4/18/12

1

Formal Approaches to Heritage Languages April 21-22, 2012 University of Massachusetts Amherst

Maria Polinsky Harvard University

1

  Heritage languages bear significant resemblance to the languages from which they were formed (the baseline)   They tend to amplify certain trends that are

already present in these languages

2

  Heritage languages deviate from the baseline in a number of ways   Contrary to expectations, they do not look

enough like the baseline   Heritage languages bear significant

resemblance to each other   They deviate from the baseline in similar ways

which call for a principled explanation

3

  While there are some parallels between structures/forms in the heritage language and in the dominant language, such parallels are not exhaustive   What prevents heritage languages from

transferring all they need from the dominant language?

4

4/18/12

2

  Viewpoint A: Learning about heritage languages   Arriving at a comprehensive description of

heritage languages, understanding their structure, processing, and origins

  Viewpoint B: Learning from heritage languages   Using heritage languages as a new source of

data feeding into theory construction

5

  New material for understanding language in time and space

−  Language origins

−  Language

acquisition

Better theory of acquisition, development, and evolution

6

  The Bickerton Quadriga

  Adult heritage languages   Definitely a more accessible source than

members of the Quadriga   Allow researchers to collect data and conduct

experiments on a much broader basis

7

  New angle on the core of human language capacity   Hence, new window on Universal Grammar

8

4/18/12

3

  New data for testing our theories of language structure and language processing

−  Language

universals

−  Language structure

Better theory of language

9

  Agreement (phi-features)   Case licensing

10

  Well-established:   [person]   [number]   [gender]

11

  Well-established:   [person]   [number]   [gender]

  Somewhat more tentative:   [status] (honorification)   [wh-agreement]

12

4/18/12

4

  Person agreement never/rarely appears on adjectives

  Probing for [person] and [number]/[gender] occurs in separate derivation steps   (Anagnostopoulou 2003, Béjar 2003,

Chomsky 2000, Laka 1993, Shlonsky 1989, Sigurdsson 1996, Taraldsen 1995, a.o.)

13

  Person agreement never/rarely appears on adjectives

  Probing for [person] and [number]/[gender] occurs in separate derivation steps   (Anagnostopoulou 2003, Béjar 2003,

Chomsky 2000, Laka 1993, Shlonsky 1989, Sigurdsson 1996, Taraldsen 1995, a.o.)

  [person] is probed first 14

  Phi features are internally structured in a hierarchical way   (Harley & Ritter 2002, McGinnis 2005, Béjar &

Rezac 2009, Coon & Preminger 2010; Preminger 2011)

15

  Person > number/gender

  Production: various errors, often with the default person and number/gender

  Comprehension: greater sensitivity to person mismatches than to number or gender mismatches

16

4/18/12

5

  Auditory experiment evaluating sensitivity to gender, number, or person mismatch in the verbal paradigm

  Rating task (1-7 scale), subject and verb separated by two words; all verbs stem-stressed (to minimize cuing)   Too bad that she quite often think…   24 subjects (avg. age 26), intermediate

proficiency 17

  The asymmetry is confirmed: person > number > gender

  If person is probed first, the failure of probing on other features may be accounted for by the need for a truncated structure

  Possible mechanism: shallow structure building, with only closest nodes accessed?

18

ClassP

[ ] NumP

[ ] πP

[ ] …

… …

Subject

…

19

X X

  Unexpected: gender is different from number

  Two possibilities:   Number is probed earlier than gender, and the

hierarchy needs to be more fine-grained   Number and gender are still equal as phi-

features but the computation of number is determined situationally ▪  the computation of gender requires going back to

the lemma, which imposes extra processing costs

20

4/18/12

6

  Within HLs:   optimize processing conditions for heritage

speakers to see if the contrast disappears in comprehension

  make comparison conditions for controls more difficult

21

  Outside HLs: look for evidence that gender and number probe sequentially

  A possible candidate: Highest conjunct agreement (HCA), as in Arabic or Hindi (Aoun et al. 1999, 2004; Benmamoun et al. 2009, Bhatt & Walkow 2012; Polinsky 2012)

22

  Agreement marking on a probe depends on the properties of only one conjunct, structurally or linearly the closest one (HCA)   Also known as single conjunct or partial

agreement   In contrast with full or resolved agreement

(RA): agreement marking on the probe results from the computation over the properties of all the conjuncts

