macw at cmu.edu
Tue Jun 10 18:23:02 UTC 2008
This posting is not directly about the Newmeyer analysis, which
seemed reasonable enough to me, but rather about the issue of the rule-
list fallacy. Basically, my point here is that thinking of this as a
dichotomy is useful for a first pass, but soon becomes inadequate when
you think of actual mental processing.
Within psycholinguistics, the interplay between lists and rules
is the centerpiece of the debate between single and dual-route
models. In the dual-route account, one route is rule (or rather
combination with the result adjusted by rule) and the other is rote
(or lists). Evidence for lists usually relies on the existence of
exceptions to rules. Evidence for rules usually relies on
productivity for new forms. In child language, productivity is also
evidenced by overregularizations and other errors. This interplay
between combination and rote occurs on every linguistic level.
The most powerful model of dual-route interactions sees the two
processes as engaged in a horse race. Both operate all the time, but
the winner in a given case is determined by item strength and support.
In 1986, Joe Stemberger and I provided evidence indicating that
even regular forms such as "wanted" are occasionally produced by rote
(or lists), since high frequency regulars are more resistant to speech
errors phenomena then low-frequency regulars. If all regulars were
produced by rule, this type of frequency effect should not obtain.
One can also describe the rule-list interplay in a single-route
connectionist model. However, in my opinion, such models simply
involve recharacterizing rules as gang effects. And these models have
to enforce special procedures to guarantee that lists can survive. In
the end, they just recharacterize the dual route . So, I see no real
way out of the idea of an interplay between processes and a horse race
What does this mean for the FunkNet discussion? Basically, it
means that, in a given usage of a particular form by a given person at
a given moment, one seldom knows whether rules or lists applied. Only
if a clear productive overgeneralization occurs, and this is very
rare, can one know for sure that a rule or gang effect applied. Given
this inherent ambiguity of real data and the essentially competitive
nature of the underlying process, only models that provide a
fundamental role of dual-process interplay make contact with
psychological reality. Fortunately, by examining the numbers in large
corpora, one can get an idea of the overall strengths of the various
horses, but that is about the best one can do and doing that right
takes some care.
Given this horse race view, I find it difficult to understand how
one can view either rules or lists as the unmarked or default case or
suggest that there is any null hypothesis regarding this interplay.
One can entertain opposing null hypotheses, of course, but their half-
life would be measured in seconds.
More information about the Funknet