Speech style and education distinguish the grammatical classes of (ING)
Sarah Horwitz
Introduction
The stability and salience of variable (ING) (workin’ ~ working) (Labov 2001a:86) have foregrounded (ING)’s role in the study of language variation and change (Fischer 1958, Labov 1966/1982, Trudgill 1974, Campbell-Kibler 2007, Houston 1985). Yet, we have limited knowledge of (ING)’s cross-varietal patterns of usage (Tagliamonte 2004:394). We analyze constraints on (ING) across grammatical category in Philadelphia English and show that nominal, verbal and quantifier (ING) are conditioned differently. We argue that (ING)’s variant stylistic conditioning supports a conception of (ING) as more than one variable and venture that (ING) may be less stable than is widely believed. In revealing the sensitivity of sub-surface aspects of (ING) to external factors like style, this study demonstrates the value of nuanced investigations. As a larger scale corpus study, this work also represents a convergence between the findings of corpus studies and the findings of qualitative (including Houston 1985; Hazen 2008) and experimental (including Campbell-Kibler 2007, 2011) work on (ING)’s indexicality.
Despite the -ing suffix’s superficial uniformity, there is evidence that nominal and verbal (ING) evolved from discrete sources (Tagliamonte 2004), which makes these grammatical forms’ divergent behavior less surprising. -ing resulted from the combination of at least two grammatical forms: the present participle, which in Old English was -ende, and the verbal noun, which in Old English was formed with the -ung suffix. In Middle English, through a sequence of reduction processes -ende was gradually replaced by -end, from which point it evolved to -en and finally to -in. Concurrently, spellings of “-ing” or “-ynge” began to replace the Old English -ung spelling. An explosion of variation connected to social evaluation occurred once -ing and -in became commonly understood as variants of the same (ING) suffix. Our study probes one of the factors that emerged to condition variation within (ING), style. We seek to determine if the rate of (ING) usage is sensitive to style, and test the hypothesis that (ING) usage is indeed sensitive to variation across stylistic contexts.
Data
Data come from a roughly age/gender-balanced sample of 40 speakers from the Philadelphia Neighborhood Corpus (Labov & Rosenfelder 2011). Every instance of (ING) was coded for pronunciation using a binary system of classification that consisted of “0s” or “1s”. The “0” code corresponds to the nonstandard, apical [ɪn] variant; and the “1” code corresponds to the standard, velar [ɪŋ] variant. Each instance of (ING) was also coded for grammatical class: nominal (including monomorphemes and nominal gerunds; n = 1015), verbal (participles and progressives; n = 2165) or quantifier (something/nothing; n = 424). The preceding segments (velar-nasal, coronal-nasal, coronal-obstruent, or other) and following segments (velar, pause, or other) were accounted for, and were coded based on their place and manner of articulation.
We follow the model for grammatical coding of (ING) developed by Tamminga (2014:44-49):
Nominal (ING) includes words that would be conventionally defined as monomorphemes and gerunds, and refers to forms that serve a nominal function in a sentence. (1) Good morning! (2) Watch out for the low ceiling. (3) What a magnificent apartment building!
Verbal (ING) includes participles and progressives. These forms are alike in each being partially composed of a verbal head; however, participles, unlike progressives, can be substituted with nouns. (1) She does computer coding. (akin to “She does computer work.”) (2) The winner was jumping for joy.
Finally, something/nothing quantifiers are notable because they can be articulated in more reduced ways, with glottal stops and syllabic nasals, in addition to with (ING)’s common apical and velar variants. Although we acknowledge the existence of these reduced articulations, we have excluded them from our data in order to focus on variation across (ING)’s apical and velar forms.
Analysis
We adopt Labov (2001b)’s Style Decision Tree for classifying speech in the sociolinguistic interview into 1 of 8 contextual styles, further grouped into “Careful” or “Casual” for analysis. We expect that more [ɪŋ] than [ɪn] will appear in Careful speech, as [ɪŋ] is the standard variant of (ING) and has indexical meanings including formality, effortfulness, articulateness and education (Campbell-Kibler 2007). Conversely, we expect that the nonstandard [ɪn] variant, which has been found to index inarticulateness/unpretentiousness, relaxation, and uneducation (Campbell-Kibler 2007), will appear with a higher frequency than [ɪŋ] in Casual speech.
