The L2 Impact on the Acquisition of Dutch: The L2 Distance Effect Job Schepens 1, 2 Frans van der Slik 1 Roeland van Hout 1 ¹ Centre for Language Studies, Radboud University Nijmegen, the Netherlands ² International Max Planck Research School for Language Sciences, Nijmegen, the Netherlands j.schepens@let.ru.nl 1
Linguistic differences 56% of citizizens in the EU member states are able to hold a converstaion in at least one language apart from their mother tongue (Euro Barometer 243, 2006) As compared to 47% 5 years earlier (Euro Barometer 55.1, 2001) 2
Introduction • TOPIC The effect of language background on the acquisition of Dutch • SITUATION It is unclear how similarity affects acquisition of an additional language Quantifying similarity is not straightforward (e.g. McMahon & McMahon, 2005; WALS online, 2011) • RESEARCH QUESTION How does the mother tongue (L1) and additional language background (L2) influence the acquisition of Dutch? • HYPOTHESES Two separate L1 and L2 distance effects vs. one interactive effect of language background 4
Empirical Measure of Proficiency in Dutch • State exam “Dutch as a Second Language” - Produced by CITO, a Dutch organization that produces tests and exams - For non-native speakers who intend to start a higher level education / occupation - 1995 – 2010 - 50,000 test scores available - Enough data to test learning differences across 74 mother tongues • Available data - Individual differences: gender, age of arrival, length of residence, years of education - Contextual differences: educational quality, language background 5
Empirical Measure of Proficiency in Dutch • The speaking exam - 14 tasks in 30 minutes - e.g. provide information, give instructions, and so on - in Dutch television, a lot of ads are made for all kinds of products, even in the middle of a program. What is your opinion about ads on TV? • Evaluation on content and correctness • Passing level 0.010 - 500 points ( ≈ upper -intermediate / B2 level) 0.008 Density 0.006 0.004 0.002 0.000 300 400 500 600 700 Speaking 6
HITTITE Greek Armenian Albanian Evolutionary Measure of Kashmiri Singhalese Nepali List Linguistic Differences Bengali Hindi Panjabi • Cognates contain “ genetic” information Gujarati Marathi - tomaat (NL), tomato (EN), Tomat (DE) Persian List Afghan - Cognate based account of evolutionary relatedness Lithuanian Latvian (Gray & Atkinson, 2003) using expert judgements of Slovenian Bulgarian genetic relatedness in 200 lexical item lists Serbocroatian Slovak • Phylogenetic tree Ukrainian Byelorussian Russian - Branch lengths are proportional to ML estimates of Polish evolutionary change per cognate Breton Irish Romanian French • Linguistic distance Spanish Portuguese - The sum of branch lengths joining one language to the Catalan Italian other (via the most recent common ancestor) represents German Dutch List the amount of evolutionary change between two Afrikaans Flemish languages English Swedish Icelandic Faroese Danish 7
Investigation 1: Influence of the Mother Tongue • Sample • Fixed Effects - 35 Indo-European mother tongues - Gender, age of arrival, length of residence, - 89 different countries years of daily education, additional language, - 33,000+ learners educational quality, linguistic distance • Dependent variable - speaking proficiency Mother Country Mean Linguistic Schooling Group Size Tongue of Birth Speaking Distance Quality Kurdish Syria 487 .426 423 63 Kurdish Turkey 490 .426 454 185 French Congo 491 .398 350 65 French France 531 .398 497 936 French Switzerland 550 .398 517 37 German Germany 558 .037 510 4434 German Switzerland 571 .037 517 190 8
Methods • Linear Mixed Effects Regression L1 C • to model dependencies in variation by estimating group level adjustments to the intercept Kurdish Farsi Iran Afghanistan • assumes adjustments are: • normally distributed, • centred around 0 , and • orthogonal to the individual level noise Learner 1 Learner 2 Learner 3 • With partially crossed random effects • We included languages with at least 20 learners per country only 9
