Volume 1 | Issue 1 | Volume 3 | Issue 2| November 2019 · 2(b) and 2(d) show this for necessives, impersonal passives and possessives. All of these latter verbs take different subject

Volume 1 | Issue 1 | Volume 3 | Issue 2| November 2019

LingUU Volume 3 | Issue 2 | November 2019

Editors-in-chief Senior editorMandy de Kuiper & Myrthe van der Veen Anique Schüller

EditorsAnne Beltman, Dylan Bonga, Bjorn Lichtenberg, Levi Remijnse & Nadine Roos

Secretary Finances & PR Layout & Website Dylan Bonga Levi Remijnse Tessa Portier & Nadine Roos Advisory board Fieneke Jochemsen, prof. dr. René Kager, dr. Marijana Marelj, dr. Iris Mulders, Sophie Slaats & prof. dr. Henriette de Swart

Contact ISSN 2589-286X (print)[email protected] | linguujournal.nl ISSN 2589-2878 (online)

Please visit our website if you want to publish in LingUU, if you want to become a reviewer or if you want to subscribe. Contact Bjorn Lichtenberg for inquiries con-cerning PR. All other communications should be addressed to Iris van der Wulp.

Thanks toour anonymous reviewers BA Linguistics Utrecht University UiL-OTS Utrecht Institute of Linguistics

AboutLingUU is the official, peer reviewed, student journal of Linguistics at Utrecht Uni-versity (UU). In this journal students, both undergraduate and graduate, as well as PhD-students, can publish their papers.Students sometimes come up with ideas that are worth sharing, and develop creative theories that more people should read, other than just the teacher. LingUU provides a way to make this happen, while giving students the opportunity to get to know the world of academic pub-lishing small-scale. Apart from the author-side of publishing, the journal offers students the possibility of developing the skill of peer reviewing in a real setting. The journal aims to feature articles from the different sub disciplines of Linguis-tics and publishes in both Dutch and English. Apart from research articles, LingUU features articles on internships and studying abroad, and book notices or reviews.

Welcom

e to our new board m

embers! Vanessa Kuhn (Senior Editor), Roos Kerrebijn, Liesje van der Linden, M

arian Marchal and Vera N

ijland ( Editors), Iris van der Wulp (Secertary), Moja Kloos (Finances) and Tessa Portier (Layout and website)

Contents

Research There and back again: An accusative’s journey for case 4 Walther Glödstaf

Comprehension of reflexive pronouns in language impaired children 14with ASD: the acquisition of the locality constraint

Sanne Ditewig

Phonetic Convergence of Students from Limburg in Utrecht: 25(De)voicing of /v/ and Hard and Soft /ɣ/

Judith Brinksma & Nelleke Jansen

Meertaligheid binnen Brabant 37Milou Leeijen

Prepositions and language transfer 50Dylan Bonga

Guest paper

A Natural Language Processing based text analysis of populist rhetoric 59 in social media text messages Nizar Hirzalla

Intern Extern Taal en psychiatrie 75 Janna van Egmond

Book reviewThe Infinite Gift by Charles Yang: Een desillusie voor de ouder, maar 77een ode aan het kind

Roos Kerrebijn

The Daily Linguist 81

4 LingUU | 3.2 | 2019 Research

There and back again: An accusative’s journey for caseA Phase Impenetrability Condition-consistent account of accusative variation in Finnish rationale adjuncts

W.A. (Walther) GlödstafPhD in Linguistics, University of Illinois, Urbana-Champaign (United States) Manuscript written during his RMA Linguistics, Utrecht University, Utrecht

KEYWORDSFinnishLDCAPICcase realisation variabilityrationale adjunctsrelative clauses

ABSTRACTA previous analysis of long distance case assignment on Finnish direct ob-jects was claimed to violate the Phase Impenetrability Condition (Brattico, 2014) due to the matrix verb’s T clearly influencing the case assignment (Brattico & Vainikka, 2014) inside rationale adjuncts. A raising analysis will be proposed here to show that a Phase Impenetrability Condition-consistent analysis is possible, as data from relative clauses show that movement of the direct object to the edge of the phase would explain the lack of case form variation in relative pronouns in contrast to non-relative clauses.

1. IntroductionFinnish rationale adverbial clauses (‘in order to verb’), vary which form the accusa-tive case of a singular non-pronominal direct object takes, depending on whether the matrix clause’s verb has morphology showcasing overt-phi-agreement (hence-forth OPA).1 As shown below, when the matrix verb has OPA, the accusative can only be realised as ACC(n) (1a). When it lacks OPA (1b), variation between ACC(n) and ACC(0) is possible. ACC(n) refers to the accusative form that is homophonous with the genitive case and identified by the '-n' case suffix, while ACC(0) refers to the accusative form homophonous with the nominative and identified through a ‘zero’ case suffix.

(1) a. Frodo ja Sam. läht-i-vät löytää-ksee-n Frodo and Sam.NOM left-PAST-3P find-KSE2-Px3S/P3 Bilbon/*Bilbo. Bilbo.ACC(0)/ACC(n) ‘Frodo and Sam needed to leave in order to find Bilbo.’

1 This is a descriptive label for the morphological correlation between verbs and accusative form, not a claim about a theory of phi-agreement!2 Marks rationale adjuncts through the ‘-kse’ affix that identifies it.3 Possessive-suffix specified for person and number (here the affix is shared between 3rd singular and plural).

5Walther Glödstaf

b. Frodon ja Samin. täyty-i lähteä löytää-ksee-n Frodo and Sam.GEN need-PAST leave find-KSE-Px3S/P Bilbo/ Bilbon. Bilbo.ACC(0)/ACC(n) ‘Frodo and Sam needed to leave in order to find Bilbo.’

2. Brattico (2014)’s AnalysisAs introduced above, rationale adjuncts in Finnish alternate between ACC(0) and ACC(n) case when the matrix verb lacks OPA. Verbs that lack OPA are necessives (need to verb), possessives (I have X), and impersonal passives. The presence or absence of OPA, resulting in ACC(n) and ACC(0) respectively, also holds in simple clauses. 2(a), 2(b) and 2(d) show this for necessives, impersonal passives and possessives. All of these latter verbs take different subject cases including nominative. Thus, nominative case not having been assigned on the subject (Jahnsson’s Rule (Jahnsson, 1871)) cannot be the explanation for the variation (Brattico & Vainikka, 2014).

(2) a. Bilbo ja Gandalf löys-i-vät tie-n. Bilbo ja Gandalf.NOM find-PAST-3P road.ACC(n) ‘Bilbo and Gandalf found a/the road.’ b. Bilbon ja Gandalfin täyty-i löytää tie. Bilbo and Gandalf.GEN need-PAST find road.ACC(0) ‘Bilbo and Gandalf needed to find a/the road.’ c. Me löydett-iin tie. We.NOM find.IMPASS4-PAST road.ACC(0) ‘We found a/the road.’ d. Bilbolla ja Gandalfilla ol-i kiire löytää tie. Bilbo and Gandalf.ADE have.PAST hurry.ACC(0) find road.ACC(0) Literally: ‘Bilbo and Gandalf possessed a hurry to find a/the road.’

This case assignment is analysed as a two-step process of ‘First Agree’ and ‘Second Agree’ (Brattico, 2014).5 ‘First Agree’ is the agreement process where the unvalued phi-features on T are valued and the subject receives its case and the partitive-ac-cusative distinction on the DO based on telicity is made. During ‘Second Agree’, the probe from T with its now valued phi-features travels further, affecting other active goals in its path, as per the Chomsky-Hiraiwa Multiple-Agree hypothesis (Chomsky, 2008; Hiraiwa, 2005). This means that “Agree does not stop reading the structure when it finds the first goal; instead it keeps affecting active goals on its path” (Brat-tico, 2014, p. 325). Active goals here are for instance DPs lacking structural case. The process would thus look as below for sentence (2a) (Brattico, 2014).

4 Impersonal passive5 A reviewer remarked this proposal is already made in Bejar & Rezac (2009). While the models might work together, they are only similar by using cycles, since Bejar& Rezac (2009) is not compatible with a multiple-agreement-hypothesis (Chomsky, 2008; Hi-raiwa, 2005)). It is also not clear what type of probe Bejar & Rezac (2009) would use since the OPA feature is about the absence/presence of morphology - not feature values.


(i) Start of the derivation T.[uφ] Bilbo ja Gandalf.[uCASE] v.[ uφ] löysi- tie-.[uTelicity & uCASE] T.[uφ] Bilbo and Gandalf.[uCASE] v.[ uφ] find.PAST road.[uTelicity & uCASE] ‘Bilbo and Gandalf found the road.’ (ii) ‘First Agree’ T.[3P] Bilbo ja Gandalf.NOM v.[3P] löysivät tie-.[+Telicity & uCASE] T.[3P] Bilbo and Gandalf.NOM v.[3P] find.PAST.3P road.[+Telicity & uCASE] ‘Bilbo and Gandalf found the road.’ (iii) ‘Second Agree’ T.[3P] Bilbo ja Gandalf.NOM v.[3P] löysivät tien T.[3P] Bilbo and Gandalf.NOM v.[3P]find.PAST.3P road.ACC(n) ‘Bilbo and Gandalf found the road.’

Since the DO is an active goal until ‘Second Agree’ and sensitive to the OPA on T, Brattico (2014) assumes that there are two different phi-feature types in Finnish; one overt ([+OPAφ]) and the other null ([-OPAφ]). Therefore, due to the accusative case form’s sensitiveness to whether the verb has or lacks OPA, ‘Second Agree’ can conceptually be explained as the DO having an unvalued phi-feature. The feature’s subsequent valuation as either the [+OPA] or [-OPA] type during ‘Second Agree’ determines the accusative form. Now that we understand how accusative case assignment works in simple clauses, let us turn to how the process works in rationale clauses to explain the accusative variation we find it in it.

As explained before, ACC(0)/ACC(n) variation in rationale adjuncts only occurs when the matrix verb lacks OPA. On the surface this is easy to explain. When the matrix verb has OPA, its T and the rationale adjunct’s T can both assign ACC(n) only as the case form in (3a). Matrix T’s OPA is signaled through the ‘Ø’ third per-son singular ‘regular’ phi-morphology. The rationale adjunct’s T’s OPA is signaled through the third person singular/plural possessive suffix that denotes the ana-phoric little pro (van Steenbergen, 1991).

(3) a. Bilbo petkutt-i-Ø Klonkku-a löytää-ksee-n. Bilbo.NOM deceive-PAST-3S Gollum-PAR find-KSE-Px3S/P tie-n ulos luolista. road-ACC(n) out caves.ELATIVE ‘Bilbo deceived Gollum in order to find the road out of the caves.’

However when the matrix verb lacks OPA, then ACC(0)/ACC(n) variation ensues.

(3) b. Bilbon täyty-i petkuttaa Klonkku-a löytää-ksee-n. Bilbo.GEN need-PAST deceive Gollum-PAR find-KSE-Px3S/P tie-n ulos luolista. road-ACC(n) out caves.ELATIVE ‘Bilbo needed to deceive Gollum in order to find the road out of the caves.’

7Walther Glödstaf

Brattico (2014) explains this variation as an effect of partial intervention (Starke, 2001). Starke (2001)’s intervention framework works as follows: When the probe and the goal are intervened by functional projections bearing all the same fea-tures, a full intervention ensues. When only a subset of these features is present, a partial intervention ensues, resulting in the probe sometimes being intervened and sometimes not. If none of these features are present then the probe always matches with the goal.

Matrix T’s probe is in Brattico (2014)’s analysis a bundle of polarity- phi-, and full tense features. Full tense is defined as “a system where a single element has at least two overt tense forms, past and present” (Brattico, 2014, p. 318). They are motivated by the phi-feature determining the Accusative form, whereas the polar-ity feature assigns partitive case for instance in negations, which in Finnish act as auxiliaries specified for phi-features, while the main verb specifies the tense as below.

(4) Frodo ei-Ø tuho-nnut sormus-ta. Frodo.NOM no-3S destroy-PAST ring-PAR ‘Frodo did not destroy the ring.’

The full-tense feature is responsible for the partial intervention in the rationale ad-junct, since as shown below, it can host a negative-polarity item (such as ’hardly’ (in Finnish 'tuskin')).

(5) Frodo täyty-i paeta Konnusta tuskin anta-ksee-n. Frodo.NOM need-PAST flee Shire.ELA hardly give-KSE-Px3S/P sormus/sormuksen Sauronille vaan tuhota-ksee-n ring.ACC(0)/ACC(n) Sauron.ALLATIVE but destroy-KSE-Px3S/P se/ se. it.ACC(0)/ACC(n) ‘Frodo needed to flee the Shire, hardly in order to give Sauron the ring but to destroy it.’

Through the possessive-suffix in the rationale adjunct we also know it has OPA, but what it lacks, is full-tense. Therefore, the probe from matrix T and the active DO goal inside the adjunct would only be intervened by polarity- and phi-features resulting in a partial intervention and variation. Further proof of this variation is that a Partitive ACC(n) variation can be induced by having a negated neccessive as the matrix clause’s verb.

(6) Frodo ei-Ø men-nyt Mordoriin antaa-ksee-n. Frodo.NOM no-3S go-PAST Mordor.ILL give-KSE-Px3S/P ?sormusta/ sormuksen Sauronille. ring.?PAR/ACC(n) Sauron.ALLATIVE ‘Frodo did not go to Mordor in order to give the ring to Sauron.’


The proposal is that when variation occurs, ACC(0) is licensed by matrix T whose phi-features are valued as their null variant, while ACC(n) is licensed through the adjunct internal T’s overt variant through the possessive-suffix as in Table 1 below.