The Style Decision Tree
Stylistic variation across (ING)’s grammatical categories
#working on fixing confidence int
# Working in dplyr
library(dplyr)
# Working in ggplot (graphing)
library(ggplot2)
# Acessing data file
ing.lsa2 <- read.csv("ing.lsa2.csv")
# Turning ing.lsa2 into dplyr dataframe
ing.lsa2 <- tbl_df(ing.lsa2)
# Naming the dataframe
ing.lsa2 <- ing.lsa2 %>%
# Creating data in long format
group_by(Bin_style, newgram3) %>%
# Summarize; add the ret rate and N column
summarise(index=mean(code),N=n())
# Setting limits for confidence intervals ("geom_errorbar") in graph
limits <- aes(ymax=0, ymin=1)
# Making graph of grammatically differentiated (ing) rate across stylistic category
ggplot(ing.lsa2, aes(newgram3,index)) +
geom_bar(stat = "identity", aes(fill=Bin_style), position = "dodge") +
geom_errorbar(limits, position="dodge", width=0.25) +
scale_y_continuous(limits = c(0, 1)) +
ggtitle("Style-shifting in (ING) across grammatical class") +
labs(x = "Grammatical Class", y = "Rate of velar (ING)",
title = "Style-shifting in (ING) across grammatical class",
fill = "Style")
Mean (ing) rate by stylistic category
The mean (ing) rate for each stylistic category was used to evaluate the relative frequency of (ING) variants used in each category. Since the “0” code corresponds to the nonstandard, apical [ɪn] variant and the “1” code to the standard, velar [ɪŋ] variant, a mean (ING) rate closer to “1” indicates that the standard variant was used more often. Conversely, a mean (ING) rate closer to “0” indicates increased usage of the nonstandard variant.
This table shows the mean (ing) rates for each style category. Again, a mean (ing) rate closer to 0 indicates greater usage of the nonstandard variant, whereas a mean (ing) rate closer to 1 indicates greater usage of the standard variant. The table must be interpreted carefully, however, because all of (ING)’s grammatical categories are lumped into the discrete style categories of the Style Decision Tree. These aggregate mean (ing) rates therefore seem to not accurately reflect stylistic differentiation that occurred within grammatical subsets of (ING). This is suggested by the fact that the mean (ing) rates do not cluster into the decision tree’s “Casual” and “Careful” style categories. One would expect the “Casual” categories to have the lowest mean (ing) rates; strikingly, the “Kids” and “Tangent” categories, both “Casual” according to the Style Decision Tree, have the highest mean (ing) rates. In contrast, the “Careful” categories of “Language” and “Soapbox” have lower mean (ing) rates.
## style Bin_style ing.rate
## 1: R Careful 0.4497817
## 2: L Careful 0.2631579
## 3: S Careful 0.3506098
## 4: C Careful 0.4227156
## 5: N Casual 0.2828125
## 6: G Casual 0.3828125
## 7: K Casual 0.4782609
## 8: T Casual 0.4710947Regression Analysis
We fit a logistic regression model to predict (ING) variant used from birth year, level of educational attainment, preceding segment, following segment, speaker gender, style, grammatical class and lexical frequency. A logistic regression was run after a generalized linear mixed effects model failed to converge. We tested for birth year/education, speaker gender/style/grammatical category, and grammatical category/lexical frequency interactions. We use the three-way speaker gender/style/grammatical category interaction to evaluate the differentiated conditioning that occurred across (ING)’s grammatical classes, instead of fitting separate models for each grammatical class as done by Tagliamonte (2004).
Every predictor except style and lexical frequency yields significant main effects, and significant birth year/education, style/verbal (ING), and speaker gender/style/verbal (ING) interactions appear.