23

  VS order: ✓HCA ža Omar w Karim

Moroccan came.3MSg Omar and Karim

Arabic ‘Omar and Karim came’

  SV order: ✓RA Omar w Karim žaw Omar and Karim came.3PL ‘Omar and Karim came’

24

4/18/12

7

  VS order: ✓RA žaw Omar w Karim came.3PL Omar and Karim ‘Omar and Karim came’

  SV order: *HCA *Omar w Karim ža Omar and Karim came.3MSg

25

Resolved Agreement

Highest Conjunct Agreement

Subject before verb (SV)

Yes No

Subject after verb (VS)

Yes Yes

26

  [number] probes for the highest target, the coordinate DP inactivating it (cf. Rackowski & Richards 2005)

  The next probe, [gender], then targets the highest remaining target:

in this case, the DP in [Spec,&P].

27

  Unexpected: gender is different from number

  Two possibilities:   Number is probed earlier than gender, and the

hierarchy needs to be more fine-grained   Number and gender are still equal as phi-

features but the computation of number is determined situationally

28

4/18/12

8

  HL hierarchy of phi features: person > number > gender

  Implications for the existing hierarchies: class (gender) is not on the same level as number

29

  Structural case: assigned in a certain structural configuration

  Inherent case: assigned depending on theta-marking

  Lexical case: assigned based on an idiosyncratic property of the assigning head

30

Baseline   Structural cases:

NOM (unmarked), ACC

  Inherent cases:   Assigned by a verbal

head: DAT   Assigned by a nominal

head: GEN   Assigned by a P head:

INSTR, LOC

Heritage, production   Structural cases: null

forms   Inherent cases:

  DAT ACC   Cases assigned by non-

verbal heads: null forms, confusion of oblique cases (noisy production)

31

0

10

20

30

40

50

60

70

80

90

100

NOM ACC DAT exp DAT goal GEN

% C

orre

ct

suppliance omission overgeneralization

82 subjects avg. age 21

32

4/18/12

9

Baseline   Structural cases:

NOM (unmarked), ACC

  Inherent cases:   Assigned by a verbal

head: DAT   Assigned by a nominal

head: GEN   Assigned by a P head:

INSTR, LOC

Heritage, comprehension

  Structural cases: high tolerance of null forms

  Inherent cases:   DAT and ACC both

accepted   Cases assigned by non-

verbal heads: low tolerance of null forms

33 34 36 subjects, avg age 18;6

Production Comprehension

Attrition?

Structural case

Replaced by an unmarked form or a frequent form, many errors

Mismatches tolerated

Yes

Inherent case Maintained as in the baseline or replaced by another marked case

Mismatches not tolerated

No

Lexical case Replaced by a more regular form

Mismatches ignored

May vary depending on lexical item

35

  Structural case gets replaced/attrited, inherent case is retained

  Both production and comprehension are necessary to evaluate case in HLs   inherent cases assigned by P and N were truly

testable only in comprehension

36

4/18/12

10

  Applying these generalizations to case configurations where primary data have not resulted in a conclusive analysis

37

  Most agree: assigned in spec, vP   Disagreement concerning the licensing of

ERG:   Inherent case (Woolford 2000, 2006, a.o.)

  Structural case (Bobaljik & Branigan 2006, a.o.)

38

  Heritage Hindi:   High rate of omission of the ergative postposition

–ne and replacement by null form in production (G. Mahajan 2009, Montrul et al. 2012)

  Heritage Hindi:   Inappropriate ergative receives high ratings

(Montrul et al. 2012)

  ERG is treated as a structural case 39

  ERG undergoes significant attrition, symptomatic of structural case

  GEN and DAT are regularized but well preserved, symptomatic of inherent case

40

4/18/12

11

  Avar: Nakh-Dagestanian (NE Caucasian) language   One of the largest in Dagestan (about 750K speakers)   A large number of speakers are bilingual in Russian

and Avar   Core cases: ABS, ERG, DAT, GEN (and many

locatives), agglutinative marking

ABS ERG DAT GEN LOC1 was was:-as was:-as-

ul was:-as-e was:-as-

da

was ‘son, boy’

41 42

0

200

400

600

800

1000

1200

1400

ABS ERG

Mill

isec

onds

Controls Heritage

42 controls, 18 HL speakers

*

  Production: relatively low error rate   Ergative replaced by the absolutive and dative   Dative is replaced by the absolutive and

ergative   No case omission

  Morphological leveling of case allophones   (N=5 speakers, video clip story)

43

  Most agree: assigned in spec, vP   Disagreement concerning the licensing of

ERG:   Inherent case (Woolford 2000, 2006, a.o.)   Structural case (Bobaljik & Branigan 2006,

a.o.)   Solution:

  Two different ergatives, structural in some languages, inherent in others

44

4/18/12

12

  Differential properties of the ergatives across HLs   Some languages have structural ergative

(Hindi, Dyirbal)   Other languages have inherent ergative (Avar)

  Next question, addressed to theorists: what parametric properties are correlated with the two different ergatives?

