-
1
Martin Schweinberger http://martinschweinberger.de Universität
Hamburg Email: [email protected] Freie
Universität Berlin Email: [email protected]
Particles and Priming - Combining sociolinguistic and
psycholinguistic
determinants of variation
1. Setting the stage & Research question(s)
Sociolinguistics
- Branch of linguistics that investigates the correlation
between social factors and language use, e.g. it looks at
differences between the language use of women and men or between
young speakers and old speakers.
Psycholinguistics
- Branch of linguistics that focuses on the relationship between
our brain and language production, comprehension and processing.
Psycholinguists investigate the cognitive underpinnings of language
and analyze to which extent language use can be explained by mental
processes e.g. processing difficulty, priming, etc.
Q1: Is our understanding of processes of language change and
variation enhanced, if we include psycholinguistic factors in
sociolinguistic analyses of linguistic variation?
Q2: To which extent do advanced statistical methods
(mixed-effects regression models) improve our understanding of
determinants of language change and variation (compared to
conventional statistics, e.g. fixed-effects regressions,
GoldVarb)
2. Theoretical framework
Modern sociolinguistic theory
- Most of the linguistic changes in progress studied in the 2nd
half of the 20th century show a high degree of gender
differentiation and social stratification (Labov 1994, 2001,
2010).
mailto:[email protected]�mailto:[email protected]�
-
2
- Linguistic variation is not random, but correlates
systematically with certain social factors (age, gender, class,
ethnicity, etc.).
- In fact, there seems to be an underlying systematicity not
only in the correlation between social variables and language use
but there are also certain reoccurring patterns with regard to
which groups are leaders of linguistic change and which lag
behind.
Systematic and reoccurring patterns in language change
- Adolescent, lower middle-class females are typically the
leaders of change.
- In cases of change, female speakers are typically one
generation ahead of male speakers.
- Males exhibit higher rates of non-standard features while
female speakers are more sensitive to stigmatization.
- Successful changes typically spread from the center outward,
i.e. from the middle classes, not from top to bottom or vice
versa.
- Our current models of language variation and change have the
advantage that they are …
- based on many studies (highly stable)
- high predictive and explanatory power
- But these models also have shortcomings…
- neglect of language contact and multilingualism
- do not take psycholinguistic factors into account
- rely on overcome statistical methods (GoldVarb, fixed-effects
models)
- Sociolinguists are aware of these shortcomings and have begun
to look at priming (persistence) (cf. Szmrecsanyi 2006)
- Also, the fact that classic fixed-effects (GoldVarb)
regression analyses are likely to produce unreliable results has
been criticized and discussed (Johnson 2009).
- The question is to which degree is the high degree of social
stratification of language use an artifact of inappropriate
statistics and ignoring psycholinguistic factors.
-
3
Traditional sociolinguistics Modern sociolinguistics Tomorrow’s
sociolinguistics
Methodology Frequency analysis & bivariate
statistics Multivariate statistics Sophisticated
statistical modelling (non-)parametric tests GoldVarb analysis
Advanced statistics (R)
χ2-test, t-test, Wilcoxon Sign-Rank test Logistic Regression
Generalized linear (mixed-effects) models, PCA, MDS
Proponents Trudgill Labov Szmrecsanyi
Chambers Rickford ? Preston Tagliamonte ?
Table 1: Methodological evolution of (Variationist)
Sociolinguistics.
3. What is priming (or persistence)?
- Priming refers to the tendency of a speaker to reuse
linguistic material which occurred in the previous discourse, i.e.
previous exposure to a linguistic element will make it more likely,
all other things being equal, that the same element will be used
again. (cf. Szmrecsanyi 2006:2)
- For example, if some says “hello” when greeting someone else,
it is more likely that the other speaker will also say “hello” and
not “hi” or “How are you”.
- The idea is that once the neuronal network for a certain
linguistic expression is activated (by hearing that expression or
using it actively), the activation first peaks and then starts to
decay but the decay will take a while.
- If an appropriate chance to use that expression comes up again
before the activation of the neuronal network has fully decayed,
the “primed” expression is more likely to be used again because it
needs less effort to reach the activation threshold.
-
4
Figure 1: Decay of neuronal activity (priming effect) after
activation event at time t0.
4. The phenomenon
Pragmatic particles
- Not verb + particle constructions like get up
- Subtype of discourse markers
- Prototypical examples
- sort of, kind of, like, so, you know
- Syntactically optional
- Semantically (rather) empty
- Pragmatic and/or social meaning (multifunctional)
- Do not change truth conditions of a proposition
-
5
eh in New Zealand English: Examples
(1) ICE-NZ:S1A-002#B: remind me to ring up early in the morning
not early about er … what time does the … ICE-NZ:S1A-002#B: oh be
up there at ten eh ICE-NZ:S1A-002#Q: i'm out of here at nine
(2) ICE-NZ:S1A-004#M: they're frightened that they might stop
them sitting states ICE-NZ:S1A-004#M: it's a really bad buzz eh
ICE-NZ:S1A-004#G: oh yeah yeah
(3) ICE-NZ:S1A-004#M: er she's working with piki teaching
ICE-NZ:S1A-004#G: oh is she ICE-NZ:S1A-004#M: cos you know her eh
ICE-NZ:S1A-004#G: marie yeah i know her
eh in New Zealand English: Properties
- Typically in turn-final position (non-polar tag)
- Much stereotyped and evaluated negatively in NZE (vernacular
and not standard NZE)
- Used more by Maori men than by Maori women or Pakehas
(British/European New Zealanders)
- Young Pakeha women, though, seem to be the next highest users
of eh.