# having trouble writing dataframe i'm using into .csv file; once i do that, will clean up this coding
# Working in dplyr
library(dplyr)
# Working in lme4 (regressions)
library(lme4)
# Acessing data file
ing.lsa4 <- read.csv("ing.lsa2.csv")
#getting "logfreq" category in data frame
ing.lsa4$logfreq <- log(ing.lsa4$subtlex.count)
#filtering out -infinity
ing.lsa4 <- subset(ing.lsa4, !logfreq=="-Inf")
#rearrange factor values
ing.lsa4$school.cat <- factor(ing.lsa4$school.cat, levels = c("Not HS", "Just HS", "Some college"))
#rearrange factor values
ing.lsa4$newgram2 <- factor(ing.lsa4$newgram2, levels = c("s", "p", "grm"))
# Making model: birthyear * education & gender * style * grammar & grammar * lexical frequency
mod.ing.lsa.all.fin3 <- glm(code ~ birthyear * school.cat + Pre_Seg + Post_Seg + gender *
Bin_style * newgram2 + newgram2 * logfreq, ing.lsa4, family = "binomial")
# Seeing the results of mod.ing.lsa.all.fin3
summary(mod.ing.lsa.all.fin3)##
## Call:
## glm(formula = code ~ birthyear * school.cat + Pre_Seg + Post_Seg +
## gender * Bin_style * newgram2 + newgram2 * logfreq, family = "binomial",
## data = ing.lsa4)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -2.2696 -0.8819 -0.5389 0.9828 2.2923
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -29.261820 10.350797 -2.827 0.00470
## birthyear 0.011680 0.004620 2.528 0.01146
## school.catJust HS 45.863746 10.830354 4.235 2.29e-05
## school.catSome college 22.544750 10.587274 2.129 0.03322
## Pre_Segcoronal.obs -0.863869 0.188000 -4.595 4.33e-06
## Pre_Segother -1.445304 0.178302 -8.106 5.23e-16
## Pre_Segvelar.N -2.523184 0.552184 -4.569 4.89e-06
## Post_Segpause 0.594542 0.092602 6.420 1.36e-10
## Post_Segvelar.C 0.026993 0.168665 0.160 0.87285
## genderm -0.788527 0.296483 -2.660 0.00782
## Bin_styleCasual 0.400998 0.327850 1.223 0.22129
## newgram2p 8.587300 5.132611 1.673 0.09431
## newgram2grm 10.164209 5.138355 1.978 0.04792
## logfreq 0.619697 0.462630 1.340 0.18041
## birthyear:school.catJust HS -0.023745 0.005596 -4.243 2.21e-05
## birthyear:school.catSome college -0.011303 0.005463 -2.069 0.03854
## genderm:Bin_styleCasual -0.723908 0.466663 -1.551 0.12084
## genderm:newgram2p -0.541727 0.326953 -1.657 0.09754
## genderm:newgram2grm 0.074374 0.347275 0.214 0.83042
## Bin_styleCasual:newgram2p -0.811991 0.352685 -2.302 0.02132
## Bin_styleCasual:newgram2grm -0.388350 0.386753 -1.004 0.31532
## newgram2p:logfreq -0.727677 0.463354 -1.570 0.11631
## newgram2grm:logfreq -0.848041 0.464225 -1.827 0.06773
## genderm:Bin_styleCasual:newgram2p 1.117152 0.510005 2.190 0.02849
## genderm:Bin_styleCasual:newgram2grm 0.763067 0.548266 1.392 0.16399
##
## (Intercept) **
## birthyear *
## school.catJust HS ***
## school.catSome college *
## Pre_Segcoronal.obs ***
## Pre_Segother ***
## Pre_Segvelar.N ***
## Post_Segpause ***
## Post_Segvelar.C
## genderm **
## Bin_styleCasual
## newgram2p .
## newgram2grm *
## logfreq
## birthyear:school.catJust HS ***
## birthyear:school.catSome college *
## genderm:Bin_styleCasual
## genderm:newgram2p .
## genderm:newgram2grm
## Bin_styleCasual:newgram2p *
## Bin_styleCasual:newgram2grm
## newgram2p:logfreq
## newgram2grm:logfreq .
## genderm:Bin_styleCasual:newgram2p *
## genderm:Bin_styleCasual:newgram2grm
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 4852.6 on 3598 degrees of freedom
## Residual deviance: 4046.4 on 3574 degrees of freedom
## AIC: 4096.4
##
## Number of Fisher Scoring iterations: 4Discussion
We discuss three interesting results from our regression analysis - grammatical conditioning, gender and style interaction, and effects of lexical frequency - and then transition into a deeper discussion of (ING)’s grammatical conditioning and the birth year and education interaction.