45

  Heritage languages are a valuable source of data that feed back into linguistic theory

  We need to learn how to mine this source

46

47

  Four- way comparison:   HL adults   HL children   Monolingual adults   Monolingual children

  This allows us to separate attrition from incomplete acquisition

48

4/18/12

13

  Do child learners (future heritage speakers) and adult heritage speakers have the same morphosyntactic deficits?   If a child and an adult deviate from the

baseline in the same way, the feature has not been acquired

  If a child and an adult perform differently, the feature has been acquired but lost/reanalyzed

49

Adult heritage language = fossilized child language, with the level of fossilization roughly corresponding to the age of interruption?

50

  Feminine: cerkov’ ‘church’, tetrad’ ‘notebook’, krovat’ ‘bed’, sol’ ‘salt’, ten’ ‘shadow’

  Masculine: put’ ‘way’, dožd’ ‘rain’, portfel’ ‘briefcase’, kalendar’ ‘calendar’

  Standard child language error:   feminine nouns are interpreted as masculine,

up to age 7;0 (Gvozdev 1961)   independent of frequency

51

  Gender of feminine nouns in palatal consonant is acquired late and poses a problem for monolingual and heritage children alike

  This incompletely acquired feature then persists in HL adults

52

4/18/12

14

Adult incomplete grammar undergoes attrition and is different from the “initial state” represented by heritage child grammar

53

  Acquired early (2;0-2;6)   Universal preference for subject relatives   Error rate (wrong head choice), ages 4-6:

  English : 10%-13% (multiple studies)   Indonesian: 11% (Tjung 2006)   Mandarin Chinese: 3.9% (Hsu et al. 2006,

2009)   Turkish: 4% (Slobin 1985)   Russian: 3.7%-4.2% (Fedorova 2005, Polinsky

2008, 2011)

54

55 Adults (C/H): 17/21, age 24; children (C/H): 6/23, age 7

  HL children perform on par with age-matched monolingual controls and significantly outperform HL adults

  The syntax of relative clauses undergoes a reanalysis across the lifespan and presents a case of attrition

56

4/18/12

15

  Same HL with a different dominant language: minimize the effect of transfer

  Structuring the tests in such a way that we could go against the transfer   (Russian relative clauses, Polinsky 2011)

57

  Distinguish heritage speakers from heritage language learners

  So far, no direct comparison between heritage speakers “in the wild” and HL re-learners   Many subjects of HL studies are drawn from

HL classes (a self-selected group)

58

  10 English-speaking learners of German in Switzerland   16.5 – 18 years old (mean 17.5)

  Tested   After three weeks (following an introductory

course)   Five months later

  Test performance correlated with increase in gray matter density

59

  Increase in gray matter density over five months correlated positively with difference in proficiency (measured by improved test scores) in   Left inferior frontal gyrus (IFG)   Left anterior temporal lobe (ATL)

60

4/18/12

16

61 62

  Gray matter density in language-related areas increases in as little as five months of instruction in country (even with a huge dialect difference)

  This increase correlates with the amount learned

  This again suggests brain growth stimulated by effective interaction with the second language

63

  Even a small amount of input in the target language changes its neurological representation and may also have behavioral consequences

  By focusing on HL learners we do not always tap into the depths of language attrition/incomplete acquisition

64

4/18/12

17

  Production can be used for preliminary data mining

  In assessing production, aim for a controlled setting   Video descriptions   Maps   Sentence completion   Elicited imitation

65

He wante- he go- he took from zeh garbage a cigarette, and, and zen he saw zeh police, said hello, and zen he, just, em, just, frew zeh garbage can- can, zen, eh, zeh rabbit, em, how it’s called…flowered his flowers, and zen he wanted to eat him, so he took a rope and went up, an- and zeh rabbit saw him, and he was wif scissors, so he cut ze- cut zeh rope, and zen he fell into zeh police…’s car. (So how did he notice the rabbit in the first place?) Because eh, zeh rabbit wan- eh, wer- because he flowered zeh, his flowers, uh, one, on- two drops went on him. (So where did the drops go?) One on his cigarette, and zeh, zeh fire, eh…not burned…blew out? And one on his nose.