- Typical feature of working-class speech (Stubbe & Holmes
1995:84)
- Functions as an in-group signal of ethnic identity for these
speakers. (Meyerhoff 1994:371)
5. Data
International Corpus of English (ICE)
- ICE New Zealand
- Most informal register (S1A): face-to-face conversation,
telephone calls (highest frequency of non-standard and discourse
features)
-
6
Summary Statistics
Table 2: Survey of the data of this study.
Sex
male female Total
speakers 86 144 230
files --- --- 100
turns 11,821 19,394 31,215
words 89,935 164,695 254,630
eh 185 257 442
Data Plotting
Figure 2: Distribution of eh over AGE and SEX.
-
7
Figure 3: Boxplot showing the number of words between prime eh
(in previous context) and target eh (at the end of the respective
turn) (left boxplot) and the number of words preceding the end of
the turn if no prime occurs divided by 2 (control group) (right
boxplot). The difference is statistically significant; significance
testing was performed by a one-sided t-test for independent samples
without equal variance.
-
8
Figure 4: Histograms and density plots showing the distribution
of eh across speakers (top) and files (bottom). Table 3:
Description of variables, their coding, and their levels.
Dependent variable HIT nominal yes/no occurrence of turn-final
eh
Independent variables FILE categorical File header SPEAKER
categorical Speaker ID AGE numeric Age groups in ascending order
SEX nominal Male/female ETHNICITY categorical Pakeha/Maori/other
PRIMING nominal Yes/no occurrence of eh within last 30 words before
turn
OCCUPATION categorical Unskilled manual labor (UML), skilled
manual labor (SML), clerical (CLE), managerial (MAN), professional
(PRO)
-
9
6. Statistical designs
What is a Regression?
- A regression or regression model is a statistical method which
measures if there is a meaningful relationship between some
phenomenon (dependent variable) and various factors (independent
variables).
- For instance, a regression can tell us whether monthly income
of someone (dependent variable) is influenced by her or his degree
of education (independent variable), i.e. whether people that have
attended university earn, on average, more than people who have not
attended university.
- If there is a meaningful relationship between the dependent
and the independent variable(s), then this relationship is called
„significant“ and what is called „effect size“ tells us how strong
this relationship is.
- For example, having A-levels (Abitur) correlates significantly
with attending a university and this „predictor“ (having A-levels)
has a substantial effect size, i.e. it tells us that if you have.
Your A-levels, then it is very likely that you will also have
attended university.
- To make this clear, having a beard is correlated significantly
with being male, but since not all men have beards, the effect size
of the predictor (having beard) is high but not perfect.
Figure 5: Schema depicting the main difference between
fixed-effects and mixed-effects regression models.
-
10
Specifics of the statistical modeling
- The statistical analyses were performed by using R.
- The classic Logistic Regression Model was validated using
bootstrapping and a correction for ‘overfitting’ was performed
based on AIC comparisons The correction was implemented by means of
penalty-factors (.95).
- Model fitting followed a step-wise step down protocol.
- Everything I present today, i.e. the data, the annotated R
code, plotting scripts and spreadsheets are freely available at my
homepage (martinschweinberger.de).
7. Results
Table 4: Summaries of the results of three regression models:
The left panel summarizes the results of a fixed-effects logistic
regression model, the mid panel summarizes the results of
generalized mixed-effects model specified as a binomial logistic
model. The right panel summarizes the final minimal adequate
mixed-effects model.
Traditional statistics (Logistic Regression
Model)
Modern statistics (Generalized Linear Mixed
Model)
Tomorrow’s statistics (Gerneralized Linear
Mixed Model)
odds ratio
increase/ decrease (%)
p-value odds ratio
increase/ decrease (%)
p-value odds ratio
increase/ decrease (%)
p-value
Intercept 0.0442 -95.5
-
11
- If a turn is uttered by a Pakeha, then the chance that there
is an eh at the end of the turn decreases by 72.3 percent.
- There are neither significant interactions nor do the
occupation of speakers or priming have a significant effect.
- The study did not detect a significant correlation between eh
use and the occupation of a speaker and thus challenges that eh is
typically a feature of working class speech.
- However, eh is typically male and is exhibits substantial age
grading, i.e. younger speakers use it more than older speakers.
- While eh is not an indicator of class membership, it is an
ethnic marker and significantly more common among Maoris than among
Pakehas.
A statisticians interpretation
- Bivariate statistics are oftentimes misleading because they do
not correct for confounding factors.