It is striking that style alone did not significantly predict which (ING) variant was more likely to be used, yet when grouped with gender and grammatical class, style did play a role in predicting the (ING) variant. This reinforces the capacity for nuanced analyses to highlight patterns of conditioning that are obscured in broader analyses.
It is also interesting that significant differences tended to appear for verbal (ING), but not for nominal (ING). While the style and verbal (ING) interaction was significant, indicating that a Casual stylistic context and verbal (ING) significantly decrease a speaker’s likelihood of using the velar /ing/ variant, the style and nominal (ING) interaction was not significant. The apparently divergent sensitivity of verbal (ING) to external factors, compared to nominal and quantifier (ING), underscores (ING)’s grammatically-differentiated conditioning.
The results of the speaker gender/style/grammatical category interaction reveal that women style-shift within verbal (ING), and that women style-shift more dramatically than men do. This divergence, shown by the greater difference for women than for men in rate of velar /ing/ between Careful and Casual speech, is graphed below.
## same problem as above with turning dataframe into .csv - working on fixing it!
# Acessing data file
ing.lsa4 <- read.csv("ing.lsa2.csv")
#getting "logfreq" category in data frame
ing.lsa4$logfreq <- log(ing.lsa4$subtlex.count)
#filtering out -infinity
ing.lsa4 <- subset(ing.lsa4, !logfreq=="-Inf")
#rearrange factor values
ing.lsa4$school.cat <- factor(ing.lsa4$school.cat, levels = c("Not HS", "Just HS", "Some college"))
#rearrange factor values
ing.lsa4$newgram2 <- factor(ing.lsa4$newgram2, levels = c("s", "p", "grm"))
# Turning ing.lsa4 into dplyr dataframe
ing.lsa4 <- tbl_df(ing.lsa4)
# Naming the dataframe
ing.lsa4 <- ing.lsa4 %>%
# Creating data in long format
group_by(gender2, Bin_style) %>%
# Summarize; add the ret rate and N column
summarise(index=mean(code),N=n())
# Making the graph - /ing/ rate across stylistic categories by gender
ggplot(ing.lsa4, aes(gender2,index)) +
geom_bar(stat = "identity", aes(fill=Bin_style), position = "dodge") +
scale_y_continuous(limits = c(0, 1)) +
ggtitle("/ing/ rate across stylistic categories by gender") +
labs(x = "Speaker gender", y = "Rate of velar (ING)",
title = "/ing/ rate across stylistic categories by gender",
fill = "Stylistic Category")
It is curious that while these speech patterns suggest women are more sensitive than men are to differences in stylistic context, males demonstrate greater sensitivity than females to differences in grammatical class. This greater male sensitivity to distinctions in grammatical category is reflected in the greater difference in velar /ing/ rates across grammatical classes for men than for women, graphed below.
