this post, courtesy of Gary Marcus. Although arguably grammatical, the sentence does not correspond to the way human brains actually think.

Its meaning, said in a more human-friendly way, is (more or less): “Some people, who were left by some other people, left.” English grammar. with its enormous flexibility, allows us to phrase this more tersely as “People people left left”, but that doesn’t mean anybody actually either talks or listens that way. In other words, if you try to write down formal rules for generating natural language, you will end up generating some sentences that are not really natural speech — in the sense that nobody would ever say them, and nobody would understand them.

Conversely, there are English sentences one might think should be convoluted and hard to understand, but that nobody has any trouble with. In his book The Language Instinct: How the Mind Creates Language, Steven Pinker quotes a great old example:

Daddy trudges upstairs to Junior’s bedroom to read his son a bedtime story. Junior spots the book, scowls, and asks, “Daddy, what did you bring that book that I don’t want to be read to out of up for?”

As a grammatical structure, the kid’s sentence is nested four levels deep, and yet English speakers — including children — generally have no problem understanding it. As Pinker points out, this is because the four nested levels of the boy’s sentence are all different.

In computer science terms, and putting the two examples together, people have no trouble with sentences structured like this: “( [ { < > } ] )” — as in Pinker’s example — but we get stuck on seemingly simpler sentences structured like this: “( ( ) )” — as in the example from Gary Marcus.

In other words, human brains don’t really work as stack machines when creating and understanding sentences. Our brains operate by matching tokens — and those tokens need to all be distinct.

Circling back around to my previous post about Getting to know your robot, I’ve been playing around with software that works with what we know about natural language, as a way to — potentially — bridge the enormous gulf between the way people naturally think and the way computers need to be told how to follow instructions.