One of the benefits of having good colleagues and a communal
department printer is that you get to discover interesting stuff you would
never have run across. The process (I call itresearch,” btw) is easy: go get
what you have just printed out for yourself and look at the papers that your
colleagues have printed out for themselves that are lying in the pick-up tray.
If you are unscrupulous you steal it from the printer tray and let your
colleague print another for her/himself. If you are honest you make a copy of
the paper and leave the original for your collegial benefactor (sort of a copy
theory of mental movement). In either
case, whatever the moral calculus, there is a potential intellectual adventure
waiting for you every time you go and get something you printed out. All of
this is by way of introducing the current post topic. A couple of weeks ago I
fortuitously ran into a paper that provoked some thought, and that contained
one really pregnant phrase (viz. “encoding-induced load reduction”). Now, we
can all agree that the phrase is not particularly poetic. But it points to a
useful idea whose exploration was once quite common. I would like to say a few
words about the paper and the general idea as a way of bringing both to your
attention.
The
paper is by Bonhage, Fiebach, Bahlmann and Mueller (BFBM). It makes two
main points: (i) to describe coding for the structure features of language
unfold over time and (ii) to identify the neural implementations of this
process. The phenomenal probe into this
process is the Sentence Superiority Effect (SSE). SSE is “that observation that
sentences are remembered better than ungrammatical word strings” (1654). Anyone
who has crammed for an exam where there is lots of memorization is directly
acquainted with the SSE. It’s a well know device for making otherwise disparate
information available to concoct sentences/phrases as mnemonic devices. This is
the fancy version of that. At any rate, it exists, is behaviorally robust and
is a straightforward bit of evidence for online assignment of grammatical
structure where possible. More exactly, it is well known that “chunking”
enhances memory performance and it seems, not surprisingly, that linguistic
structure affords chunking. Here is BFBM (1656):
Linguistically based
chunking can also be described as an enriched
encoding
process because it entails, in addition to the simple sequence of items,
semantic and syntactic relations between items… [W]e
hypothesize that the WM benefit of sentence structure is to a large part
because of enriched encoding. This enriched encoding in turn is hypothesized to
result in reduced WM [working memory, NH] demands during the subsequent
maintenance phase…
BFBM identifies the
benefit specifically to maintaining a structure in memory, though there is a
cost for the encoding. This predicts increased activity in those parts of the
brain wherein coding happens and reduced activity in parts of the brain
responsible for maintaining the coded information. As BFBM puts it (1656):
With respect to
functional neuroanatomy, we predict that enriched encoding should go along with
increased activityin the fronto-temporal language network for semantic and
syntactic sentence processing.
During the
subsequent maintenance period, we expected to see reduced activity for sentence
material in VWM systems responsible for phonological rehearsal because of the
encoding-induced load reduction.
The paper makes
several other interesting points concerning (i) the role of talking to oneself
sotto voce in memory enhancement (answer: not important factor in SSE), (ii)
the degree to which the memory structures involved in the SSE are language
specific or more domain general (answer: both language areas and more general
brain areas involved) and (iii) the relative contribution of syntactic vs
semantic structure to the process (somewhat inconclusive IMO). At any rate, I
enjoyed going through the details and I think you might as well.
But what I really
liked is the program of linking linguistic architectures with more general
cognitive processes. Here, again, is BFBM (1666):
But how does
the involvement of the semantic system contribute to a performance advantage in
the present working memory task? One possible account is chunking
of information.
From cognitive psychology, we know that chunking requires the encoding of at
least two hierarchical levels: item level and chunk level (Feigenson & Halberda,
2004). The grammatical information contained in the word list makes it possible
to integrate the words (i.e., items) into a larger unit (i.e., chunk) that is
specified by grammatical relationships and a basic meaning representation,
as outlined
above. This constitutes not only a syntactically but also a semantically
enriched unit that contains agents and patients characterized, for example,
by specific
semantic roles. Additional encoding of sentence-level meaning of this kind,
triggered by syntactic structure, might facilitate the following stages (i.e.,
maintenance and
retrieval) of the working memory process.
So, there is a (not surprising) functional interaction
between grammatical coding and enhanced memory (through load reduction) through
reduced maintenance costs in virtue of
there existing an encoding of linguistic information above the morpheme/word
level. Thus, the existence of G encodings fits well with the cognitive
predilections of memory structure (in this case maintenance).
Like I said, this general idea is very nice and is one
that some of my closest friends (and relatives) used to investigate
extensively. So, for example, Berwick and Weinberg (B&W) tried to
understand the Subjacency Condition in terms of its functional virtues wrt efficient
left corner parsing (see, e.g. here).
Insightful explorations of the “fit” between G structure and other aspects of
cognition are rarish if for no other reason that it requires really knowing
something about the “interfaces.” Thus, you need to know something about
parsers and Gs to do what B&W attempted. Ditto with current work on the
meaning of quantifiers and the resources of the analogue number system embedded
on our perceptual systems (see here).
Discovering functional fit requires really understanding properties of the
interfaces in a non-trivial way. And this is hard!
That said, it is worth doing, especially if one’s
interests lie in advancing minimalist aims. We expect to find these kinds of
dependencies, and we expect that linguistic encodings should fit snugly with
non-linguistic cognitive architecture if “well-designed.”[1]
Moreover, it should help us to understand some of the conditions that we find
regulate G interactions. So, for example, Carl De Marcken exhibited the virtues
of headedness for unsupervised learners (see here
for discussion and links). And, it seems quite reasonable to think that
Minimality is intimately connected with the fact that biological memory is
subject to similarity based interference effects. It is not a stretch, IMO, to
see minimality requirements as allowing for “encoding-induced load reduction”
by obviating (some of) the baleful effects of similarity based interference
endemic to a content addressable memory system like ours. Or, to put this
another way, one virtue of Gs that include a minimality restriction is that it
will lessen the cognitive memory load on performance systems that use these Gs (most likely for acts of
retrieval (vs maintenance)).
Minimalism invites these kinds of non-reductive functional
investigations. It invites asking: how does the code matter when used? Good
question, even if non-trivial answers are hard to come by.
[1]
Yes, I know, there are various conceptions of “good design” only some of which
bear on this concept. But one
interesting thing to investigate is the concept mooted here for it should allow
us to get a clearer picture of the structure of linguistic performance systems
if we see them as fitting well with the G system. This assumption allows us to
exploit G structure as a probe into the larger competence plus production
system. This is what BFBM effectively does to great effect.
No comments:
Post a Comment