## same problem as above with turning dataframe into .csv - working on fixing it!
# Acessing data file
ing.lsa4 <- read.csv("ing.lsa2.csv")
#getting "logfreq" category in data frame
ing.lsa4$logfreq <- log(ing.lsa4$subtlex.count)
#filtering out -infinity
ing.lsa4 <- subset(ing.lsa4, !logfreq=="-Inf")
#rearrange factor values
ing.lsa4$school.cat <- factor(ing.lsa4$school.cat, levels = c("Not HS", "Just HS", "Some college"))
#rearrange factor values
ing.lsa4$newgram2 <- factor(ing.lsa4$newgram2, levels = c("s", "p", "grm"))
library(ggplot2) #for graphing
# Turning ing.lsa4 into dplyr dataframe
ing.lsa4 <- tbl_df(ing.lsa4)
# Naming the dataframe
ing.lsa4 <- ing.lsa4 %>%
# Creating data in long format
group_by(gender2, newgram3) %>%
# Summarize; add the ret rate and N column
summarise(index=mean(code),N=n())
# Making the graph - /ing/ rate across grammatical categories by gender
ggplot(ing.lsa4, aes(gender2,index)) +
geom_bar(stat = "identity", aes(fill=newgram3), position = "dodge") +
scale_y_continuous(limits = c(0, 1)) +
ggtitle("/ing/ rate across grammatical categories by gender") +
labs(x = "Speaker gender", y = "Rate of velar (ING)",
title = "/ing/ rate across grammatical categories by gender",
fill = "Grammatical Class")
Finally, it is worth drawing attention to lexical frequency’s effect in predicting which (ING) variant is more likely to be used. The lexical frequency and grammatical class interaction suggests that, for both verbal and nominal (ING), speakers become more likely to use (ING)’s apical /in/ variant as words occur more frequently. Although this interaction was not significant, it is interesting that a word’s lexical frequency can have a stronger effect on the (ING) variant likely to be realized than the word’s grammatical category, which was here demonstrated to most strongly condition realization of (ING). This highlights a phenomenon that can be more carefully probed in future studies, such as those involving priming (similar to Tamminga 2014).
General Discussion
These findings resemble the grammatically differentiated constraint ranking for (ING) observed by Tagliamonte in a 2004 study of York English. Of the internal and external factors that Tagliamonte evaluated, grammatical category most significantly influenced patterns of (ING) usage (Tagliamonte 2004:398). Subsequent analysis of constraint ranking for nominal compared to verbal (ING) revealed that “NOUNS and VERBS have entirely separate and unique linguistic and social profiles” (Tagliamonte 2004:399). Although differences exist in the factors that Tagliamonte and I found to be significant, in both studies, of two distinct varieties of English, grammatical category primarily conditioned variation within (ING). The significant role of grammar in mediating variation within (ING) is not new (see Houston 1985; Marsh 1866), but as Tagliamonte emphasizes sufficient attention has not been dedicated to examining this trend.
It is notable that stylistic constraints primarily distinguish each grammatical category; these categories have not consistently been differentiated in literature on (ING)’s indexical meaning (Eckert 2012, Podesva 2007, Kiesling 2008). This evidence supports the claim that (ING) is more than one variable (Tagliamonte 2004:400, Tamminga 2014:5). The significant birth year and education interaction, shown by decreasing /ing/ rates over time for speakers who only attended high school and who attended some college, reflects college education’s changing prevalence, as finishing high school is no longer a definitive social achievement and a certain level of social prestige and opportunities are increasingly only available to those with college degrees. This interaction is particularly striking given the effects of the individual birthyear and education predictors. The individual predictors suggest that younger speakers and speakers who have completed high school and/or some college are more likely to use the velar /ing/ variant than older speakers and speakers who have not attended high school. Rather than amplifying this effect, the interaction reverses the effect’s direction, underscoring the complex social phenomena that are reflected in patterns of language use and, once again, the value of nuanced investigations.
Birth year’s interaction with education (as in Horvath 1985 and Labov 1972) and its significant main effect counter (ING)’s corroborated stability (summarized Labov 2001a:86), which suggests a change in progress towards increased /in/ usage by people with lesser levels of educational attainment (i.e. less than a college degree). That is, while the social meanings of the /ing/ and /in/ variants, including “education” and “lack of education” respectively, are remaining relatively constant over time, the social definition of what constitutes “high educational attainment” has been changing. This change is reflected in patterns of language use, as over time people whose educational backgrounds are no longer considered to reflect high educational achievement seem increasingly less likely to use linguistic forms, such as the velar /ing/ variant, that index “high education”.
