Hi everyone,
Thanks for your comments. I see that some more details that are didn’t
include for fear of blowing up the opening post are in order.
The gender system of the L1 is as follows. In the written standard and
the Northern part of the language area, only two genders exist: common
and neuter. In Southern substandard varieties, a three-way distinction
exists: masculine, feminine and neuter. For all intents and purposes,
standard neuter maps perfectly onto substandard neuter. The L2, like the
Southern substandard, has a three-gender system, and I wondered whether
Southerners, who are literate and were schooled in the two-gender
standard, make use of their three-gender system when assigning gender in
the L2.
I have a list of standard common-gender nouns which I know to be
predominantly masculine or feminine in the substandard. The RQ is
whether Southern participants assign L2 gender to stimuli whose standard
cognates are common-gender on the basis of the substandard gender; the
standard doesn’t provide clues as to how to distinguish masculine from
feminine gender here. In reality, the correct L2 gender (masc., fem.,
neuter) for these nouns cuts across the L1 gender categories – at least
in this experiment –, but it proved impossible to fully balance them
(simply not enough stimuli available for all cells).
By my reckoning, I could code the participants’ responses to stimuli
with a common-gender standard cognate for substandard congruency:
congruent if they assign masculine gender and the substandard gender is
masculine or if they assign feminine gender and the substandard gender
is feminine, incongruent in other cases. In this case, transfer from the
substandard is a likely explanation if Southern tend to provide more
congruent responses. (An alternative would be to code the participants’
responses for their correctness and add substandard L1 gender as a
predictor; Southerners should then be sensitive to this predictor to a
greater degree than Northerners. That is, I’d need an interaction
between Region and substandard gender. But I think the two approaches
accomplish the same thing, and fewer terms would be needed if I code for
congruency rather than correctness. A less satisfying alternative would
be to forget about the North and just model the Southerners’ responses
and include substandard gender as a predictor. But such an effect may be
driven by formal or semantic stimulus-related characteristics; I need
the Northerners as a control group.)
But I’ve been thinking that an alternative or additional explanation for
a larger congruency effect for Southerners (or a Region*substandard
gender interaction) might be that (a) something about the stimuli (all
monomorphemic, monosyllabic and inanimate) nonetheless says ‘masculine’
or ‘feminine’ to Northerners and Southerners alike, too, along the same
lines as the Southern substandard distinction but (b) Northerners know
more about L2 and therefore ignore point (a). To rule this latter
explanation out as best I might, some measure of the participants’ L2
knowledge seemed useful.
The task also contains 15 L2 stimuli whose L1 cognates are neuter in
both the standard and the substandard. For five of these stimuli, the L2
gender is neuter, for six it is masculine, for four it is feminine
(inanimate feminine L2 words with neuter L1 cognates are few and far
between, unfortunately). These are all fairly transparent cognates, so I
presume that most participants will ‘know’ these words receptively. My
reasoning was that participants with little to no L2 knowledge will
essentially pursue one of these three strategies: (a) random guessing;
(b) tick neuter everywhere (transfer); (c) tick another response
everywhere (e.g. masculinisation strategy). Either strategy should
result in a fairly low score compared to participants with more
extensive L2 knowledge. True, strategy (b) would be transfer-based, but
not uniquely substandard-based (the strategy is available to both
Northerners and Southerners).
@Florian: Does that seem reasonable?
I do have an alternative, but less direct L2 proficiency measure at my
disposal: self-assessments on a CEFR-based scale. I guess a cloze test
or a vocabulary task with non-cognates may be alternatives, but these,
too, are indirect. The test will be run online, so I’m doing all I can
not to scare away customers mid-task. (I’ve considered including
non-cognates in the gender assignment task, but low-level participants
seem to have a knack for seeing cognate relationships were there are
none, especially if many cognates occur elsewhere in the task, so what I
intend as non-cognates may not be non-cognates to the participants.) The
suggestions about non-linear proficiency effects and word frequency are
excellent, but seeing as writing up the analysis without the data at
hand is proving difficult enough as it is, I went with Scott’s
suggestion and reserved those for an exploratory analysis. (I guess one
could also imagine that early on, learners aren’t particularly keen on
transferring L1 gender to the L2 but grow more fond of it when acquiring
more and more positive examples.)
Oh, and one more thing - said Columbo:
“If the effect of L2 knowledge is asymmetric for congruent and
incongruent stimuli, then that's how it is empirically (which could have
a variety of interpretations, since presumably there will be many
additional differences between these stimuli that could plausibly
explain the difference in the effects size of L2 knowledge).”
I understand what you mean, Florian, but wouldn’t you also be able to
account for such ‘many additional differences’ by forcing the model to
produce one fixed parameter estimate for ‘effect of L2 knowledge on
response congruency’ but allowing this effect to vary between stimuli
(random slope)? Whatever many additional differences exist, they’re
unlikely to precisely follow the masculine/feminine distinction in the
substandard?
Thanks a lot for your contributions, Scott, Florian and (privately) Dan!