Print

Print


On 4 January 2013 08:18, Jörg Rhiemeier <[log in to unmask]> wrote:
> Hallo conlangers!
>
> On Friday 04 January 2013 07:36:27 Gary Shannon wrote:
[...]
>> So I guess my
>> question is this: In natlangs, how deep does the deferred elements
>> stack generally go? What depth does it never exceed? Does anybody have
>> a handle on these questions?
>
> At any rate, "stack depth" (I sincerely doubt that "stack" is the
> right concept here, we are rather dealing with tree structures here)
> in human languages is quite limited, and deep center-embedding is a
> no-no.  Most people feel uncomfortable with clauses embedded more
> than three deep, I think, though some people are capable of handling
> more.

Optimal parsing algorithms like PCKY certainly make no use of a stack
structure, but aren't 100% cognitively plausible because a) they
assume unbounded memory and b) it's simple to observe that humans are
not optimal parsers.
I have seen one example (though I'm sure there are probably more) of
research into a general-purpose parser with human-like memory
constraints (http://www-users.cs.umn.edu/~schuler/paper-jcl08wsj.pdf)
which assumes that parsing occurs mainly in short-term working memory,
you can have only 3-4 "chunks" (containing partial constituents) in
working memory at any given time, and memory can be saved by
transforming partial trees to maximize how much stuff you can put into
one chunk by ensuring that you never have to store complete but
unattached constituents. The parser is actually implemented as a
hierarchical hidden markov model where shirt-term memory locations are
represented by a small finite set of random variables whose values are
partial syntactic trees, but access patterns look the same as access
patterns for a stack structure, such that it could be equivalently
represented by a bounded push-down automaton with a maximum stack
depth of 3-4.
That model can explain why some examples of center-embedded sentences
cause interpretation problems in human while other
structurally-identical models don't because the probability of
constructing a certain syntactic structure changes in different
contexts; thus, garden-path constructions that you are very familiar
with (and thus which have been programmed into the transition
probabilities of the HHMM) don't feel like garden-path constructions
anymore.

-l.