Table 1

Type of Phi-feature valuation Matrix T Adjunct T6 DO Case

Overt ACC(n)ACC(n)

ACC(n)Covert ACC(0) ACC(0)/ACC(n)

The prediction of this analysis is thus that in a construction where matrix T’s probe and an active DO goal are intervened by polarity-, phi-, and full tense features, a matrix verb lacking OPA should not be able to influence the case of a DO further down the syntactic structure (Brattico, 2014). As shown below, this is the case for instance in temporal adjectives (Brattico, 2014) that carry phi-features in the form of a possessive-suffix (7a), can host a negative polarity item (7b), and h ave full tense features as the past/present alternation between (7a) and (7c) show.

(7) a. Sam ilahtu-i-Ø näh-tyää-n haltia. Sam.NOM rejoice-PAST-3S see-TEMP(PAST)-Px3S elf.ACC(n) ‘Sam rejoiced after seeing an elf.’ b. Sam ilahtu-i-Ø tuskin näh-tyää-n haltian. Sam.NOM rejoice-PAST-3S hardly see-TEMP(PAST)-Px3S elf.ACC(n) ‘Sam rejoiced hardly after seeing an elf.’ c. Sam ilahtu-i-Ø nähd-essää-n haltian. Sam.NOM rejoice-PAST-3S see-TEMP(PRES)-Px3S elf.ACC(n) ‘Sam rejoiced while seeing an elf.’

Yet, as shown below, even when the matrix T lacks OPA, ACC(0) is an ungrammati-cal form for the DO inside the temporal adjunct.

(8) Samin täyty-i ilahtua nähd-essää-n haltian/* haltia, Sam.NOM need-PAST rejoice see-TEMP(PRES)-Px3S elf.ACC(n)/*ACC(0 tai Frodo petty-i-si-Ø. or Frodo.NOM disappoint-PAST-COND-3S ‘Sam needed to rejoice while seeing an elf or Frodo would become disappointed.’

Similarly, manner adverbials lack polarity-7, phi-8 , and tense features as shown in (9). They are thus ‘transparent’ for the probe (Brattico, 2014).

6 The rationale Adjunct T always has OPA obligatorily.7 Or at the least the sentence seems ‘odd’ with one (Brattico, 2014:320).8 Some native Finnish speakers in Helsinki interviewed by the author spontane-ously produce overt phi-morphology on the manner adverbial. The data here is presented as in Vainikka & Brattico (2011:49) and Brattico (2014).

9Walther Glödstaf

(9) a. ?*Gandalf viehett-i-Ø lapsia tuskin ampu-malla. Gandalf.NOM charm-PAST-3S child.PAR hardly shoot-MA ilotulitteen. firecracker.ACC(n) *‘Gandalf charmed the children hardly by shooting a firecracker.’ b. *Gandalf viehätt-i-Ø lapsia ampu-mallaa-*n. Gandalf.NOM charm-PAST-3S child.PAR shoot-MA-*Px3S/P ilotulitteen. firecracker.ACC(n) ‘Gandalf charmed the children by shooting.Px3S/P a firecracker.’

Further the adjunct internal DO’s Case is always decided according to whether the matrix verb has or lacks OPA or is negated. Predictably when the matrix verb has OPA, the DO can only be ACC(n) (10a). When the matrix lacks OPA, the DO can only be ACC(0) (10b), and when the matrix verb is negated, the DO is always assigned Partitive Case (10c).

(10) a. Gandalf viehätt-i-Ø lapsia ampu-malla Ilotulitteen. Gandalf.NOM charm-PAST-3S child.PAR shoot-MA firecracker.ACC(n) ‘Gandalf charmed the children by shooting a firecracker.’ b. Gandalf täyty-i viehättää lapsia ampu-malla lotulite. Gandalf.NOM need-PAST charm child.PAR shoot-MA firecracker.ACC(0) ‘Gandalf charmed the children by shooting a firecracker.’ c. Gandalf ei-Ø viehättä-nyt lapsia ampu-malla ilotulitetta. Gandalf.NOM no-3S charm-PAST child.PAR shoot-MA firecracker. PAR ‘Gandalf did not charm the children by shooting a firecracker.’

Given this evidence, it is tempting to assume that the accusative variation is caused by partial intervention and violates PIC, since matrix T’s probe travels into the adjunct to give the DO case and we have no proof the DO was ever at the phase edge for this process. This analysis thus predicts that all DOs in rationale adjuncts should exhibit this accusative variation. But as will be shown in the next section, relative pronouns do not conform to this prediction.

3. CounterevidenceThis section will provide counter evidence to the analysis in Brattico (2014), by showing that relative pronouns lack the expected accusative form alternation in rationale clauses. These constructions are constructed with a main clause, modi-fied by a relative clause consisting of a rationale clause and the rationale clause’s licensing matrix clause. Even when the main clause and matrix clause’s verb lack OPA (11a) and are negated (11b), the rationale clause’s DO (the relative pronoun) always surfaces as ACC(n). Instead, according to the analysis in Brattico (2014), it should show variation or surface as a partitive respectively, both of which are ungrammatical. For ease of reference, the clausal structure of the following ex-


amples will be indicated as follows: the main clause of which the DO is modi-fied by the relative pronoun is marked in bold. The rationale clause is marked in underlined, and the matrix clause licensing the rationale clause is marked as underlined bold.

(11) a. Frodon täyty-i tuhota sormus, jonka/*joka Frodo.GEN need-PAST destroy ring.ACC(0) which.ACC(n)/*ACC(0) saada-ksee-n Bilbon täyty-i petkuttaa Klonkua. get-KSE-Px3S/P Bilbo.GEN need-PAST cheat Gollum.PAR ‘Frodo needed to destroy the ring which in order to get Bilbo had to cheat Gollum.’ b. Frodon ei-Ø täyty-nyt tuhota sormusta, jonka/*jota Frodo.GEN no-3S need-PAST destroy ring.PAR which.ACC(n)/*PAR saada-ksee-n Bilbon ei-Ø täyty-nyt petkuttaa Klonkkua. get-KSE-Px3S/P Bilbo.GEN no-3S need-PAST cheat Gollum.PAR ‘Frodo did not need to destroy the ring which in order to get Bilbo did not have to cheat Gollum.’

As per Huhmarniemi (2012) and Huhmarniemi and Brattico (2013), we also know that Finnish relative pronouns start head-externally in accordance with Smith (1964) and Chomsky (1965). This means that the head of the relative clause (the DP being modified by it, in our case ring), was never inside the relative clause (Huhmarniemi & Brattico, 2013). The relative pronoun is thus a DO comparable to one in a non-relative rationale adjunct, as both have the same base position.

Evidence for the head-external analysis is the so called ‘snowballed pied-piping’ (Huhmarniemi, 2012) that is evident in (12). In a regular head-external analysis of relative clauses, the relative pronoun moves to the front of the relative clause via wh-movement as below (Huhmarniemi & Brattico, 2013).

(12) Elrond called [DP a council [CP[+WH] Pippin attended a council in secret]] Elrond called [DP a council [CP[+WH] which1 Pippin attended a council1 in secret]]

For (11a), the same operation applies, but now we get much more radical pied-piping, that results in the linear sentence order between the rationale adjunct and its matrix clause being reversed. The DO targeted for wh-movement, therefore, moves to the edge of each phase and moves everything below itself with it at each movement step. The operation starts off with (13a). The first step is to move the DP to the edge of the rational adjunct (CP2) as in (13b). Then, the entire rational adjunct is moved to the edge of its matrix clause (CP1) as in (13c), where both the rational adjunct and its matrix clause are attached to the DP and delete the +WH-feature that triggered the movement.

11Walther Glödstaf

(13) a. Frodon täytyi tuhota [DP sormus, [CP1[+WH] [C1 Bilbon täytyi Frodo needed to destroy [the ring [Bilbo needed petkuttaa Klonkua [CP2 saadakseen jonka]]]] to cheat Gollum [in order to get which]]] b. Frodon täytyi tuhota [DP sormus, [CP1[+WH] [C1 Bilbon täytyi Frodo needed to destroy [the ring [Bilbo needed petkuttaa Klonkua [CP2 jonka1 [C2 saadakseen ______1]]]]] to cheat Gollum [which in order to get ]]] c. Frodon täytyi tuhota [DP sormus, [CP1[+WH] [CP2 jonka1 Frodo needed to destroy [the ring [ which [C2 saadakseen 1]]2 [C1 Bilbon täytyi petkuttaa Klonkua]] 2] [in order to get]] [Bilbo had to cheat Gollum.]]

So far, the analysis for Finnish LDCA presented here has not taken PIC into con-sideration. Yet I will argue in the next section that the difference in case variation between relative pronouns and other DOs in rationale clauses merits a considera-tion of phases. For if we take phases into consideration, then the DO would only be able to alternate in rationale adjuncts if it were able to move to the edge of the rational adjunct phase and from there be visible to the probe from matrix T. Consequently, if the DO is forced to stay in its base position, unable to move into the phase-edge, then no variation should be possible.

4. Towards a PIC-consistent AnalysisRecent developments in how to account for the free word order in Finnish by Brat-tico (2018) help explain why the DO has to stay low in precisely those environ-ments where the DO is being targeted for wh-movement. According to Brattico (2018), Finnish sentences are built in three stages that follow each other in a fixed order:

Stage 1 is the stage where Narrow Syntax (thus Case assignment) takes place.

Stage 2 satisfies information structural needs that allow Finnish to have free word order, by attaching arguments left-/rightwards as adjuncts, to satisfy the information structure.

Stage 3 is where A-bar movement such as wh-movement takes place and only once stage 3 is completed, the structure is sent to LF and PF.

Evidence for this distribution is, for instance, that informational structural move-ment such as topicalization or focalization (roughly indicating that certain infor-mation is new or old) does not change the case of arguments. Hence case must be assigned before stage 2. Likewise, as shown previously, wh-movement results in pied-piping phenomena, meaning the word order inside the pied-piped constitu-


ents must be fixed before wh-movement is initiated. Stage 2 operations, however, change the word order significantly so wh-movement must happen after it.

If we assume phases exist, both the +WH-probe and the probe from matrix T would target the same position, namely the edge of the adjunct phase. Thus if a DO were to move to the edge of the adjunct phase during stage 1 to receive case, it would also match with the +WH-probe and start stage 3 prior to stage 2, which is prohibited. My suggestion is therefore that the lack of variation in envi-ronments where the DO of the rationale adjunct is wh-moved, is a sign that when ACC(n)/ACC(0) variation occurs, the DO can move to the phase edge to receive ACC(0), as expected by Phase theory. If it is prohibited to move there, we expect it to only be able to receive ACC(n), as shown by the data in the last section. Likewise, when the DO is WH-moved in questions, an ACC(0) marked question-word is marginal.

(14) Minkä/ ??Mikä pelastaa-ksee-n Frodon täyty-i What.ACC(n)/??ACC(0) rescue-KSE-Px3S/P Frodo.GEN need-PAST tuhota sormus? destroy ring.ACC(0) ‘What in order to save did Frodo have to destroy ring?’

This, however, does not explain why the DO would move to the phase edge to re-ceive ACC(0). After all, it has already received ACC(n) adjunct-internally; satisfying the case filter. As explained in section 2, the rationale adjunct’s T has OPA but lacks full-tense features. Usually in Finnish, having OPA entails having full-tense features. Therefore, I suggest that the adjunct internal case assignment is ‘weak-er’ due to this and sometimes fails, leading to the DO being moved to the edge of the phase as a last-resort mechanism to avoid a case filter violation. However, it would result in a preverbal DO word order prior to information structure being applied, which, while not entirely ungrammatical, is lyrical and marked. But it can be solved by appealing to stage 2’s domain over information structure, because there are many processes in Finnish and other languages where marked struc-tures are used, such as diary pro-drop in English or adjective-noun word order reversals in Finnish songs.

Accusative variation in rationale adjuncts is thus summarized as follows: During stage 1, when Case is assigned, the DO is assigned ACC(n) adjunct-internally. However, because the adjunct lacks a full tense feature, this valuation fails some-times and the DO is moved to the edge of the adjunct phase. There, it can be seen by matrix T’s probe, which, if the matrix verb lacks OPA, assigns ACC(0). This results in a lyrical preverbal word-order, so that during stage 2, the preverbal DO is moved back to its postverbal position in neutral contexts. Similarly, DO’s targeted by WH-movement stay low and thus do not showcase variation.

13Walther Glödstaf

5. ConclusionIn conclusion, this paper proposed a PIC-consistent analysis of accusative vari-ation in Finnish rationale clauses. The motivation for the analysis was to explain the lack of the expected variation in environments where the DO was the target of wh-movement, such as in relative clauses. In order to account for this data, it was proposed that the variation in accusative case is PIC-consistent, since in this analysis, the DO would have to raise to the phase edge in order to receive ACC(0). Lack of ACC(0) in environments where the DO was the target of wh-movement was thus explained through the three stages of Finnish sentence compilation, as proposed in Brattico (2018), where wh-movement is the last stage of the compiler. Moving the DO to the phase edge, where it could be targeted by wh-movement would have triggered the last stage prematurely in relative clauses. Therefore, lack of variation is caused by the DO being forced to stay low.

The DO’s movement to the phase edge outside of relative clauses was motivated through occasional failure of the case assignment inside the adjunct, as the adjunct lacked the usual full-tense feature that accompanies overt phi-morphology, which go-verns whether ACC(0) or ACC(n) is assigned. Failure to assign ACC(n) thus would lead to the raising of the DO. Since this is only occasional, it produces a variation pattern.

The next steps for this analysis would be to see if it can be generalized across LDCA in Finnish and whether similar phenomena of back-and-forth movement can be found in related languages such as Hungarian and Estonian that are also hy-pothesized to share some of the underlying three-stage-compilation machinery. ■

Received April 2018; accepted April 2019.