66

  Speaker describes a map to a confederate who moves objects on the screen (Gómez Gallo et al. 2007)

  Speakers produce spontaneous instructions to the confederate

  Confederate does not give verbal feedback

67

  Mandarin Chinese, baseline: Beijing dialect   13 native speakers and 17 heritage

speakers of advanced proficiency in spoken Mandarin   However, five HL speakers do not have the

knowledge of formal registers

68

4/18/12

18

  Basic word order is SVO, but attributes including relative clauses precede the head noun

  Nouns occur with classifiers   Use of numerals with nouns is associated

with indefiniteness   Locative expressions: in most cases, locative

PPs appear before the VP   Serial verbs are widely used, and some

serialization correlates with the presence of the ba-construction 69

  Proxy: noun filling the gap after the relative clause when the real head precedes the relative clause.   For instance:

把 小 三角形 [RC角上 有 圆点的] 图形 移 到 北京

ba small triangle corner have ball de figure move to Beijing REAL HEAD RELATIVE CLAUSE PROXY

‘Move the small triangle with a ball on its corner to Beijing.’

70

  In Mandarin, numeral phrases include numerals and classifiers:

一 个 三角形 ‘a triangle’ yi ge sanjiaoxing NUM CLF NOUN

  Generally, numeral phrases are considered indefinite

  In our corpus, native controls used numerals less, esp. when the theme expression had modifiers (hence, was more likely to be definite)

  Heritage speakers show lack of awareness of this subtle semantic feature

72

4/18/12

19

  Controls use the ba-construction, which is problematic for heritage speakers

  Controls use prenominal relative clauses, heritage speakers use postnominal relatives

74

  Relative clauses (RC) precede the head noun

  Most native speakers strictly follow this rule

[RC角上 有 菱形 的 小] [Head Noun 正方形] corner has a diamond ADN small

square ‘a small square that has a diamond at its corner’

75

  Heritage speakers tend to put the relative clause after the head noun

在 北京，放一个大的三角形 边上 有一个点的 In Beijing, put a big triangle the hypotenuse has a dot ADN In Beijing, put a big triangle that has a dot on its hypotenuse

  Possible reasons:   Late planning in production, due to the overall

complexity of the theme description   Interference from English

76

4/18/12

20

Native speakers Heritage speakers

Proxy construction Yes No

Numerals Less likely More likely

Word order 1) ba-construction 2) Attributes before

head noun

1) Rigid SVO 2) Attributes after

head noun Verb complexity Verb compound Single verb

  The Fruit Cart experimental design is an effective method if eliciting production form heritage speakers in such a way that their output is well constrained

78

  Allows researchers to focus on the areas that may cause difficulty

  Borrow from the playbook of other fields: L1, L2, clinical populations, fieldwork experiments   Types of phenomena   Methodologies

79

  Grammaticality Judgment Tasks (GJTs)   Points of general concern:

  What is the exact nature of grammaticality?   Dichotomous or gradient?   What is the role of extragrammatical factors?

  Point of methodological concern:   Absence of rigorous control techniques

  An additional worry:   Heritage speakers show a notoriously high rate of

null responses on GJT (Polinsky 2006) 80

4/18/12

21

  Avoiding GJT (Schütze 1996, Tremblay 2005)

  Alternatives:   Response time   Rating or magnitude estimation   Eye tracking (Irina Sekerina’s work)

81

  Optimize the comprehension conditions for heritage speakers   Allows us to make sure we are not dealing

with bottleneck effects   Speed/add distractions to comprehension

conditions for the controls

82

  Self-paced reading (SPR), an established tool

(Just et al. 1982, Mitchell 2004)   Timing is regular except for areas of

difficulty   Problem: HL speakers have difficulty

reading, even if the alphabet is the same   How can one extend the SPR paradigm to

populations that do not read? 83

  Taking lessons from researchers for whom reading is irrelevant, inappropriate, or an unwelcome confound   Sign language research   Child language acquisition research   Research on clinical populations   Phonological investigations

84

4/18/12

22

  Sentence-picture matching (SPM), also well-established   (Bamber 1969, Carey & Lockhart 1973, Clark

& Chase 1972, Frost 1972, Seymour 1974, Shepard 1967, a.o.)

