Sigh. I’m familiar with this stuff. Yes, the experiments are definitely revealing and interesting. As for the analysis I’ve seen of it…
I find it fairly upsetting that it’s not informed in the slightest by an understanding of the space of parse algorithms that work. “Serial” and “parallel” as two possible strategies? What? The way they’re explained isn’t even close to bottom-up vs. top-down, which is an actual distinction that parse algorithms make (neither of those choices is “parallel”). It’s a little bit akin to unique-parse algorithms vs. algorithms that handle ambiguity explicitly, but the questions as stated would not be meaningful for any algorithm I’m aware of, since everything does both at once to varying extents.
This is an emotional reaction which, if somebody takes me to task for it, I’m sure I’ll deserve it, but - It feels like psychologists and linguists are hoping that if they ask enough of these questions they’ll discover what parsing is, since they seem determined to start from a position of assuming nothing about the problem.
Certainly, the fact that something is done some particular way by a computer is of no relevance to how the brain does it. But this is a fifty-year-old field now, and we actually know a great deal about possible and impossible ways to do it. It turns out that it’s a problem which parallelizes very, very badly, so it seems as though the ways in which the brain is different from a uniprocessor, though not fully understood, are not relevant to the fundamental constraints. Which means that whatever humans do (and I wouldn’t assume there’s one answer that’s true for everybody or all the time), it’s a polynomial algorithm in the common case. Natural-language grammars experience exponential blow-up only a few words into a sentence, unless one aggressively resolves ambiguity, so computational complexity definitely constrains the solution space.
And, definitely, garden-path sentences which produce pathological cases are a powerful tool to figuring out what someone’s brain is doing. I just wish linguists doing this research would recognize that they are not parse-theory experts, and that that expertise is relevant. I’ve been told a few times that it’s pretty much deliberately excluded from consideration, and…
Just in general, the school of experimentation that comes up with N algorithms to do X and then conducts experiments to decide which one the brain is using just seems problematic - the answer could almost always be “all”, “something else” or “your understanding of the algorithm is wrong” or “the algorithm is implemented inefficiently for God-Knows-What-Reason”.
The problem is it’s hard to think up other ways to research the brain.
That’s absolutely true. I do feel that experimenters in this area could do a better job of asking questions to which the answers, in the best case, would be useful at all. :)
I thought this was particularly interesting: “Garden path sentences are used in psycholinguistics to illustrate the fact that when human beings read, they process language one word at a time.”