ReferencesBejar, S., & Rezac, M. (2009). Cyclic Agree. Linguistic Inquiry 40 (1), pp. 35-73.Brattico, P. (2014). Long Distance Case Assignment and Intervention. Lingua 148, pp. 309-336.Brattico, P. (2018). Word Order and Adjunction in Finnish. Charleston, SC.: Aquila & Celik.Brattico, P., & Vainikka, A. (2010). The Finnish Accusative: Long Distance Case Assignmentby φ-Agreement submitted for publication. Retrieved from: research.uni-leipzig.de/lomo/

literatur/vainikka-brattico2010.pdfChomsky, N. (2001). Derivation by Phase. In M. Kenstowicz (ed.) Ken Hale: A Life in Lan-

guage. Cambridge, MA: MIT Press, pp. 1-52.Chomsky, N. (2008). On phases. In: Freidin, R., Otero, C., Zubizarreta, M.-L. (Eds.), Founda-

tional Issues in Linguistic Theory: Essays in Honor of Jean-Roger Vergnaud. MIT Press, Cambridge, MA, pp. 133-166.

Hiraiwa, K. (2005). Dimensions of Symmetry in Syntax: Agreement and Clausal Architecture(Dissertation). MIT.Huhmarniemi, S. (2012). Finnish A -movement: Edges and Islands. Studies in Cognitive

Science 2. Helsinki: University of Helsinki.Huhmarniemi, S., & Brattico, P. (2013). The Structure of the Finnish Relative Clause. Finno-

Ugric Languages and Linguistics 2(1), pp. 53-88.Starke, M. (2001). Move Dissolves into Merge: A Theory of Locality (Ph.D. thesis). University of Geneva.Steenbergen van, M. (1991). Lond-distance binding in Finnish. In: Long-distance anaphors,

ed. by J. Koster and E. J. Reuland. Cambridge: Cambridge University Press, pp. 231-244.


Comprehension of reflexive pronouns in language impaired children with ASD: theacquisition of the locality constraint

S. (Sanne) DitewigManuscript written during her RMA Linguistics, Utrecht University, Utrecht

KEYWORDSAutism Spectrum Disorderreflexiveslanguage comprehensionbinding

ABSTRACTAlmost 3% of children aged four to twelve are diagnosed with Autism Spectrum Disorder (ASD) in the Netherlands. Some of these children have language impairments, referred to as Autism Language Impaired (ALI). Studies by Perovic et al. (2013a; 2013b) show that the comprehension of reflexive pronouns is impaired in these children. It is hypothesized that children with ALI do not have full knowledge of the constraints on the binding of reflexives, known as Principle A (Chomsky, 1986). However, the nature of this impairment remains unknown, as only the c-command constraint of Principle A has been tested. The locality constraint of Princi-ple A remains uninvestigated. This article proposes methodology to bridge this gap in knowledge and answer the following question: 'Do children with ALI have knowledge of the locality constraint on the interpretation of reflexives?' It is expected that children with ALI lack all knowledge of principle A. The proposed research will result in valuable insights into the linguistic symptomatology of children with ALI, and thus has clinical impli-cations. Moreover, it has a theoretical contribution by implicitly examining the notion of binding and its underlying constructs, through investigat-ing which syntactic constraints are impaired in children with ALI and thus group together.

1. IntroductionAlmost 3% of children aged four to twelve are diagnosed with a form of Autism Spectrum Disorder (ASD) in the Netherlands9. This population is highly hetero-geneous, but a distinction is generally made between ASD individuals with pre-served language abilities, known as Autism Language Normal (ALN) and ASD in-dividuals with language impairments, known as Autism Language Impaired (ALI) (First argued for by Kjelgaard & Tager-Flusberg, 2001). The latter group shows symptomatology included in the general autism criteria in the DSM-5 (e.g. social interaction deficits and repetitive behavior without the presence of global devel-opmental delays; see American Psychiatric Association, 2013). Children with ALI

9 As found by the Centraal Bureau voor Statistiek, in their yearly report on the wellbeing of children aged 4-12 in the Netherlands: CBS. (2018). Ervaren gezondheid, gebruik en leefstijl bij kinderen tot twaalf jaar. Retrieved from:http://statline.cbs.nl/Statweb/publication/?VW=T&DM=SLNL&PA=83716ned&D1=12&D2=a&D3=0&D4=0&D5=0-3&HD=180627-1313&HDR=T&STB=G1,G2,G3,G4.

15Sanne Ditewig

can language-wise be distinguished from children with other developmental disor-ders that lead to language problems, such as Developmental Language Disorder (DLD), mainly because children with ALI show a more severely impaired ability to use gestures and have greatly impaired language comprehension. Children with ALI have better language production than comprehension, while children with DLD and Specific Language Impairment (SLI) do not show this dissociation (e.g. Paul, Chawarska & Volkman, 2008).

Most research on children with ALI is concerned with pragmatic and social impair-ments, whereas little is known about other linguistic subsystems and the causes of their problems with language comprehension. However, there has been a re-cent interest in the (morpho)syntactic abilities of children with ALI. For example, difficulties have been found with the comprehension of grammatical morphemes marking tense (Eigsti & Bennetto, 2009; Roberts, Rice & Tager-Flusberg, 2004). In addition to this, there are indications that children with ALI have impaired com-prehension of reflexive pronouns (Perovic, Modyanova & Wexler, 2013a; 2013b). Perovic et al. (2013a; 2013b) argue that this is caused by a ‘Principle A deficit’; children with ALI do not have complete knowledge of the binding Principle A (as proposed by Chomsky, 1986), which restricts the binding of reflexive pronouns. Reflexives require local, c-commanding and agreeing antecedents. However, Per-ovic et al.’s (2013a; 2013b) research only tested children with ALI’s knowledge of the c-command constraint on the binding of pronouns, and did not investigate the locality constraint. There thus remains a gap in knowledge on the nature of the impaired comprehension of reflexive pronouns in children with ALI: 'Do children with ALI have a locality constraint on the interpretation of reflexives?'

This paper proposes research to answer this question. It is expected that children with ALI do not have any knowledge of Principle A, including the locality constraint. The proposed research will provide insights into the linguistic symptomatology of children with ALI, and could thus have clinical implications for treatment. It could both help in clinical assessment and diagnosing ALI, and in indicating what sort of treatment regarding reflexives is feasible. Moreover, this research will give us insights in typical development and even deliver a theoretical contribution by im-plicitly examining the notion of binding and its underlying theoretical constructs, c-command and locality, through investigating which syntactic constraints are im-paired in children with ALI and thus group together.

This article is organized as follows: In section 2, the binding of reflexives and its acquisition in typically developing children and in children with ALI are discussed. Section 3 contains some methodological considerations, followed by a descrip-tion of the research question and hypotheses in section 4. Section 5 includes an overview of the proposed methodology for this study, followed by a discussion of the analysis and anticipated results in section 6. This leads to the conclusion in section 7.


2. The binding of reflexives and its acquisition2.1 Principle A and its acquisition in typically developing children

As discussed in the introduction, the comprehension of reflexives is restricted. Ac-cording to Chomsky (1986), in the Universal Grammar framework, reflexives are bound by their antecedent. The binding of reflexives is governed by Principle A and they require local, c-commanding and agreeing antecedents (Chomsky, 1986).

(1) John’s dad washes himself.

(2) John wants Peter to paint himself.

The c-command constraint on binding can be explained using example (1). In this sentence, himself must refer to John’s dad and not to John, because John does not c-command the reflexive. Example (2) illustrates the locality constraint on the binding of reflexives. Here, the reflexive himself is not bound by the c-commanding NP John, because this antecedent is not in the same clause as the reflexive. Him-self has to be bound locally, by the c-commanding NP Peter.

The current study is concerned with the Dutch language. Since the 1990’s several studies have shown that Dutch children have knowledge of the constraints on the interpretation of reflexive pronouns from around the age of four (i.e. Bergmann, Paulus & Fikkert, 2009; Sigurjónsdóttir & Coopmans, 1996; Spenader, Smits & Hendriks, 2009; Van Koert, Hulk, Koeneman & Weerman, 2013; Van Rij, Van Rijn & Hendriks, 2010) .

(3) Bert zegt dat Ernie zich/hem krabt. Bert says that Ernie himself/him scratches. ‘Bert says that Ernie scratches himself/him.’

For example, Sigurjónsdóttir & Coopmans (1996) used a truth-value judgement task with sentences like the one in (3). Their results showed that four- and five-year-olds had the right interpretation of 73% and 78% of the sentences, respec-tively, and six-year-olds had adult-like interpretations for all stimuli.

2.2 The acquisition of Principle A in ALI childrenThe acquisition of Principle A in children with ALI has only been studied experi-mentally by Perovic et al. (2013a; 2013b), as far as I am aware. In their first study on binding in children with ASD, Perovic et al. (2013a) tested fourteen children with ALI, aged six to sixteen. A picture selection task (PST) was used. Children saw a screen with two pictures on it, heard a sentence, and were asked to point to the picture that showed the meaning of the sentence. An example of the experimental stimuli used is the following:

(4) Bart’s dad is pointing to himself.

17Sanne Ditewig

As sentence (4) illustrates, knowledge of Principle A was tested by providing sen-tences with reflexives and possessive subjects. As explained in section 2.1, under Principle A the embedded NP Bart in (4), inside the subject Bart’s dad, does not bind the reflexive, because Bart does not c-command himself. The children heard a total of eight sentences containing reflexives. The results showed that the chil-dren performed at chance level. This means that the children with ALI wrongly chose the embedded NP as the antecedent of the reflexive around 50% of the time. They made considerably more mistakes than the control group with typically developing (TD) children.

Based on the results of their first study, Perovic et al. (2013b) conducted a follow-up study. This study included more syntactic phenomena than just reflexives; the comprehension of raising and object control structures was included as well. 48 children with ASD were tested in total. 26 of these children were categorized as ALI and 22 as ALN, based on their scores on language production and perception tests. The participants completed exactly the same PST as described for Perovic et al.’s (2013a) earlier study. The results of the follow-up study showed that the children with ALI performed around chance level, just as in Perovic et al.’s (2013a) earlier study. The ALN children, however, behaved like the control group of TD children and did not show a Principle A deficit. Thus, as expected, children with ALI show problems in the comprehension of reflexives, whilst ALN children do not.

2.3 Proposed explanations for the Principle A deficitBased on their studies, Perovic et al. (2013a; 2013b) propose two main explana-tions for the principle A deficit found in children with ALI. As illustrated above, the only part of Principle A tested in Perovic et al.’s (2013a; 2013b) studies is the c-command constraint. Perovic et al. (2013a; 2013b) only tested sentences like (4), and not sentences like (2), with a possible local and a possible distant anteced-ent. Therefore, nothing can be concluded about possible knowledge of the locality constraint in the ALI population. Perovic et al. (2013a; 2013b) propose that chil-dren with ALI are at least insensitive to the c-command constraint of Principle A. They hypothesize that children with ALI might have a different version of Principle A, with only the locality constraint. This would mean that they only require an ante-cedent to be in the same clause as the reflexive, but do not require the antecedent to be c-commanded by its referent.

However, an insensitivity to c-command raises many questions related to other syntactic phenomena that rely on c-command relationships. If children with ALI do not show sensitivity to the c-command constraint in binding constructions, the question would be if they do show sensitivity to c-command in other structures. Importantly, Perovic et al.’s (2013a; 2013b) experiments did show that children with ALI performed well on another structure which interpretation depends on c-command: possessive sentences without reflexives (e.g. ‘Bart’s dad is petting a dog’). They showed the right interpretation of 77% of these sentences, which sug-gests some knowledge of c-command. Thus, the hypothesis that children with ALI


lack the c-command constraint on binding leads to problems related to learnabil-ity. It should be noted, however, that knowledge of c-command in the ALI popula-tion has not been investigated apart from Perovic et al.’s (2013a; 2013b) studies, as far as I am aware.

The second hypothesis Perovic et al. (2013a;2013b) propose is that children with ALI might not have knowledge of Principle A at all. Children with ALI lack a filter on the binding of reflexives and allow a range that is too big. This would mean that children with ALI might have knowledge of the general principles of c-command and locality, but do not know to apply these principles as constraints to the inter-pretation reflexives. In order to understand this hypothesis, consider the results of the picture selection task used in both of Perovic et al.’s (2013a; 2013b) stud-ies. Children with ALI performed at chance level. Under the hypothesis that these children have no constraints on the binding of reflexives, this could be explained in the following way: Children with ALI allow both possible interpretations in the picture selection task. So, if they hear sentence (4) and see a picture where Bart points at himself and a picture where Bart’s father points at himself, both pictures are actually depictions of sentence (4) for them.

3. Methodological considerationsIt should be noted at this point that the studies by Perovic et al. (2013a; 2013b) had multiple methodological drawbacks. First of all, their PST consisted of pairs of only two pictures, showing two possible interpretations and giving children a 50% chance at a correct answer. This absence of fillers is problematic. Furthermore, the use of a PST is unfitting to answer Perovic et al.’s (2013a; 2013b) research question. A PST only tests which interpretation is preferred by participants. Perovic et al. (2013a; 2013b) did not test if their participants actually allow both inter-pretations of a given sentence like (4). However, in order to test whether children have a certain linguistic constraint, it is vital to test if multiple interpretations are possible for them. The children need to be put in a position where they can deny the appropriateness of a sentence given an ambiguous situation.

An experimental method that allows participants to be put in this position is the Truth Value Judgement Task (TVJT) (first proposed by Crain & McKee, 1985). The fundamental characteristic of this task is that it requires the child to give a bi-nary (yes/no) judgement about whether target statements provide an accurate description of a particular situation. TVJT’s are widely used to test knowledge of constraints, including binding, in TD children (e.g. Baauw, Zuckerman, Ruigendijk & Avrutin, 2011; Spenader et al., 2009; Van Koert et al., 2015; Van Rij et al. 2010). TVJT’s are used to test the ASD population as well (e.g. on c-command and binding in High-Functioning children with ASD, Khetrapal & Thornton, 2017; and on scope, Durrleman et al., 2016).

