garyfeng Blog

I've been re-reading L&T03, which is perhaps the more comprehensive language training study on ToM. The conclusion is fairly clear -- language helps, no language, no help. Sentence complements are useful, so are simply sentences that pointed out what the child failed to perceive. L&T argued for a discourse-based interpretation.

As much as I like this paper, the question in my mind is their ToM measure -- an appearance-reality test. They were asked 3 questions

1. What do you think is inside (the box before its content is revealed)

2. What did you think was in the box when you first saw the box (after the surprise content was revealed)

3. What your friend X will say in the box when s/he sees the box for the first time

Notice the shift of tense and aspect in the three cases. I wonder if 3.5 year old (German speaking) children will get these.

Also in the sentential complement test II (Hale & Tager-Flusberg, 2003), the child was asked the following question:

"What did the girl say she was cutting?"

which has a deep structure: "the girl say-ed she is-ed cutting WHAT".

This would be easily confused with

"what did the girl .... was cutting?" --> "what did the girl cut" or "what was the girl cutting?"

Did anybody tried to do a shadowing test, and ask the child to repeat the question after certain delays? I suspect what they perceived (or memorized) is quite different the input.

Or, instead of a production, do a forced choice comprehension test.

The point is -- the appearance-reality task hinges on the correct interpretation of a tricky sentence, which involves shifting tense and aspect of a past event. Is this ToM or language comprehension? Or from the Whorfian point of view, does this matter? It's all linguistically constructed.

I start to notice ads placed in the middle of the page on CUL. There were some ads at the side bar, which I don't care. But I don't want any ads in the middle of the page. Please don't!

Looks like CUL has gone through a page redesign. Some of the DOM structure has changed, which has broken my GM script for adding the custom 1-click-add-to-my-library buttons.

The problem has to do with:

// old format:

//var links = document.evaluate("//li/a[@class='title']",

// new format: they added a layer of DIV

var links = document.evaluate("//div[@class='list']/li/div[1]/a[@class='title']",

Now it works again.

Ross, D. H. sapiens as ecologically special: what does language contribute? Language Sciences, In Press, Corrected Proof.

Reference: * My library * Original article

http://www.citeulike.org/user/garyfeng/article/712006

posted a couple of notes there. My biggest question has to do with how PSS might enable and then interact with a linguistic system, which has to be discrete.

I think Barsalou took a convenient shortcut in 2.4.3 to avoid the problem of individualizing, dissecting experience, the classic question raised by E. Sapir. There Barsalou shifted from talking about analogical perceptual symbols -- networks of neurological representations of experience -- to discrete, countable simulators, so that he can equate simulators to concepts. Presumably concepts are what linguistic symbols represent and where they came from.

But after many exposures of four-legged, flat-surfaced, knee-high, man-made objects, how does the simulator (assuming it's individualizable) know that it's its job to become the "chair" simulator? And who draws the boundary between the "chair" simulator and the "stool" simulator, and maybe the "armchair" simulator and the "loveseat" simulator? In essence, why does Barlasou think there is a farm of individual simulators (=concepts), rather than a "simulator soup" (analogical all the way down)?

Not that I think a perceptual system cannot be discrete or categorical, but unless one opens up a mechanism for some sort of supervised learning, it's hard to see how one could arrive a set of "concepts" that is stable and meaningful enough inter-individually to afford symbolic exchanges, linguistically or now.

One possible argument is that linguistic or other kinds of symbols are simply part of the input of the simulator. When they are present, they activate the corresponding simulators (which probably predate the symbols). That's fine as a model of how dogs learn verbal commands, but it falls way short of an account of (a) linguistic communication and (b) hierarchical conceptual systems.

This is obligatory. Gotta try it.