# Working in dplyr
library(dplyr)
# Working in ggplot (graphing)
library(ggplot2)
# Accessing data file
ing.lsa2 <- read.csv("ing.lsa2.csv")
# Turning ing.lsa2 into dataframe
ing.lsa2 <- data.frame(ing.lsa2)
# Reordering educational attainment values
ing.lsa2$school.cat <- factor(ing.lsa2$school.cat, c("Not HS","Just HS","Some college"))
# Creating graph
ggplot(ing.lsa2, aes(birthyear, code, color = school.cat)) +
geom_point(aes(color=school.cat)) +
scale_colour_manual(values = c("blue","red","green")) +
stat_smooth(aes(color=school.cat, fill=school.cat), method = "lm") +
scale_fill_manual(values = c("blue","red","green")) +
labs(x = "Birth year", y = "Rate of velar (ING) (%)",
color = "Educational attainment", fill="Educational attainment") +
ggtitle("/ing/ rate by education over time")
That the rate of (ING) use is changing over time with educational achievement is striking, and reinforces how deeply patterns of language usage are entrenched in greater social phenomena. This is interesting given (ING)’s purported stability (summarized Labov 2001a:86), but more significantly because it implies further investigations of (ING) can be used to probe socially constructed realities that are evolving below the level of consciousness. Eckert (2012:94) identifies “indexical mutability” as the core property of linguistic variables in the third wave study of variation. That (ING)’s usage was shown to change over time suggests it is not immune to the processes of bricolage by which speakers reconstruct variables to reflect their community’s changing social concerns.
Conclusion
Grammatically differentiated conditioning within (ING) draws attention to what the study of (ING) has yet to teach us, emphasizing the variable’s continued relevance in third wave studies of variation. It appears subsequent studies of (ING) have the potential to deepen our understanding of indexicality and of processes of speaker agency such as stance (Kiesling 2008). At a moment when understanding “the dynamics of variation in individuals” (Tamminga, MacKenzie & Embick 2) is of emergent scientific interest, (ING) presents a unique opportunity to investigate variation at the individual level, in many individuals simultaneously. Analysis of larger, more variegated corpora, from diverse contexts, may be necessary to further understand (ING)’s socio-indexical meanings.
References
Campbell-Kibler, K. (2007). Accent, (ING), and the social logic of listener perceptions. American Speech, 82, 1, pp. 32-64. ● Campbell-Kibler, K. (2011). The sociolinguistic variant as a carrier of social meaning. Language Variation and Change, 22, pp. 423-441. ● Eckert, P. (2012). Three waves of variation study: the emergence of meaning in the study of sociolinguistic variation. In Annual Review of Anthropology, 41, 87-100. ● Fischer, J. (1958). Social influences on the choice of a linguistic variant. In Word, 14, 1, pp. 47-56. ● Hazen, K. (2008). (ING): A vernacular baseline for English in Appalachia. American Speech, 83, 2, pp. 116-140. ● Horvath, B.M. (1985). Variation in Australian English. Cambridge: Cambridge University Press. ● Houston, A. (1985). Continuity and change in English morphology: the variable (ING). PhD thesis, University of Pennsylvania. ● Kiesling, S.F. (2008). Style as stance: Stance as the explanation for patterns of sociolinguistic variation. In Alexandra Jaffe (ed), Sociolinguistic Perspectives on Stance. Oxford University Press. pp. 171-194. ● Labov, W. (1966/1982). The Social Stratification of English in New York City. Washington, D.C.: Center for Applied Linguistics. ● Labov, W. (1972). Sociolinguistic patterns. Philadelphia: University of Pennsylvania Press. ● Labov, W. (2001a). Principles of Linguistic Change: Volume 2, Social Factors. Blackwell Publishers. ● Labov, W. (2001b). The anatomy of style-shifting. In Eckert & Rickford (Eds.), Style and Sociolinguistic Variation. Cambridge University Press. ● Labov, W. & I. Rosenfelder. (2011). The Philadelphia Neighborhood Corpus. ● Podesva, R. (2007). Three sources of stylistic meaning. Texas Linguistic Forum 51. ● Tagliamonte, S.A. (2004). Someth[in]’s go[ing] on!: Variable ing at ground zero. In Gunnarsson et al. (eds.), Language Variation in Europe: papers from the Second International Conference on Language Variation in Europe, ICLA VE 2 Uppsala, Sweden, June 12-14, 2003. ● Tamminga, M. (2014). Persistence in the Production of Linguistic Variation. Ph.D. Dissertation, University of Pennsylvania. ● Tamminga, M., MacKenzie, L. & D. Embick. (forthcoming). The dynamics of variation in individuals. Paper under review at Language Variation. ● Trudgill, P. (1974). The social differentiation of English in Norwich. Cambridge: University of Cambridge Press.