19Sanne Ditewig

4. The current studyPerovic et al. (2013a; 2013b) cannot exclude either of their possible explanations discussed above, because they did not test the locality constraint on the binding of reflexives. Therefore, the nature of the deficit in the comprehension of reflexives is still unknown. In order to breach this gap in knowledge, the current research proposal proposes methodology to investigate the following question: 'Do children with ALI have knowledge of the locality constraint on the interpretation of reflex-ives?' The two possible explanations for the Principle A deficit in children with ALI discussed in section 2.3 lead to the following hypotheses regarding sentences like (2), with both a possible local and distant antecedent:

H0: Children with ALI have no locality constraint and lack knowledge of Principle A altogether. This is reflected in high acceptance rates of both distant and local antecedents for reflexives.

H1: Children with ALI do have a locality constraint on the binding of reflexives, and thus have an altered version of Principle A. This is reflected in high acceptance rates of local referents for reflexives and low acceptance rates of distant anteced-ents for reflexives.

5. Method5.1 Participants

The current study includes 40 monolingual Dutch speaking children, aged six to twelve. Half of these children have been diagnosed as ALI. The other half are the TD control group. The children with ALI are recruited from schools for special education and treatment centers. They are categorized as ‘ALI’ based on their diagnosis of language problems, shown by scores below the tenth percentile on language production and perception tests taken during their diagnostic process (following Perovic et al., 2013b). The TD children are recruited from schools and daycares.

5.2 General design and ProcedureThe design of the study is a reward/punishment variation of the Truth Value Judge-ment Task (TVJT). This is in contrast to the design used by Perovic et al. (2013a; 2013b), for the reasons explained in section 3. Target sentences are presented with short stories. These stories are presented by acting them out with toys. After the story a puppet describes the situation that was just acted out. The children are instructed to either reward or punish the puppet if he describes the situation correctly/ incorrectly. The stories are checked for (pragmatic) felicity in a pretest with adults and TD children. The experiment takes place during three sessions, in order not to strain the children. All sessions start with a training phase, in which children are introduced to the task by showing them two practice stories. This is meant to ensure the child understands the task. The order of the stories is semi-randomized in each session; making sure that the child does not hear too many sentences with the same characters and sentence types in a row.


5.3 Experimental materials 5.3.1 Target sentences

Dutch has two reflexive pronouns, zichzelf and zich. The primary reflexive is zich-zelf. Zich occurs only with inherently reflexive predicates (e.g. zich schamen, 'to be ashamed') (Wijnen & Verrips, 1998). There is no gender marking on the Dutch re-flexives, as there is in English. To keep the stimuli consistent only zichzelf is used in the sentences; since differences in the comprehension of the two reflexives have been found in children, related to verb types (Sigurjónsdóttir & Coopmans, 1996).

The experimental stimuli are eight Dutch sentences with reflexives and a possible distant and local antecedent, as in examples (5) to (7).

(5) Elsa vraagt Sneeuwwitje om zichzelf te schilderen. Elsa asks Snow White to herself to paint. ‘Elsa asks Snow White to paint herself.’

(6) Dora zegt dat Assepoester een broodje voor zichzelf moet smeren. Dora says that Cinderella a sandwich for herself must butter. ‘Dora says that Cinderella has to make herself a sandwich.’

(7) Olaf vraagt Kristoff om zichzelf te wassen. Olaf asks Kristoff to himself to wash. ‘Olaf asks Kristoff to wash himself.’

The reflexive ‘zichzelf’ (‘herself’) in example (5) can be bound by the local anteced-ent ‘Sneeuwwitje’ (‘Snow White’), but under the locality constraint of Principle A, the reflexive cannot be bound by the distant antecedent ‘Elsa’. The structure of the sentences is loosely based on a study by Chien & Wexler (1990) in which they tested the locality constraint on the binding of reflexives in TD children, using an act-out task. The Dutch equivalents of three verbs used by Chien & Wexler (1990) are used, ‘vragen’(‘ask’), ‘zeggen’(‘say’) and ‘willen’ (‘want’). This way, the stimuli include both finite and infinitival sentences. Additionally, the stimuli are balanced on having an extra object in the story, like in (6).

Ten control sentences are used with a similar structure as the experimental stim-uli, but the reflexive is replaced by either an animate or an inanimate object. The sentences are specifically constructed not to contain any of the morpho-syntactic structures that have been found to be problematic for children with ALI (such as raising or object control; see Perovic et al., 2013b). An example of a control sen-tence is the sentence in (8):

(8) Dora wil dat Ariel de kroon op haar hoofd zet. Dora wants that Ariel the crown on her head puts. ‘Dora wants Ariel to put the crown on her head.’

21Sanne Ditewig

5.3.2 TVJT StoriesAll stories in the TVJT are similar in size and set-up. All eight target sentences described above are accompanied by two stories; one corresponding to an inter-pretation of the sentence in which the reflexive refers to the local referent (here-after: local stories) and one corresponding to the interpretation of the sentence in which the reflexive refers to the distant referent (hereafter: distant stories). I will illustrate the story types with sentence (5), repeated here:


In the local story Elsa asks Snow White to paint herself (Snow White), whereas in the distant story Elsa asks Snow White to paint her (Elsa). These stories contain-ing the experimental target sentences are complemented by stories containing the ten control sentences. Four of these stories show a correct interpretation of the sentence, and the other six clearly deviate from the content of the sentence. This ensures that some of the stories elicit a ‘no’-response in the children, regard-less of their ability to comprehend reflexives. This way, a possible yes-bias (or no-bias) is controlled for.

The design thus incorporates three different conditions: stories for control sen-tences (control condition), stories which represent an interpretation in which a reflexive is bound by a local referent (local condition) and stories which represent an interpretation in which a reflexive is bound by a distant referent (distant condi-tion). Apart from in the target sentences, no reflexives are used in the stories, to ensure this will not hamper general understanding of the story.

6. Analyses and anticipated resultsThe data are analyzed by taking two factors into account; Locality and Group. Locality refers to the type of story (distant or local), and Group refers to the experi-mental group of children with ALI and the control group of TD children. Within each Group, the mean percentages of correct judgements for both the local, distant and control stories are compared. The Locality means are compared using Mann-Whitney tests1, to test whether the differences are statistically significant. These comparisons show how accurate Children with ALI and TD children are in their judgements of the target sentences, and, if their ability to judge the sentences correctly differs between sentence type. In the next step, the group means for the ALI Group are compared to that of the TD control Group, to test whether there is a difference in correct judgements between the groups.

What is expected is that if the null hypothesis holds true, there is no locality con-straint on the interpretation of reflexives by children with ALI. This is the most

1 The Mann-Whitney test is a non-parametric test, primarily used to compare means of ordinal variables.


likely outcome, as discussed in section 2.3. Remember that, under the locality constraint, all distant stories should elicit a ‘no’-response and all local stories should elicit a ‘yes’-response. The crucial difference between the two hypotheses will thus be found in the correctness percentages of the stimuli presented with distant stories. To illustrate, I once again use example (5).


The distant story, where Snow White paints Elsa, has the correct response no. The local story, where Snow White paints herself, has the correct response yes. If children do not have a locality constraint, this means that they will accept distant referents for reflexives as well as local referents. The children will thus have high acceptance rates for both distant and local stories, and a significantly lower cor-rectness score on the distant stories than on the local stories.

If the experimental hypothesis holds true, there is a locality constraint on the inter-pretation of reflexives in children with ALI. This means that the children with ALI will accept the stimuli presented with local stories significantly more often than those presented with distant stories. Their overall correctness percentages for both categories thus will be high, and there will not be a significant difference be-tween these mean scores. This pattern of results is expected for the TD group as well, as they are expected to have knowledge of the locality constraint. A pretest of the proposed material with three TD children shows that this is indeed the case.

Additionally, individual patterns should be considered to check whether all chil-dren with ALI pattern the same and whether they show understanding of the task, or a yes-bias, in the control sentences.

7. ConclusionThis paper proposed research into the comprehension of reflexives in children with ALI. Research by Perovic et al. (2013a; 2013b) indicates that children with ALI have problems understanding reflexive pronouns. However, the extent to which reflexive pronouns are problematic and the cause of these problems remain un-known. The studies by Perovic et al. (2013a; 2013b) only investigated one aspect of the comprehension of reflexives; they must be c-commanded by their anteced-ents (Chomsky, 1986). A second important constraint, the locality constraint, has not yet been investigated in the ALI population. Moreover, the studies by Perovic et al. (2013a; 2013b) have some methodological drawbacks, resulting from the use of methodology that does not allow testing of possibly ambiguous interpretations.

The proposed research aims to investigate the comprehension of the locality con-straint in children with ALI, whilst overcoming the methodological drawbacks of Perovic et al.’s (2013a; 2013b) studies. The proposed research has both clinical

23Sanne Ditewig

as well as theoretical contributions, providing knowledge on the linguistic symp-tomatology of children with ALI and indicating which syntactic constraints group together in first language acquisition. ■

Received October 2018; accepted July 2019.

ReferencesAmerican Psychiatric Association. (2013). Diagnostic and statistical manual of mental dis-

orders (5th ed.). Arlington, VA.Baauw, S., Zuckerman, S., Ruigendijk, E., & Avrutin, S. (2011). Principle B delays as a pro-

cessing problem: Evidence from task effects. Production-comprehension asymmetries in child language, 247-272.

Bergmann, C., Paulus, M. & Fikkert, P. (2009). A Closer Look at Pronoun Comprehension: Comparing Different Methods. In: J. Costa, A. Castro, M. Lobo & F. Pratas (Eds.) Lan-guage Acquisition and Development: Proceedings of GALA 2009, 53-61. Cambridge Scholars Publishing, Cambridge.

Chien, Y., & Wexler, K. (1990). Children’s knowledge of locality conditions in binding as evi-dence for the modularity of syntax and pragmatics. Language Acquisition, 1. 225–295.

Chomsky, N. (1986). Knowledge of language: Its nature, origin and use. New York, NY: Praeger.

Crain, S., & McKee, C. (1985). The acquisition of structural restrictions on anaphora. In Pro-ceedings of NELS 15, 94-110.

Durrleman, S., Burnel, M., Thommen, E., Foudon, N., Sonié, S., Reboul, A., & Fourneret, P. (2016). The language cognition interface in ASD: Complement sentences and false belief reasoning. Research in Autism Spectrum Disorders, 21, 109-120.

Eigsti, I. & Bennetto, L. (2009). Grammaticality judgments in autism: Deviance or delay. Journal of Child Language, 19, 1–23.

Khetrapal, N., & Thornton, R. (2017). C-Command in the Grammars of Children with High Functioning Autism. Frontiers in Psychology, 8, 402, Retrieved from: https://doi.org/10.3389/fpsyg.2017.00402

Kjelgaard, M., & Tager-Flusberg, H. (2001). An investigation of language impairment in au-tism: Implications for genetic subgroups. Language and cognitive processes, 16(2-3), 287-308.

Noveck, I. A., Guelminger, R., Georgieff, N., & Labruyere, N. (2007). What autism can reveal about every… not sentences. Journal of Semantics, 24(1), 73-90.

Paul, R., Chawarska, K., & Volkmar, F. (2008). Differentiating ASD from DLD in Toddlers. Per-spectives on language learning and education, 15, 101–111.

Perovic, A., Modyanova, N. & Wexler, K. (2013a). Comprehension of reflexive and personal pronouns in children with autism: A syntactic or pragmatic deficit?. Applied Psycholin-guistics, 34(4), 813-835.

Perovic, A., Modyanova, N. & Wexler, K. (2013b). Comparison of grammar in neurodevel-opmental disorders: The case of binding in Williams syndrome and autism with and without language impairment. Language acquisition, 20(2), 133-154.

Roberts, J. A., Rice, M. & Tager-Flusberg, H. (2004). Tense marking in children with autism. Applied Psycholinguistics, 25, 429–448.

Sigurjónsdóttir, S. & Coopmans, P. (1996). The acquisition of anaphoric relations in Dutch. In W. Philip & F. Wijnen (eds.), Connecting children’s language and linguistic theory (Amsterdam series in child language development 5), 52-72. Amsterdam: Institute of General Linguistics.

Spenader, J., Smits, E. & Hendriks, P. (2009). Coherent Discourse Solves the Pronoun Inter-pretation Problem. Journal of Child Language 36, 23-52.

Van Koert, M., Hulk, A., Koeneman, O., & Weerman, F. (2013). How do Turkish-Dutch Bilin-


gual Children Interpret Pronouns and Reflexives in Dutch?. In Proceedings of the 12th Generative Approaches to Second Language Acquisition Conference, 85-99.

Van Rij, J., van Rijn, H. & Hendriks, P. (2010). Cognitive Architectures and Language Ac-quisition: A Case Study in Pronoun Comprehension. Journal of Child Language 37(3), 731-766.

Wijnen, F., & Verrips, M. (1998). The acquisition of Dutch syntax. In S. Gillis & A. de Hou-wer (eds.), The acquisition of Dutch, 223-300. Amsterdam: John Benjamins Publishing Company.

25Judith Brinksma & Nelleke Jansen

Phonetic Convergence of Students from Limburg in Utrecht:(De)voicing of /v/ and Hard and Soft /ɣ/

J.H. (Judith) Brinksma & N.G. (Nelleke) JansenRMA Linguistics, Utrecht University, Utrecht.