  Present acoustic stimuli and record response time for a stimulus-to-picture matching task

  Common in the fields of aphasiology and child language acquisition

85

  An unknown: Do SPR and SPM produce comparable results?

  Test case: Relative clause processing

86

  Subject relatives are easier to process   (SPR: Traxler et al. 2002; ERP: King & Kutas

1995; PET: Stromswold et al. 1996; fMRI: Just et al. 1996; Eye-tracking: Traxler et al. 2002…)

  Cross-linguistic advantage of subject relatives   (Dutch: Frazier 1987; German: Mecklinger et

al. 1995; Hebrew: Arnon 2005; Japanese: Miyamoto & Nakamura 2003; Korean: Kwon et al. 2006; Russian: Polinsky 2011…)

87

  Subject preference in the processing of relative clauses in Russian (Levy et al. 2007, submitted; Polinsky 2011, 2012)

  Subject and object RCs can have the same word order

NPi [whichNOM __i Verb NPACC] = Subject Relative

NPi [whichACC __i Verb NPNOM] = Object Relative

88

4/18/12

23

Tim

e in

m

s

Subject Object

Polinsky 2012; Polinsky & Fedorova in prep.

*

89

  Subjects see two pictures on computer screen followed by a sound file

90

Tim

e in

m

s

Subject Object

*

Polinsky & Fedorova in prep. 91

  Heritage studies can re-appropriate well-established paradigms from other experimental fields   Picture matching in lieu of SPR   Possible use of self-paced listening, also used

in L1 and L2 research (Marinis 2003)

92

4/18/12

24

  Heritage studies can re-appropriate well-established paradigms from other experimental fields

  Using visual world paradigms in general will provide rich results: more need for eye-tracking in HL studies

93

  From behavior to brain   From modules to interfaces   From homogenous sub-populations to

assessing variance

94

  ERP measures: brain/behavior dissociations

  Grammar   University students learning Spanish for

the first time

95

  Violations of tense (similar to English) Su abuela *cocinando/cocina muy bien His grandmother *cooking / cooks very well

  Violations of gender (no parallel in English) Ellos fueron a *un/una fiesta They went to a party

  Violations of number (English has it, but not here) *El/los niños están jugando The boys are playing

96

4/18/12

25

97 98

99

  University students again showed a brain/behavior dissociation:   Their acceptability judgment responses were

at chance   But their brain responses reliably differentiated

grammatical from ungrammatical sentences   In this respect, their brain responses

looked like those of native speakers

100

4/18/12

26

  NPIs are licensed by negation and negative-like contexts   Nobody expects Congress to ever change.   Voters expect Congress to ever change.   Few people expect Congress to ever change.   Voters doubt that Congress will ever change.

  But negation can’t be just anywhere, it must be structurally higher than the NPI   *The people [rel. cl. who can’t stand it] expect

Congress to ever change.

101

  No bills [that the democratic senators supported] will ever become law.

  *The bills [that the democratic senators supported] will ever become law.

  *The bills [that no democratic senators supported] will ever become law.

(German: Drenhaus et al. 2005; English: Xiang et al, 2009)

102

  U of Chicago undergraduates tested in Autism-spectrum Quotient (AQ)   Focus on communication subscale (known to

correlate negatively with pragmatic comprehension)

  Lower score better pragmatic inferencing   NPI Illusions: Lower AQ score ( pragmatic

skills) – more illusions   Agreement Illusions: no correlation with AQ

score

103

  Autism-spectrum Quotient (AQ) can be used on healthy HL populations to separate the syntax-semantics or syntax-pragmatics interface from purely grammatical phenomena

104

4/18/12

27

  Extreme variation with regard to proficiency in heritage speakers   C.f. three-stage model (Polinsky & Kagan 2007)   (i) Acrolectal HS: high proficient, near-native

speakers of Russian, maximally close to competent monoling

  (ii) Mesolectal HS: clear deficencies if compared to monolingual

  (iii) Basilectal HS: lowest-proficiency speaker, maximally removed from native attainment, may have never acquired literacy in Russian

105

  Rate of speech (measured in words per minute) may serve as a predictor of heritage speakers’ overall language proficiency

  Advantages:   Does not rely on literacy skills   A very simple measure

  Disadvantages:   More proof of the concept needed   Unclear what RoS actually reflects

106