- Traditional statistical models tend to overestimate
significance and studies relying on their results may have
overestimated the impact of extra-linguistic variables (of sex,
age, priming, …).
- Studies which have not taken hierarchical /nested variable
structures into account –although nesting tokens would have been
required– may well be fatally flawed.
- Traditional models have been the basis for sociolinguistic
theorizing for the past 30 or so years, but these models may have
led to flawed generalizations about the workings of language…
Conclusion
Converging psycholinguistics & sociolinguistics
- Although priming did not turn out to affect the use of eh,
including psycholinguistic factors into models that analyze
language variation and change is recommendable.
- Advanced statistical models outperform traditional
fixed-effects models and lead to more robust and reliable
interpretations of complex data sets
- This case study of eh has shown that it has a distinct
sociolinguistic profile and that implementing psycholinguistic
factors into sociolinguistic models of variation and change has
great potential to improve our understanding of determinants of
linguistic variability.
Outlook
- What to do next and what to do better
-
12
- I would like to apply the approach I have presented here to
other varieties of English and more phenomena which are currently
undergoing change, for instance
- General Extenders (stuff like that, and so on, and shit)
- Intensifiers (totally, so, wicked)
- I will definitely take a closer look at priming in the context
of Construction Grammar and particular with respect to
collostructional and collexeme analyses.
- I would also like to determine the exact the decay rate of
priming effects by combining experimental and corpus-driven
analyses.
9. References & Appendix
References
Baayen, Harald. 2008. Analyzing linguistic variation – a
practical introduction using R. Cambridge: Cambridge University
Press.
Johnson , Daniel Ezra. 2009. Getting off the GoldVarb Standard:
Introducing Rbrul for Mixed-Effects Variable Rule. Language and
Linguistics Compass 3(1). 359–383.
Galway, Nick. 2006. Introduction to Mixed Modelling: Beyond
Regression and Analysis of Variance. Chichester: Wiley.
Gries, Stefan Th. 2009. Statistics for linguistics with R. A
practical introduction. Berlin & New York: Mouton de
Gruyter.
Maier, Georg. 2012. The Distribution of Subject Pronoun Case
Forms in Subject Predicative Complements in Varieties of English: a
corpus- and web-based study of pronoun case variation. University
of Hamburg PhD dissertation.
Harrington, Jonathan. 2012. Generalised linear mixed models
(GLMM) und die logistische Regression.
http://www.phonetik.uni-muenchen.de/~jmh/lehre/sem/ss10/stat/glmm.pdf,
access October 10th, 2012.
Meyerhoff, Miriam. 1994.Sounds pretty ethnic, eh? A pragmatic
particle in New Zealand English. Language in Society 23.
367-388.
Manning, Christopher. 2007. Generalized Linear Mixed Models.
(illustrated with R on Bresnan et al.’s dativs data).
http://nlp.stanford.edu/~manning/courses/ling289/GLMM.pdf, access
October 10th, 2012.
Stubbe, Maria & Janet Holmes. 1995. You know, eh, and other
‘exasperating expressions’: an analysis of social and stylistic
variation in the use of pragmatic devices in a sample of New
Zealand English. Language & Communication 15(1). 63–88.
-
13
Szmreczanyi, Benedikt. 2006. Morphosyntactic Persistence in
Spoken English. A Corpus study at the Intersection of variationist
Sociolinguistics, Psycholinguistics, and Discourse Analysis. Berlin
& New York: Mouton de Gruyter.
Labov, William. 1994. Principles of Linguistic Change. Vol. 1,
Internal Factors. Oxford: Blackwell.
Labov, William. 2002. “Driving forces in linguistic change”. In
Proceedings of the 2002 International Conference on Korean
Linguistics. http://www.ling.upenn.edu/~wlabov/Papers/DFLC.htm,
access October 14th, 2012.
Appendix
Figure 6: Graphical display of the effect sizes of the sex, age
and ethnicity of speakers according the final minimal mixed-effects
model.
http://www.ling.upenn.edu/~wlabov/Papers/DFLC.htm�http://www.ling.upenn.edu/~wlabov/Papers/DFLC.htm�
-
14
Figure 7: Original age groups and their mean use of ‘eh’ with
95% confidence intervals.
Figure 8: Cluster dendrogram showing the clustering of age
groups according to their mean frequency of ‘eh’ use and their
size.
Particles and Priming - Combining sociolinguistic and
psycholinguistic determinants of variation1. Setting the stage
& Research question(s)SociolinguisticsPsycholinguistics
2. Theoretical frameworkModern sociolinguistic theorySystematic
and reoccurring patterns in language change
3. What is priming (or persistence)?4. The phenomenonPragmatic
particleseh in New Zealand English: Exampleseh in New Zealand
English: Properties
5. DataInternational Corpus of English (ICE)Summary
StatisticsData Plotting
6. Statistical designsWhat is a Regression?Specifics of the
statistical modeling
7. Results8. Discussion, Conclusion &
OutlookDiscussionSociolinguistic interpretationA statisticians
interpretationConclusionConverging psycholinguistics &
sociolinguisticsOutlook
9. References & AppendixReferencesAppendix