KEYWORDSphonetic convergencefricativesDutchsociolinguistics

ABSTRACTPhonetic convergence refers to speakers’ tendency to imitate other’s pro-nunciation. This convergence has been demonstrated for both interlocu-tors within conversations (e.g. Pardo, 2006) and for groups of speakers over longer periods of time (e.g. Quené et al., 2017). While previous stud-ies have established convergence in different time-related contexts, few studies have yet considered what kind of variables are taken up in con-vergence and what factors influence this. The current study considers the potential convergence of students of Utrecht University who grew up in the south of the Netherlands (Limburg) to fellow students who grew up in or around Utrecht. This was investigated on the basis of a less salient fea-ture, the (de)voicing of /v/, and a more salient feature, the hard and soft /ɣ/ distinction. Participants were recorded while reading out a word list and answering selected open questions. The latter was not only used for analysis of /v/ and /ɣ/, but also provided information on factors possibly influencing convergence. Results indicated that one of four participants originally from Limburg regularly devoiced /v/, similar to participants from Utrecht, whereas one other participant consistently produced a uvular /ɣ/. Register (word-reading task vs interview) did not appear to be a rel-evant factor. Time of residence and language attitudes are suggested to play a role in convergence.

1. IntroductionMultiple studies have investigated speakers’ tendency to imitate each other’s pro-nunciation, a phenomenon known as phonetic convergence. This convergence has been demonstrated for both interlocutors within conversations (e.g. Pardo, 2006) and for groups of speakers over longer periods of time (e.g. Quené, Orr, & Van Leeuwen, 2017). The study of convergence gives insight into the interac-tion between speech perception and production (see Pardo, 2006). While conver-gence has been found in different time-related contexts, few studies have con-sidered which variables influence the extent to which convergence takes place. Pardo, Gibbons, Suppes and Krauss (2012) found that reported closeness to the interlocutor correlates with degree of convergence, suggesting that attitudinal fac-tors play a role. Besides attitude, the current study also investigates register as a factor possibly influencing convergence.


The current pilot study investigates phonetic convergence within the Dutch stu-dent community in Utrecht. The focus is on the realisation of /v/ and /ɣ/ by stu-dents who were born and raised in the province of Limburg (the most southern province) but who had been living in Utrecht for at least half a year. Previous stud-ies suggest that devoicing of /v/ is less strong in Limburg than in more northern areas like Utrecht (e.g. Van de Velde, Gerritsen & Van Hout, 1996; Pinget, 1996). The degree of voicing of /v/ was thus considered to be an adequate measure of possible phonetic convergence. The realisation of /v/ was chosen as speak-ers are generally not aware of the different realisations regarding voicing. This contrasts with the realisation of /ɣ/, which is more salient. In Limburg /ɣ/ has a (palato )velar pronunciation, whereas northern areas are characterised by a uvular realisation (e.g. Van de Velde, Van Hout & Gerritsen, 1997; Van der Harst & Van de Velde, 2008). Investigating both the (de)voicing of /v/ and the realisation of /ɣ/ will provide insight into what kind of variables are taken up in convergence in different contexts. Fricative realisations were assessed for a word-reading task and a sociolinguistic interview. The interview also collected attitudinal informa-tion, which was employed in a qualitative analysis of the data.

Results indicated that one of four participants originally from Limburg regularly devoiced /v/, like participants from Utrecht, whereas one other participant con-sistently produced a uvular /ɣ/. Register did not appear to be a relevant factor. We suggest time of residence and language attitudes play a role in convergence.

2. Theoretical background2.1 Phonetic convergence

In 1981, Trudgill found that his pronunciation shifted towards that of the person he was interviewing. More than twenty years later, Pardo (2006), investigating convergence on a larger scale, found that interacting speakers increase their similarity in phonetic repertoire. This can be explained by the social function of language, as it could for instance “contribute to mutual comprehension and/or rapport through a decrease in social distance” (p. 2391). Whereas her study demonstrates convergence within conversational interaction, Pardo suggests that further research is needed to investigate this phenomenon across speaker com-munities. She also argues that phonetic convergence “can form the basis for phe-nomena such as accent change and dialect formation” (p. 2382).

In recent years several studies have examined phonetic convergence over longer periods of time and in community contexts. Student populations are well suited for investigations into convergence, as they usually consist of people from a va-riety of origins and backgrounds. For example, Pardo et al. (2012) consider pho-netic convergence in pairs of college roommates from different regions in the US over the course of one academic year. Their findings were variable and correlated moderately with the self-reported closeness of the relationships. Some partici-pants indeed showed longer-term adjustments in their phonetic repertoire. A lon-gitudinal study by Orr et al. (2011) looked at phonetic convergence in native and


non-native speakers of Dutch and English attending University College Utrecht, throughout their three-year period of undergraduate study. The results show that the Dutch speakers had converged in their pronunciation of /s/ in both Dutch and English (Quené et al., 2017). These studies suggest that phonetic convergence indeed takes place on a larger scale and may lead to more permanent changes in pronunciation. The present study extends previous investigations by investi-gating convergence of students from Limburg towards the pronunciation of local students in Utrecht.

2.2 The devoicing of /v/The devoicing of voiced fricatives is a change affecting both Northern Standard Dutch (NSD), spoken in the Netherlands, and Southern Standard Dutch (SSD), spoken in Flanders. Van de Velde et al. (1996) analysed archived radio-recordings from 1935-1993, finding that in both NSD and SSD the devoicing of /v/ increased during this time period, although the increase was weaker in SSD than in NSD. Van de Velde and Van Hout (2001) investigated the devoicing by Dutch language teachers in a reading task. Again, the primary different regions looked at were the Netherlands and Flanders, but regional variation within the Netherlands was also considered. Results indicated that /v/ was devoiced least often in the south. A recent study by Pinget (2015) studied the perception and production of the la-biodental fricatives by 18- to 28-year-olds from different regions. She found more voicing of /v/ in Limburg than in other regions in the Netherlands. These results provide good reason to believe that /v/ is more voiced in Limburg than in Utrecht. Pinget (2015) also found more devoicing in the case of a word reading task than in the case of spontaneous speech. The difference in results could be due to devoic-ing being considered part of the standard nowadays. A word reading task provides greater opportunity for reflection upon pronunciation, and thus more standard language may be used.

2.3 The pronunciation of /ɣ/Although the difference between the Northern ‘hard g’ and the Southern ‘soft g’ is well known, few researchers have investigated this topic. Another paper on the radio study described earlier considers the realisation of /ɣ/. Van de Velde, Van Hout and Gerritsen (1997) investigated several linguistic variables, including the place of articulation and scrape of /ɣ/, and found that the pronunciation in NSD became harder in the period investigated, especially towards 1993. This means that /ɣ/ is pronounced further back (i.e., uvular) and with heavy scrape, instead of a softer velar or palato-velar pronunciation. Van de Velde et al. suggest these changes first took place in middle- and upper-class speech in the larger cities in the Randstad area, including Utrecht (p. 384). A more recent study by Van der Harst and Van de Velde (2008) investigating the pronunciation of /ɣ/ by teachers of Dutch has demonstrated a difference between the Randstad region and Lim-burg. Although the /ɣ/ is voiceless throughout the Netherlands, the results show a clear distinction in place of articulation and scrape above and below the great rivers. In the regions Randstad, Northern Netherlands and Middle Netherlands,


the uvular /ɣ/ with scrape is mostly used, whereas in Limburg, velar and palato-velar fricatives are most common.

In summary, previous studies have demonstrated different norms for the pronun-ciation of /v/ and /ɣ/ between Limburg and the Randstad, as well as speakers’ tendency to converge to the pronunciation common in their surroundings. The current study therefore investigates if students from Limburg adapt their pronun-ciation of the aforementioned sounds to the pronunciation common in Utrecht. We considered students’ production of these fricatives in the reading of a word list and in an interview.

3 Research questions and hypothesesWe aimed to answer the following research question: Do students from Limburg living in Utrecht demonstrate phonetic convergence towards the pronunciation of Standard Dutch in Utrecht regarding the (de)voicing of /v/ and the pronunciation of /ɣ/? Two subquestions were considered: (1) Does the degree of convergence depend on the register? (2) Is there within-group variance, and can this be ex-plained by sociolinguistic factors?

Based on studies of phonetic convergence among student populations by Pardo et al. (2012) and Quené et al. (2017), we hypothesized that students from Lim-burg living in Utrecht converge towards the Randstad pronunciation, as demon-strated by a similar extent of (de)voicing of /v/ as speakers raised in Utrecht, as well as a hardening of /ɣ/. Based on Pinget (2015), we expected participants to demonstrate more devoicing in a word list reading than in spontaneous speech. As words are presented in isolation in the word reading task, it might be difficult for participants to produce voicing from the start of the fricative onwards. Addi-tionally, as the interview concerns continuous speech, voiceless fricatives may be harder to produce due to coarticulation (Pinget, 2015, p. 46). Participants were expected to use a harder /ɣ/ in the word list reading than in the sociolinguistic interview, because of a possible greater focus on standard pronunciation in the word list, with hard /ɣ/ generally being considered more standard in the Nether-lands than soft /ɣ/ (Van de Velde et al., 1997). As for the second subquestion, within-group variance was expected, which would be explained by sociolinguistic differences between participants including the duration of residence in Utrecht and language attitudes. This means that participants who had lived in Utrecht for several years had converged more than participants who had moved there more recently, and that negative attitudes towards a Limburgian accent would increase degree of convergence.

4 Methodology4.1 Participants

The participants of this pilot study were seven students living in Utrecht, all native speakers of Dutch. Four students grew up in central Limburg, forming the experi-mental group; the other three students were raised in or nearby Utrecht, forming


the control group. The students who had moved to Utrecht had not lived in any provinces other than Limburg and Utrecht. They had lived in Utrecht for varying pe-riods of time, ranging from half a year to three and a half years. Previous studies found evidence of convergence within the year (e.g. Pardo et al., 2012), thus stu-dents who had not lived in Utrecht for multiple years could possibly demonstrate convergence. Length of residence is considered as a variable in the qualitative analysis. For detailed demographics of the participants see appendix A.

4.2 MaterialsA word list and a sociolinguistic interview were designed to assess phonetic con-vergence. The word list consisted of 50 randomized items (appendix B). 10 of these items were designed to measure the realisation of /v/ and 19 were in-cluded to measure the pronunciation of /ɣ/.1 For both /v/ and /ɣ/ we considered the position of the fricative and whether the syllable containing the fricative was stressed (see Tables 1 and 2). It was deemed important to code for the aforemen-tioned variables to enable comparison with the spontaneous speech data, where fricatives could appear in any of these environments. To avoid coarticulation of /ɣ/ with other consonants, target words were chosen such that /ɣ/ was preceded or followed by a vowel.

Table 1 Examples of the conditions for /v/ in the word reading task.

Initial /v/ (n=5) Medial /v/ (n=5)Unstressed Vakantie GravenStressed Vinden Activisme

Table 2 Examples of the conditions for /ɣ/ in the word reading task.

Initial /ɣ/ (n=7) Medial /ɣ/ (n=5) Final /ɣ/ (n=6)

Unstressed gezond jarige handig

Stressed gasfornuis legaal bedrag

A sociolinguistic interview (appendix C) was designed to elicit spontaneous speech. The interview simultaneously elicited data on participants’ attitudes and awareness of possible convergence. These data were considered relevant for the analysis of results of the experimental group. For the control group, the questions specific to having moved were replaced by other questions.

Materials for the two tasks were combined in a PowerPoint presentation, with the

1 The experiment was also designed to consider potential devoicing of /z/ (10 items) along with the devoicing of /v/, but these data are not discussed in the current paper considering it is not clear whether this variation follows the expected regional pattern (Van de Velde & Van Hout, 2001). The remaining 11 items were fillers.


word list preceding the interview. The wordlist was presented one word per slide such that participants had to pause the articulation and phonetic assimilation could not take place from one word to the next. The interview was presented with one question per slide such that participants took their time and remained una-ware of the purpose of the study for as long as possible.

4.3 ProcedureParticipants were recorded in a soundproof room with a microphone (Sennheiser ME-64). The PowerPoint was presented on a computer screen. Participants used the mouse to click through the slides. Prior to the tasks, instructions were dis-played on a slide. The participants were instructed to take their time with the tasks and to correct themselves in case of mispronunciations. Regarding the in-terview, participants were instructed to answer as accurately as they could and to use full sentences. Audio was recorded using the Audacity software (Audacity Team, 2018).

4.4 AnalysisFor both the wordlist and the interview, the instances of /v/ and /ɣ/ were an-notated using Praat (Boersma & Weenink, 2018). Based on auditory analysis, relevant instances of /v/ (i.e. excluding ones that are standardly devoiced regard-less of region) were coded as voiced, partially voiced, or voiceless, and /ɣ/ was coded as (palato-)velar or uvular. The uvular articulations of /ɣ/ were also coded for scrape using the three-way distinction ‘no scrape’ (ns), ‘light scrape’ (ls) and ‘heavy scrape’ (hs). Scrape, which is caused by vibration of the uvula, is easily audible and reflected by sound waves on the spectrogram. For each participant the interview was transcribed in Dutch.

The results of the experimental group and the control group were compared for the word list and the sociolinguistic interview. In addition, the results of the two tasks were compared and the within-group variation of the experimental group was determined. However, due to the small sample size of this pilot study, no sta-tistical test could be performed. Information from responses to the sociolinguistic interview was used in a qualitative analysis of the results on fricative realisations.

5 Results5.1 Word Reading Task: /v/

In the word reading task, the overall percentage of voicing was 46.25 for the ex-perimental group and 28.33 for the control group. These percentages were calcu-lated as follows: if /v/ was voiceless, it was scored 0, if it was partially voiced 1, and if it was voiced 2; subsequently the sum was taken, and the percentage of the total possible points was calculated. Individual results show that the overall pat-tern does not hold for all participants. Figure 1 presents the percentage of voicing per participant (experimental group: E1-E4, control group: C1-C3). The higher voic-ing of /v/ for the experimental group holds for participants E1, E2, and E3, while participant E4 frequently devoices /v/.