Linear Mixed Effects Regression Results • Distance explains a high percentage of between mother tongue variance • Correlation of observed scores with predicted scores is higher (r = .87, p < .0001) than correlation of observed scores with linguistic distance only (r = -.77, p < .0001) R 2 Learner R 2 R 2 Mother Tongue Model Fit Country of Birth 977.51 (75.2%) 184.33 (14.2%) 137.28 (10.6%) 159,535.0 Null Model Variance (LogLik) Components 4.2% 59.0% 75.1% 1516.6 (Chisq), Model with 6 6 (Chi Df), fixed effects Explained variance p < 2.2e-16 *** 11
Linear Mixed Effects Regression Results Mixed Linear Regression measures of fit for all learners Random Effect Df logLik Chisq Chi Df Pr(> Chisq) Structure Null model L1, C 4 -247,546.18 L1, C 10 -246,336.19 2,419.978 6 < 2.2e-16 *** L1L2, C 10 -246,097.03 478.33 0 < 2.2e-16 *** 11 -246,003.81 186.43 1 < 2.2e-16 *** L1,L2, C L1,L1L2, C 11 -245,993.14 21.35 0 < 2.2e-16 *** L1,L2,L1L2, C 12 -245,945.03 96.21 1 < 2.2e-16 *** 16
P values and HPD confidence intervals Fixed Effects Estimate MCMCmean HPD95lower HPD95upper pMCMC Pr(> |t|) (Intercept) 505.0218 504.8874 498.4352 511.3588 0 0 1. Gender (1 = Female) 7.3931 7.4079 6.7446 8.0468 0 0 2. Age of Arrival -0.7248 -0.7249 -0.7661 -0.6841 0 0 3. Length of Residence 0.6183 0.6181 0.5509 0.686 0 0 4. Years of Daily Education -0.7686 -0.7817 -1.8289 0.2446 0.1368 0.143 5. Secondary School Enrollment Rate 0.1785 0.1798 0.1121 0.2493 0 0 6. interaction 4*5 0.0365 0.0367 0.0242 0.0487 0 0 Name Std.Dev. MCMCmedian MCMCmean HPD95lower HPD95upper Random Effects L1L2 (Intercept) 3.29 2.95 2.96 2.27 3.7 C (Intercept) 8.3 7.86 7.9 6.55 9.26 L1 (Intercept) 11.13 10.49 10.55 8.76 12.46 L2 (Intercept) 3.82 3.88 3.93 2.64 5.27 Residual 31.34 31.36 31.35 31.16 31.55 17
Pair Wise Comparisons Hungarian German 19.64 Hungarian Romanian 18.79, p<.0001 Hungarian English 16.93, p<.0001 Hungarian Monolingual 4.73, p<.0001 • Using the aggregated adjustments by L2 to incorporate interactional effects Polish German 9.44 Polish English 5.06, p<.0001 Kurdish English -1.77 Polish French 2.53, p<.0001 Kurdish Arabic -7.92, p<.0001 Polish Russian -.85, p<.0001 Kurdish Monolingual -9.28, p<.0001 Polish Monolingual -1.88, p<.0001 Kurdish Farsi -13.47, p<.0001 Polish Italian -3.59, p=0.008 Kurdish Turkish -19.9, p<0.0001 German French 36.44 Serbian German 10.21 German English 34.12, p<.0001 Serbian English 2.89, p<.0001 German Italian 31.20, p<.0001 Serbian French -1.14, p<.0001 German Spanish 31.06, p=0.90 Serbian Russian -4.62, p<.0001 German Russian 27.67, p<.0001 Serbian Monolingual -7.89, p<.0001 German Monolingual 26.66, p=0.11 18
(Intercept) 10 L1L2 Adjustments are Normally Distributed 5 0 Quantile to quantile plots of random intercepts -5 with HPD intervals L2 noise > L1 noise -10 • Gap at upper intermediate part of scale • German and Swedish seem extremely useful -2 0 2 Standard normal quantiles (Intercept) (Intercept) (Intercept) L2 C L1 10 20 20 5 0 0 0 -5 -20 -20 -10 -2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2 Standard normal quantiles Standard normal quantiles Standard normal quantiles 20
TOP 10 By-Mother Tongue Adjustments Pearson’s r = .994, n = 74 Language L1 only L2 added Difference German 25.93 27.07 1.14 Swedish 24.97 24.24 -0.73 Slovenian 21.67 19.97 -1.70 Afrikaans 19.27 19.09 -0.18 Danish 18.96 17.56 -1.40 Norwegian 18.90 17.20 -1.70 Estonian 16.53 14.69 -1.84 Papiamentu 15.08 15.14 0.06 English 12.90 16.29 3.39 Byelorussian 12.84 11.55 -1.29 21
TOP 10 By-Best Additional Language Adjustments Pearson’s r = .984, n = 45 Language Difference L1 and L2 Interactions only added German 10.11 10.15 0.04 Swedish 4.49 4.39 -0.10 English 3.33 2.32 -1.01 Czech 2.52 2.40 -0.12 Hindi 2.50 2.30 -0.20 Norwegian 2.46 2.30 -0.16 Hebrew 2.23 2.32 0.09 Slovak 1.94 2.00 0.06 Urdu 1.87 1.69 -0.18 Pashto 1.74 1.03 -0.71 22
Conclusions • We claim that linguistic distance is an essential concept in second language acquisition • Moreover, there may be a similar distance effect for one’s best additional language - The evidence indicates that an L2 distance effect operates side by side and interacts only slightly with an L1 distance effect. This might have consequences for our understanding of multilingual language processing. • Mixed effects modeling enables researchers to analyze interactions between different random effects • Quantitative models of linguistic differences are useful for researchers who want to model differences in language skills 23
Recommend
More recommend
