GLOSS and the Department of Linguistics Colloquium
November 1, at 3pm (not 3:30) in 202 Agate Hall
Volya Kapatsinski
Department of Linguistics
University of Oregon
"Some controversial issues in linguistic theory: Redundancy, individual
variability, and the need for multimodel inference"
Epigraph: "I am quite unwilling to grant that our brain-storage has any great use for economy; instead, I
feel that extravagant redundancy is built in all along the line, and table look-up rather than algorithm is
the normal behavior… A linguist who could not devise a better grammar than is present in any speaker’s
brain ought to try another trade."
Householder (1966:99-100)
Abstract: Cognitive and constructionist approaches to grammar agree that grammar consists primarily
(for some, exclusively) of knowledge of form-meaning pairings (a.k.a. constructions). The constructions
are language-specific and must be learned from one's experience with language rather than known
innately. This much is uncontroversial.
Nonetheless, a number of foundational issues remain unsettled. One such issue is the issue of
abstraction during learning: does acquisition of language involve acquisition of grammar, an intricate
web of predictive relationships amongst features of the experienced stimuli, or would simple storage of
the experienced stimuli coupled with analogical reasoning suffice, as exemplar models suggest (e.g.,
Skousen 1989)? I argue that pure exemplar models are fatally flawed, and that it is necessary to learn
which features are predictive of which other features. Nonetheless, exemplar models work very well in
many cases (e.g., Arndt-Lappe 2011, Eddington 2000, Ernestus & Baayen 2003). Why is that? I argue that
this is because linguistic units are multidimensional and redundant: every feature is predictable from
many other features.
This often makes it possible to predict any feature by letting all other features vote, without giving some
of them a stronger voice. Yet, recalcitrant exceptions where this strategy fails remain (e.g., Albright &
Hayes 2003, Kalyan 2012) and reveal that we actually do learn which features are predictive of others.
Another issue is whether the goal of grammatical description should be description of E-Language (the
language out there, observed in a corpus) or I-Language (the system of mental representation
generating the language out there within an individual, which is not directly observable; Chomsky 1986).
The terms "cognitive grammar" (Langacker
1987) and "cognitive linguistics" (Lakoff 1987) imply the I-Language goal. How can this goal be
reconciled with the fact that language conventionalization is at the level of E-Language, the level of
observable behavior (Weinreich, Labov & Herzog 1968), and that there is much individual variability in
the generalizations that one draws from linguistic experience (e.g., Householder 1966, Barlow 2010,
Dabrowska 2012, Misyak & Christiansen 2010, Yu 2010)? Individual variability at the level of mental
representation can be tolerated, and is inevitable, in a redundant system where many different
grammars will produce the same behavior much of the time (e.g., Ernestus & Baayen 2003). If we are
then to keep to the I-Language goals of cognitive grammar, what we should infer from a corpus is not a
single grammar but rather an ensemble of grammars, each grammar weighted by how much it is
supported by the corpus. This ensemble can then be used to make predictions about future samples of
language drawn from the same population. Fortunately, multimodel inference techniques (Burnham &
Anderson 2002, Strobl et al.
2008) are designed to do exactly that. I shall argue that multimodel inference provides a principled
solution to the problem of idiolects for the study of I-Language/cognitive grammar.