Figure 1. Percentage of voicing of /v/ per participant for the word list.

5.2 Sociolinguistic interview: /v/For most participants the first 10 relevant words with /v/ were analysed. This number was not met by participant E2 and participant C1, for whom 9 and 6 items were found respectively. Overall, the percentage of voicing was 46.15 for the experimental group and 7.69 for the control group. At group level, there was no difference for the experimental group between registers. Figure 2 presents the voicing per participant in the interview. More devoiced realisations of /v/ were found for participant E2 compared the word list reading, and thus the difference from the control group is smaller. Additionally, participant E4 again performs like the control group, while participant E1 and E3 do not.

Figure 2. Percentage of voicing of /v/ per participant for the interview.


5.3 Word Reading Task: /ɣ/Results for /ɣ/ in the word list reading are given in Figure 3. Overall, the pronuncia-tion of the experimental group had not converged towards the control group. Three of the participants in the experimental group used almost exclusively (palato-)velar variants of /ɣ/. One of them used a single uvular articulation. As expected, all participants in the control group used only uvular articulations. One of the par-ticipants in the experimental group (E3) used exclusively uvular pronunciations as well, although he produced fewer articulations with heavy scrape than the control group participants.

Figure 3. Realisation of /ɣ/ in percentages per participant for the word list.

5.4 Sociolinguistic interviews: /ɣ/The sociolinguistic interviews (Figure 4) again show that the group of students from Limburg in general had not converged towards the control group regarding their pronunciation of /ɣ/. As in the word reading task, three participants of the experimental group produced a (palato-)velar /ɣ/, whereas participant E3 had a uvular pronunciation, like the control group. The participants with uvular articula-tions used more scrape in the interview than in the reading task. In other respects, the results are similar; again, the participant from Limburg who had a uvular /ɣ/ produced fewer words with heavy scrape compared to the control group.


Figure 4. Realisation of /ɣ/ in percentages per participant for the word list.

5.5 Sociolinguistic Interviews: The ResponsesThe sociolinguistic interviews provided information regarding the participants’ so-cial circles, their attitudes towards different accents, and awareness of their own potential accent change. The participants were aware that their speech was highly influenced by their place of origin. They all expressed awareness of prejudices towards accents, and they particularly discussed the fact that the accent of Utre-cht (or the Randstad in general) was more widely accepted than more peripheral accents like the Limburgian accent. Some participants of the experimental group also mentioned specific characteristics of their speech that had changed since they moved to Utrecht, which demonstrates awareness of phonetic convergence. Appendix D2 provides a summary of the answers per participant.

6. DiscussionThe current pilot study was designed to investigate whether students who had moved to Utrecht from Limburg had adapted their accent based on their current environment. Comparison of the experimental and control group did not offer con-crete evidence for convergence. Regarding the (de)voicing of /v/, participants’ individual results indicate that participant E4 may have converged in his pronun-ciation of /v/. Participant E4 devoiced regularly in both the word reading task and the sociolinguistic interview, patterning with the control group. Participant E2 pro-duced largely voiceless realisations in the interview but appeared to converge only in this task. The results contrast with what was expected based on Pinget (2015) regarding task type. The participant’s stronger devoicing in the interview may be explained by the high number of words with initial /v/ selected. Previous studies

2 Find the online version of this article, including appendices, at linguujournal.nl.


have found stronger devoicing in initial position (e.g. Van de Velde, Gerritsen & Van Hout, 1996). In summary, it appears that participant E4 is the only participant who has converged in the realisation of /v/, which was the case independent of register.

Results on the pronunciation of /ɣ/ indicate that only one of the four speakers from Limburg had converged towards a uvular realisation. The other three partici-pants from Limburg retained a (palato-)velar pronunciation. The expected within-group variance was thus confirmed. Considering the degree of scrape used by participant E3 and the participants in the control group, it can be concluded that although the speaker from Limburg had obtained a uvular pronunciation of /ɣ/, he used less scrape than the control group speakers, indicating that his /ɣ/ was still somewhat softer. The word reading task and the spontaneous speech from the interviews generally gave similar results. However, contrary to what was ex-pected regarding possible differences in register, speakers with a uvular place of articulation used heavier scrape in natural speech than in the word list. In other words, when speakers focus on the articulation of a specific word, they sometimes produce a softer /ɣ/ than in normal speech. This is surprising, considering that a hard /ɣ/ is regarded as the more standard form in the Netherlands (Van de Velde et al., 1997). An explanation for this result might be that heavy scrape sounds rather harsh, even for the Randstad group, when speakers pay attention to their pronunciation. However, the differences between the two tasks were not greater for participant E3 than for the control group, so we can again conclude that the degree of convergence does not depend on register.

The question that remains is whether within-group variance can be explained by sociolinguistic factors, as was hypothesised for the second sub question. Firstly, we need to determine why participant E4 had converged in his realisation of /v/, while other participants had not. Possibly, the period of residence is a relevant factor. In contrast to participants E1 and E2, who had lived in Utrecht for a year or less, participant E4 had lived in Utrecht for more than three years. The same was the case, however, for participant E3. This participant had converged in his realisation of /ɣ/, but not of /v/. It should be noted that participant E3 states he had a uvular pronunciation of /ɣ/ before he moved to Utrecht and while he may thus have converged to some extent, he did not really make the switch from hard g to soft g. Another clear difference between participant E3 and the rest of the experimental group is how often they visit their families in Limburg. Whereas par-ticipants E1, E2 and E4 all go back nearly every week, participant E3 goes back about twice a month, which may explain the fact that only E3 has converged in his pronunciation of /ɣ/.

Finally, it is possible that attitudes play a role. Whereas participants E1 and E2 were hesitant to express a preference, participant E4 stated that he preferred a “neutral” accent and highlighted the negative associations of a Limburgian accent. The same explanation can be given for the phonetic convergence demonstrated


by participant E3, who had adapted a uvular articulation of /ɣ/, of which he was conscious himself. In his interview, this participant gave a slight preference for a Randstad pronunciation, because it “helps you become part of your surroundings in Utrecht”. He also expressed the belief that people may be negatively prejudiced towards Limburgian accents, which would make you sound “dumb”. These lan-guage attitudes, or even beliefs about other people’s attitudes—which most par-ticipants discussed—may thus form a deciding factor for phonetic convergence.

The collected data, specifically the sociolinguistic interviews, may also be used for further research on other linguistic variables, such as the pronunciation of /r/, which had clearly changed for one of the participants. It should be noted that the experiment described here was a pilot study, which could be elaborated on by, for example, recruiting more participants and adding a control group of students who were raised in Limburg and are still living there. However, the best research design would be longitudinal such that individual realisations can be tracked.

7. ConclusionThis study investigated the realisation of /v/ and /ɣ/ by students living in Utrecht, comparing students who grew up in Limburg to students raised in Utrecht, to see if the pronunciation of the Limburg students had converged. Analysis of a word reading task and a sociolinguistic interview indicated that, while the experimental group in general had not converged for /v/ and /ɣ/, one participant had con-verged for /v/, devoicing more often than other participants of the experimental group. This finding is likely related to the period the participant had lived in Utre-cht, although other factors such as attitudes are also suggested to play a role. An-other participant had converged towards the Randstad accent regarding the uvu-lar realisation of /ɣ/. The convergence of /ɣ/ could largely be explained by how often students returned to their families in Limburg, in addition to beliefs about attitudes towards accents. The current pilot study thus provides some evidence suggesting long-term convergence but also notes various factors convergence is dependent on besides interaction with speakers with another accent, including attitudes and the strength of connections to students’ home towns. However, fur-ther research is needed to corroborate these preliminary findings. ■

Find the online version of this article, including appendices, at linguujournal.nl

Received November 2018; August 2019.

ReferencesAudacity Team (2018). Audacity: Free Audio Editor and Recorder (Version 2.2.2) [Computer

software]. Retrieved from https://audacityteam.org/Boersma, P., & Weenink, D. (2018). Praat: Doing phonetics by computer (Version 6.0.37)

[Computer software]. Retrieved from http://www.fon.hum.uva.nl/praat/Orr, R., Quené, H., Beek, R. V., Diefenbach, T., Leeuwen, D. A. V., & Huijbregts, M. (2011).

An international English speech corpus for longitudinal study of accent development. In Twelfth Annual Conference of the International Speech Communication Association


(pp. 1889-1892).Pardo, J. S. (2006). On phonetic convergence during conversational interaction. The Jour-

nal of the Acoustical Society of America, 119(4), 2382-2393.Pardo, J. S., Gibbons, R., Suppes, A., & Krauss, R. M. (2012). Phonetic convergence in col-

lege roommates. Journal of Phonetics, 40(1), 190-197.Pinget, A. C. H. (2015). The actuation of sound change. Utrecht: LOT.Quené, H., Orr, R. & van Leeuwen, David A. (2017). Phonetic similarity of /s/ in native and

second language: individual differences in learning curves. The Journal of the Acousti-cal Society of America Express Letters, 142(6), 519-524.

Trudgill, P. (1981). Linguistic accommodation: Sociolinguistic observations on sociopsycho-logical theory. In R. Hendrick, C. Masek & M. F. Miller (Eds.), Paper from the parasession on language and behavior (pp. 218-237). Chicago: Chicago Linguistics Society.

Van der Harst, S., & van de Velde, H. (2008). 17 g's in het Standaardnederlands? Taal en Tongval, 59, 172-195.

Van de Velde, H., Gerritsen, M., & Van Hout, R. (1996). The devoicing of fricatives in Stand-ard Dutch: A real-time study based on radio recordings. Language Variation and Chan-ge, 8(2), 149-175.

Van de Velde, H., & van Hout, R. (2001). The devoicing of fricatives in a reading task. Lin-guistics in the Netherlands, 21(1), 47-57.

Van de Velde, H., van Hout, R., & Gerritsen, M. (1997). Watching Dutch change: A real time study of variation and change in standard Dutch pronunciation. Journal of Sociolinguis-tics, 1(3), 361-391.

37Milou Leeijen

Meertaligheid binnen Brabant Morfo-syntactische overeenkomsten in de Noord-Meierijse en Kleverlandse dialecten binnen Noord-Oost-Brabant

Bilingualism within Northern Brabant: Morpho-syntactic similarities between the Noord-Meierijs and Kleverland dialects

within the North-East Brabantish dialects

M.M. (Milou) Leeijen Manuscript geschreven tijdens haar MA Meertaligheid en Taalverwerving, Universiteit Utrecht, Utrecht

KEYWORDSBrabantishKleverlandsmorpho-syntacticphenomenainflection of articlesinflection of pronouns

ABSTRACTThe Brabantish dialect spoken in Northern Brabant (The Netherlands) consists of many non-standardized dialects. In fact, there are three parts within Northern Brabant where a completely different dialect is spoken. One example of such a dialect is Kleverlandish, which is spoken in Land van Cuijk (Swanenberg, 2010). Speakers of this dialect, however, nev-ertheless think of themselves as speakers of Brabantish, despite using different vowel substitutions. In this paper, morphosyntactic phenomena in the Brabantish dialect (spoken in the municipality Bernheze), and the Kleverlandish dialect (spoken in municipality Boxmeer, part of Land van Cuijk), are compared using transcribed interviews from the Nederlandse Dialectenbank (Meertens Institute). Similarities were found within five morphosyntactic phenomena: subject doubling, double negatives, inflec-tion of adjectives, inflection of pronouns and inflection of articles. Subject doubling, double negatives and inflection of adjectives were not sufficient-ly present within the dataset; however, speakers of both municipalities use inflection of pronouns and articles in many occasions. The observed similarities might explain why speakers of the Kleverlandish dialect in Northern Brabant have the impression that they are using a Brabantish dialect, as often-used inflections in their dialect greatly resemble mor-phosyntactic phenomena in the neighbouring Brabantish dialects. It is, however, still unclear whether the Kleverlandish and Brabantish varieties share the same grammatical rules considering these phenomena.

1. IntroductionHet Brabants is een verzameling van niet-gestandaardiseerde dialecten, dat gesproken wordt in de Nederlandse provincie Noord-Brabant en de Belgische pro-vincies Antwerpen en Vlaams-Brabant (Swanenberg & Brok, 2008). Er bestaan drie uitzonderingsgebieden betreft het Brabantse dialect binnen Noord-Brabant, waar-onder het Land van Cuijk. Hier wordt in plaats van Brabants, Kleverlands gespro-ken (Swanenberg, 2010). Dit gebied zou historisch gezien een geografische grens hebben gehad in de vorm van het Peelmoeras, dat het Land van Cuijk afscheidde


van de rest van Noord-Oost-Brabant (Swanenberg, 2010). Hier loopt nu ook de Mid-denpeelweg (zie Figuur 1), die tussen 1924 en 1940 is aangelegd. Swanenberg (2010) concludeert dat, ondanks het Kleverlandse dialect, inwoners van het Land van Cuijk zich Brabander voelen en vinden dat ze met (een lichte afwijkende) Bra-bantse tongval spreken. Dit is opmerkelijk, aangezien het Kleverlands ook wordt gesproken in Zuid-Oost-Gelderland, in Noord-Limburg en over de grens in Duitsland (Giesbers, 2008).

Figuur 1. Kaart Land van Cuijk en Middenpeelweg (N277) (Swanenberg & Brok, 2008, p.32).

De verschillen tussen het Brabants aan de ene kant van de Middenpeelweg, wat in dit geval de variëteit Noord-Meierijs betreft, en het Kleverlands aan de andere kant van de Middenpeelweg, vindt men vooral op het gebied van klinker-uitspraak en het lexicon (Swanenberg, 2010; Giesbers, 2008). Het Kleverlands kenmerkt zich vooral met uu-klanken in plaats van ui-klanken (huus voor huis) en ie-klanken in plaats van ij-klanken (zien voor zijn) (Swanenberg, 2010). De enige morfolo-gische eigenschap uit het Kleverlands die wordt beschreven in de literatuur, en ook in (de meeste) Brabantse dialecten bestaat, is verdubbeling van het subject. Giesbers (1989) rapporteert aan de hand van het Kleverlandse dialect dat in Ot-tersum gesproken wordt, wat tegen de grens van Land van Cuijk aan ligt, dat

39Milou Leeijen

subjectverdubbeling voorkomt in de vormen van bende gij, ziede gij, en hedde gij. De verdubbeling van het subject zit bij dit laatste voorbeeld in het enclitisch pronomen hedde, wat de persoonsvorm al uitdrukt (‘heb je’), waarna het dubbele subject gij ook nog volgt. Deze subjectverdubbeling komt enkel voor bij inversie, wat bijvoorbeeld in vraagzinnen optreedt (Visser, Nortier & Swanenberg, 2015).

In sectie 1.1 beschrijf ik welke morfo-syntactische eigenschappen typisch zijn voor het Brabantse dialect, waarna ik in sectie 1.2 inga op wat erover bekend is in het Kleverlands. In sectie 1.3 bespreek ik een mogelijk onderliggend verband tussen de twee dialecten, waarna ik in 1.4 de onderzoeksvraag en mijn verwachtingen toelicht. In hoofdstuk 2 beschrijf ik de methode die in dit onderzoek is gebruikt. In hoofdstuk 3 staan de resultaten beschreven van het onderzoek, uitgesplitst in de vijf morfo-syntactische verschijnselen. In hoofdstuk 4 volgt de discussie inclusief suggesties voor vervolgonderzoek, waarna in hoofdstuk 5 de conclusie wordt gepresenteerd.

1.1 Typisch Brabantse morfosyntactische verschijnselenEnkele morfosyntactische verschijnselen zijn typisch voor het Brabantse dia-lect. Visser, Nortier en Swanenberg (2015) categoriseren de volgende zeven verschijnselen: (1) verdubbeling van het subject (‘hedde gij’); (2) dubbele nega-tie (‘niks nie’); (3) dubbele voegwoorden; (4) dubbele voltooid deelwoorden; (5) bijvoeglijke naamwoorden die hun uitgang verliezen in de vrouwelijke vorm (‘een klein kat’) of in de meervoudsvorm; (6) de verbuiging (genusgebonden inflectie) van voornaamwoorden (‘deezen boer’ en ‘dees boerin’); en (7) de verbuiging van lidwoorden (‘unne jonge’ versus ‘un meid’) (Swanenberg & Meulepas, 2011; Vis-ser, Nortier & Swanenberg, 2015). Het verdubbelen van voltooid deelwoorden wordt ook wel hulpwerkwoordverdubbeling genoemd (Meertaligheid en Taal-stoornissen, z.d.). Een voorbeeld is ‘Ik hè da gezeed gehad’, waarbij gehad aan de zin is toegevoegd of verdubbeld van het hulpwerkwoord heb (hè).

1.2 Kleverlandse grammaticaDe morfologie als onderdeel van de Kleverlandse grammatica is tot op heden niet uitgebreid beschreven. Onderzoek naar het Kleverlands bestaat voornamelijk uit dialectgebruik versus Standaardnederlands en Standaardduits, taalattitudes tegenover het Kleverlandse dialect, uitspraak van het dialect en lexicale verschillen binnen het dialectgebied (Cornelissen, 1995). Taaldivergen-tie bij de Nederlands-Duitse staatsgrens krijgt hierbij veel aandacht (Cornelissen, 1995). Het enige morfosyntactische verschijnsel dat wordt besproken, is het ee-rdergenoemde verdubbeling van het subject.

1.3 Twee gescheiden dialecten of één grammatica?Delen het Kleverlands en het Noord-Meierijs een stukje grammatica, omdat ze dezelfde constructies gebruiken? Volgens Roeper (1999) bestaan er ‘co-existing grammars’ binnen een spreker. Dit houdt in dat meerdere grammatica’s naast elkaar bestaan. Zo kan een spreker van het Nederlands bijvoorbeeld een gram-


matica bezitten van zijn of haar kindertaal, van het Standaardnederlands en van zijn of haar dialect. Deze gaan niet verloren en kunnen ook te allen tijde gebruikt worden. Padovan et al. (2016) hebben aangetoond dat talen elkaar kunnen beïn-vloeden, zowel binnen sprekers als tussen sprekers. Zo is het ook mogelijk dat tal-en elkaar op zinsniveau (de output) kunnen beïnvloeden. Twee varianten van het Nederlands, in dit geval de dialecten Noord-Meierijs en het Kleverlands, zouden elkaar hebben kunnen beïnvloed. In eerste instantie wordt er hier gesproken over twee gescheiden dialecten: het Noord-Meierijs en het Kleverlands. Echter, wan-neer deze dialecten gebruik maken van eenzelfde stukje grammatica, zou dit kun-nen duiden op een gedeelde grammatica. Dit kan mogelijk een gevolg zijn van de aanleg van de Middenpeelweg, waarna de dialecten meer in contact zijn gekomen met elkaar.

1.4 OnderzoeksvraagIn het licht van bovenstaande, wil ik in kaart brengen of er overeenkomsten zijn op morfosyntactisch gebied tussen het Noord-Oost-Brabants, in dit geval het Noord-Meierijs dat gesproken wordt in gemeente Bernheze aan de ene kant van de Middenpeelweg, en het Kleverlands in gemeente Boxmeer, wat gesproken wordt aan de andere kant van de Middenpeelweg. Als er weinig overeenkomsten zijn, dan zou dit kunnen duiden op een grens die beide soorten dialecten uit elkaar houdt. Als er meerdere morfosyntactische verschijnselen overeenkomen, kan dit op verschillende situaties duiden. Ten eerste zouden de twee soorten gram-matica’s elkaar kunnen hebben beïnvloed door taalcontact, al is taalcontact niet noodzakelijk de oorzaak van mogelijke overeenkomsten tussen de dialecten, en ten tweede zouden beide dialecten los van elkaar gebruik kunnen maken van één grammaticasysteem op het gebied van de overeenkomende morfosyntac-tische verschijnselen. Dit leidt tot de volgende onderzoeksvraag: 'Welke van de vijf morfosyntactische verschijnselen ((1) verdubbeling van het subject; (2) dub-bele negatie; (3) bijvoeglijke naamwoorden; (4) verbuiging van voornaamwoorden en (5) verbuiging van lidwoorden) die voorkomen in het Brabantse dialect dat gesproken wordt in gemeente Bernheze (Noord-Meierijs), komen ook voor in het Kleverlandse dialect, dat gesproken wordt in gemeente Boxmeer, ook gelegen in Noord-Oost-Brabant, behorend tot het Land van Cuijk?'

Ik verwacht aan de hand van Giesbers (1989) en Visser, Nortier en Swanenberg (2015) in zowel de gemeente Boxmeer als in de gemeente Bernheze verdubbe-ling van het subject tegen te komen. Aan de hand van de literatuur kan ik geen voorspelling doen over de aanwezigheid van de overige morfo-syntactische cate-gorieën (die typisch voorkomen in het Brabants) in het Kleverlands. Wel verwacht ik, dankzij de Brabantse identiteit van de Kleverlandse sprekers, dat zij de overige morfo-syntactische categorieën ook zullen gebruiken, maar in mindere mate dan de Brabantse sprekers. Ik verwacht dat de Brabantse sprekers alle vijf de morfo-syntactische verschijnselen gebruiken in semi-spontane spraak.

41Milou Leeijen

2. MethodeOm vast te stellen of de vijf morfosyntactische verschijnselen in het Kleverlands voorkomen, dat gesproken wordt in gemeente Boxmeer, evenals in het Brabants (Noord-Meierijs) voorkomen, heb ik opnames uit de Nederlandse Dialectenbank (Zeldenrust, z.d.) getranscribeerd.

2.1 MateriaalDe opnames bestaan uit interviews die de vorm van vrije gesprekken hebben, waarin er over onderwerpen gesproken wordt die de dialectsprekers zelf kiezen. Bij de interviews zijn op zijn minst twee sprekers, naast de interviewer, aanwezig. De hele interviews duren allemaal langer dan een half uur, behalve de opname uit Boxmeer. Die duurt minder dan twintig minuten. De opnames zijn sinds de jaren vijftig gemaakt door het Meertens Instituut en zijn vrij te gebruiken voor onderzoek. De opnames zijn dus gemaakt na de aanleg van de Middenpeelweg. Op basis van de onderwerpen die besproken worden in de opnames, schat ik dat deze rond de jaren zeventig gemaakt zijn.

De meeste sprekers in de interviews zijn mannen. Ook die sprekers die het meeste aan het woord zijn (met het label SP1), zijn mannen, behalve van het in-terview uit Heesch; daar is een vrouw het meeste aan het woord. Bij het interview in Nistelrode was ook een vrouw aanwezig, dit was 'spreker 2'. De opname uit Oef-felt heeft geen genummerde sprekers, omdat de mannenstemmen dusdanig op elkaar leken, dat het onderscheid maken tussen de sprekers niet is gelukt. In de interviews bleken vaak dezelfde onderwerpen besproken te worden: werk, kermis, oorlog en doop.

2.2 ProcedureOpnames uit de plaatsen Heesch, Nistelrode en Vorstenbosch, gelegen in ge-meente Bernheze, evenals opnames uit Boxmeer, Oeffelt en Rijkevoort, gelegen in gemeente Boxmeer, zijn beluisterd en daarvan zijn gemiddeld twaalf minuten getranscribeerd. De eerste minuut is telkens overgeslagen, omdat de sprekers nog moeten wennen aan de opnameapparatuur. De transcripten omvatten twee tot drie verschillende sprekers per opname. Zinnen die onverstaanbaar waren, zijn achterwege gelaten. De volgende eigenschappen zijn gecategoriseerd en geanalyseerd: (1) verdubbeling van het subject; (2) dubbele negatie; (3) bijvoeg-lijke naamwoorden; (4) verbuiging van voornaamwoorden en (5) verbuiging van lidwoorden.

De transcripten zijn voorzien van een uniek label voor iedere spreker. Uitspraken en vragen van de interviewer zijn buiten beschouwing gelaten in de analyse. De zinnen zijn vervolgens gecategoriseerd aan de hand van de vijf morfosyntactische eigenschappen. Wanneer zinnen herhaald werden, of wanneer de ene spreker een deel van de zin van de andere spreker herhaalde, zijn deze meegeteld in de analyse. Bij de verbuiging van voornaamwoorden en lidwoorden zijn zowel Standaardnederlandse uitingen als de dialectvarianten van deze verschijnselen


bekeken, evenals bijvoeglijke naamwoorden vóór een zelfstandig naamwoord in de vrouwelijke vorm of in het meervoud. Hiermee is berekend hoe vaak de sprek-ers voor de dialectvariant kiezen en hoe vaak voor het Standaardnederlands.

3. Resultaten3.1 Verdubbeling van het subject

De sprekers uit gemeente Bernheze hebben vier zinnen geuit waarin zij het sub-ject verdubbelen. In drie gevallen gebruiken ze de constructie ‘witte gij/ge’ in een vraagzin. Één spreker uit Vorstenbosch gebruikt de subjectverdubbeling niet vol-gens de regels: ‘Moete gè maar is kijke’. Er is hier geen sprake van inversie zoals in een vraagzin, dus is het eigenlijk niet toegestaan om het enclitische pronomen moete (‘moet je’) te combineren met gè (‘je’). Echter, het woord gè lijkt hier te dienen als nadruk in de gebiedende wijs, waarbij het in vergelijkbare combinaties ook op deze manier mogelijk lijkt te zijn (‘Moete gè/gij (maar) is oplette’, ‘Moete gè/gij (maar) is proberen’). De sprekers uit gemeente Boxmeer hebben geen en-kele keer een verdubbeling van het subject geuit in de interviews (zie Tabel 1).

Tabel 1 Aantal uitingen met een verdubbeling van het subject volgens de Brabantse regels.

UitingenHeesch 1Nistelrode 1

Vorstenbosch 2

Totaal gemeente Bernheze 4

Boxmeer 0Oeffelt 0Rijkevoort 0Totaal gemeente Boxmeer 0

3.2 Dubbele negatie

De sprekers uit gemeente Bernheze hebben vijf keer een dubbele negatie geuit, waarbij de zin ook in de ontkenning bedoeld is. De sprekers uit Heesch hebben geen enkele keer een dubbele negatie geuit. De dubbele ontkenningen die geuit zijn, werden gevormd met (i) twee keer nie (‘nie…nie’), (ii) noit nie, (iii) noit niks en (iv) noit gin (‘geen’). De sprekers uit gemeente Boxmeer hebben slechts één keer een dubbele ontkenning geproduceerd. In Oeffelt is een dubbele ontkenning gevormd met twee keer nie (‘nie…nie’). Zie Tabel 2 voor het totaal aantal uitingen per dorp en gemeente met een dubbele negatie.

43Milou Leeijen

Tabel 2 Aantal uitingen met een dubbele negatie volgens de Brabantse regels.


Vorstenbosch 4



3.3 Bijvoeglijke naamwoorden Er zijn verschillende bijvoeglijke naamwoorden die hun uitgang kunnen verliezen in de vrouwelijke vorm. In gemeente Bernheze zijn er acht uitingen die volgens de Brabantse regels verbogen zijn, waaronder den/de heel/hil buurt (‘de hele buurt’), un olling week (‘een hele week’), goei soep (‘goede soep’), un mooi sprei (‘een mooie sprei’) en un heel modern fabriek (‘een heel/hele moderne fabriek’). Er zijn drie bijvoeglijke naamwoorden die hun uitgang verliezen in de meervoudsvorm: die heel klein keuterboerkes (‘die heel/hele kleine keuter-boertjes’), ‘iel èrem streken (‘heel/hele arme streken’) en da jong vrollie (‘die jonge vrouwen’/’dat jonge vrouwenvolk’). Er komen geen vrouwelijke vormen of meervoudsvormen voor in gemeente Bernheze, die volgens het Standaardned-erlands gevormd worden; ze volgen allemaal de Brabantse regels wat betreft de bijvoeglijke naamwoorden. In de getranscribeerde interviews in Vorstenbosch zijn er geen zinnen met bijvoeglijke naamwoorden gevonden.

In gemeente Boxmeer worden er ook bijvoeglijke naamwoorden verbogen. In Rijk-evoort vindt er verbuiging plaats volgens de Brabantse regels voor de vrouwelijke vorm en valt de verbuigings -e weg; ánder/ander schol (‘andere school’). Opmer-kelijk is dat in Oeffelt ‘de ánder joar’ wordt geuit, terwijl hier waarschijnlijk ‘de andere jaren’ bedoeld wordt, en niet ‘het andere jaar’. Toch vervalt hier de uitgang -e voor het ogenschijnlijke onzijdige woord ‘jaar’.

Verder zijn er bijvoeglijke naamwoorden die voor een mannelijk zelfstandig naam-woord staan, waar iets opmerkelijks mee gebeurt. Sommige bijvoeglijke naam-woorden krijgen een extra -n of -en op het eind. Spreker 1 uit Heesch heeft de volgende combinaties gemaakt: ‘nen groten dikken doek’ (‘een groot dik doek’) en ‘unnen krommen èrrum’ (‘een kromme arm’). Ook de combinatie met een onzijdig zelfstandig naamwoord gaat bij deze spreker niet geheel volgens de regels: ‘een groewte fist’, waarbij de uiting ‘een groot feest’ zou moeten zijn en niet ‘een grote feest’. Daarnaast voegen de sprekers uit Nistelrode ook een -n toe in de manneli-


jke vorm: ‘een groewten busje’ (‘een groot busje’) en ‘de lesten twee’ (‘de laatste twee’). In gemeente Boxmeer worden er ook bijvoeglijke naamwoorden geuit die tegen de regels in gaan. In Oeffelt krijgt ‘hele’ ook een extra -n: ‘hillen’ (‘hillen weg’ en ‘hillen berrum’). Tevens wordt ‘goei schik’ (‘goede schik’) geuit, waar het om een mannelijke vorm gaat en de verbuigings -e niet weg zou mogen vallen. Zie Ta-bel 3 voor een overzicht van het aantal uitingen waarin een bijvoeglijk naamwoord de uitgang –e verloren heeft.

Tabel 3 Aantal uitingen met een bijvoeglijk naamwoord waarbij de uitgang –e verloren is gegaan in de vrouwelijke vorm of meervoudsvorm volgens de Brabantse regels.


Vorstenbosch 0



3.4 Verbuiging (genusgebonden inflectie) van voornaamwoorden De volgende voornaamwoorden zijn in gemeente Bernheze volgens de Brabantse verbuigingsregels gevormd: ‘gin(ne(n))’ (‘geen’): ginnen dokter, gin boks, gin in-dustrie, gin navraag; ‘ons/onzen’: vijf keer ons moeder, onzen vodder (‘onze vad-er’), ons vrouw; ‘dees’ (‘deze’): dees stroat en ‘mun’ (‘mijn’): mun rats, mun cente. ‘Die’ wordt dertien keer in de meervoudsvorm volgens de Brabantse regels ge-bruikt. Daarnaast wordt ‘dieje(n)’ acht keer volgens de Brabantse regels gebruikt voor de mannelijke vorm: diejen oard, diejen os, diejen tijd, diejen baas, dieje ambtenarij en dieje èrme mens. Het voornaamwoord ‘geen’ is in gemeente Bern-heze ook een aantal keer niet volgens de Brabantse verbuigingsregels gevormd, maar volgens het Standaardnederlands: gin os, gin bioscoop en gin man. ‘Os’, ‘bioscoop’ en ‘man’ zijn mannelijke woorden en zouden ‘ginne(n)’ kunnen krijgen, als de spreker de Brabantse regels zou volgen. Ook ‘die’ wordt één keer gebruikt bij de mannelijke vorm: die mister. In 89 procent van de gevallen is er wel voor de Brabantse verbuiging voor de mannelijke vorm gekozen. Opmerkelijk is dat de Vorstenbosche spreker ‘hun’ verbuigt naar ‘urre’ (‘urre jas’).

De volgende voornaamwoorden zijn in gemeente Boxmeer volgens de Brabantse verbuigingsregels gevormd: ‘gin(ne(n))’ (‘geen’): ginne cént, gin uur, gin kostwin-ning en ginnen enne; ‘mien/munne(n)’ (‘mijn’): munnen vóder en munnen halfse broer en ‘ons’: ons voete. ‘Voeten’ is meervoud en daardoor kan het voornaam-

45Milou Leeijen

woord de uitgang verliezen. ‘Die’ wordt 38 keer in de meervoudsvorm en vrou-welijke vorm volgens de Brabantse regels gebruikt. Daarnaast wordt ‘dieje(n)’ zes keer volgens de Brabantse regels gebruikt voor de mannelijke vorm: die(je)n boer, diejen hillen weg, dieje hillen berrum, dieje weg, diej(e)n hoek en diejen ander schol (‘die andere school’). In zes gevallen van de mannelijke vorm, is de Standaardnederlandse vorm gekozen (‘die’ en niet ‘dieje(n))’. Het gaat hier bij-voorbeeld om de mannelijke woorden: tijd, weg en voorzitter. In 50 procent van de gevallen wordt de Brabantse verbuiging van ‘die’ gekozen. Het voornaamwoord ‘geen’ is in gemeente Boxmeer ook een keer niet volgens de Brabantse verbui-gingsregels gevormd: gin mieter (‘geen meter’). Daarnaast is ook ‘mijn’ niet altijd volgens de mannelijke vorm verbogen: mien vóder en mien vader. Opvallend is dat de sprekers uit gemeente Boxmeer in totaal vaker een verbogen voornaam-woord gebruiken dan de sprekers uit gemeente Bernheze, maar dat het percen-tage van de verbogen vorm lager is. Zie Tabel 4 voor een overzicht van het aantal uitingen waarin een verbogen voornaamwoord voorkomt.

Tabel 4 Aantal uitingen met verbogen voornaamwoorden volgens de Brabantse regels.

Uitingen

Heesch 2Nistelrode 6

Vorstenbosch 6



3.5 Verbuiging (genusgebonden inflectie) van lidwoordenIn gemeente Bernheze worden de mannelijke onbepaalde lidwoorden negen keer verbogen volgens de Brabantse regels: unnen krommen errum (‘een kromme arm’), unnen grooten kool (‘een grote kool’), unnen hillen gróete (‘een hele grote’), unne keer (‘een keer’), unne zekere middag (‘een zekere middag’), unne speek (‘een spaak’), unnen ovend (‘een avond’), unne jas (‘een jas’) en unnen derde-jaars student (‘een derdejaars student’). ‘Un’ wordt in gemeente Bernheze 34 keer volgens de Brabantse regels gebruikt voor de vrouwelijke en onzijdige vorm, evenals het meervoud. Daarnaast zijn er ook twaalf gevallen waarin ‘un’ is geuit, waar dit ook ‘unne’ had kunnen zijn, omdat het vóór een mannelijk zelfstandig naamwoord kwam. In 57 procent van de gevallen wordt er volgens het Standaard Nederlands verbogen voor de mannelijke vorm en voldoet 43 procent aan de Brabantse verbuigingsregels. ‘Enne(n)/inne(n)’ komt ook zelfstandig voor met de betekenis ‘één’ of ‘eentje’. In gemeente Bernheze is dit vijf keer geuit: drie keer


door de Heesche spreekster en twee keer door de Vorstenbosche spreker. Ook ‘één’ is twee keer voorgekomen in de uitingen van de Vorstenbosche spreker.

Gebruik van unne dat tegen de Brabantse regels ingaat, komt ook een keer voor in gemeente Bernheze; in Nistelrode wordt ‘unnen uur’ geuit, terwijl uur een onzijdig woord is. Volgens de grammatica zou het daardoor ‘un uur’ moeten zijn. Of er in de zin ‘Unne groete lange wándelhanze’, die door de Heesche spreekster geuit is, van de Brabantse regels gebruik wordt gemaakt van de inflectie op het lidwoord, kan ik niet vaststellen. Dit komt doordat ik niet weet wat een wándel-hanze is en of dit een mannelijk, vrouwelijk of onzijdig woord is.

Het bepaald lidwoord ‘den/d’n’ komt in gemeente Bernheze negentien keer vol-gens de Brabantse regels voor in de mannelijke vorm. ‘De’ komt 29 keer volgens de Brabantse regels voor in de vrouwelijke vorm en meervoudsvorm. Daarnaast is er ien keer voor het Standaards Nederlands gekozen (‘de’), waar dit de mannelijke vorm betreft. Hier had dus ‘den/d’n’ gezegd kunnen worden. In 65 procent van de gevallen is het Brabantse mannelijke lidwoord geuit. Twee keer kwam het voor dat er een naam genoemd werd waar ‘den’ in voorkwam. Deze zijn buiten de ana-lyse gelaten. Tot slot is er één keer in Bernheze tegen het patroon in gegaan; de spreekster uit Heesch gebruikte ‘den’ in combinatie met het meervoud ‘hoeken’, waar het lidwoord ‘de’ bij hoort vanwege het meervoud.

In de gemeente Boxmeer produceerde men vijf keer het onbepaald mannelijk lidwoord volgens de Brabantse regels; unne méter (‘een meter’), unnen diek (‘een dijk’), unne gewone mik (‘een gewoon brood’) en unne cent (‘een cent’). De Rijkevoortse sprekers hebben geen enkele keer het Brabantse mannelijke lid-woord ‘unne/unnen’ geproduceerd. ‘Un’ wordt in gemeente Boxmeer 31 keer vol-gens de Brabantse regels gebruikt voor de vrouwelijke en onzijdige vorm, evenals het meervoud. Daarnaast zijn er ook egen gevallen waarin ‘un’ is geuit, waar dit ook ‘unne’ had kunnen zijn, omdat het vóór een mannelijk zelfstandig naamwoord kwam. In 36 procent van de gevallen is er de mannelijke vorm gekozen die voldoet aan de verbuigingsregels van het Brabants.

Daarnaast komt ‘enne(n)/inne(n)’ in gemeente Boxmeer ook zelfstandig voor met de betekenis ‘één’ of ‘eentje’: in Boxmeer is dit niet voorgekomen, in Oeffelt is dat één keer geuit, en in Rijkevoort negen keer. Wel is ‘ene’ voorgekomen in de uitingen van een spreker uit Boxmeer: ‘ene keer’. Ook in gemeente Boxmeer komt er lidwoordgebruik voor dat tegen de Brabantse regels ingaat. In Rijkevoort wordt er ‘unne cente’ gezegd, terwijl het hier om het meervoud van cent gaat. Het zou geen onbepaald lidwoord mogen krijgen als het de Brabantse regels zou volgen.

Het bepaald lidwoord ‘den/d’n’ komt in gemeente Boxmeer twintig keer voor in de mannelijke vorm volgens de Brabantse verbuigingsregels. ‘De’ komt 53 keer vol-gens de Brabantse regels voor in de vrouwelijke vorm en meervoudsvorm. Daar-naast is er twintig keer voor het Standaards Nederlands gekozen (‘de’), waar dit

47Milou Leeijen

de mannelijke vorm betreft. Hier had dus ‘den/d’n’ gezegd kunnen worden. In 50 procent van de gevallen is het Brabantse mannelijke lidwoord geuit. ‘Den’ en ‘de’ komen ook een keer voor, waarbij ze tegen de Brabantse regels ingaan: ‘den piepken’ (‘het pijpje’) en ‘de gemeentehuus’ (‘het gemeentehuis’). Opvallend is dat in het totaal in gemeente Boxmeer vaker uitingen voorkomen met een verbo-gen lidwoord volgens de Brabantse regels dan in gemeente Bernheze. Toch is de verhouding tussen de verbogen vorm en de vorm volgens de Nederlandse gram-matica andersom. Zie Tabel 5 voor een overzicht van het aantal uitingen waarin een verbogen lidwoord voorkomt.

Tabel 5 Aantal uitingen met verbogen lidwoorden volgens de Brabantse regels.


Vorstenbosch 6



4. DiscussieIn dit paper heb ik geprobeerd de volgende onderzoeksvraag te beantwoorden: Welke van de vijf morfo-syntactische verschijnselen ((1) verdubbeling van het sub-ject; (2) dubbele negatie; (3) bijvoeglijke naamwoorden; (4) verbuiging van voor-naamwoorden en (5) verbuiging van lidwoorden) die voorkomen in het Brabantse dialect dat gesproken wordt in gemeente Bernheze (Noord-Meierijs), komen ook voor in het Kleverlandse dialect, dat gesproken wordt in gemeente Boxmeer, ook gelegen in Noord-Oost-Brabant, behorend tot het Land van Cuijk? Aan de hand van de resultaten heb ik gezien dat de voorspelling dat er in beide gemeenten ver-dubbeling van het subject plaatsvindt niet is uitgekomen; sprekers in gemeente Boxmeer hebben geen enkele keer een verdubbeling van het subject geuit. Dit wil nog niet zeggen dat ze dit nooit doen, maar in de delen van opnames die getranscribeerd zijn, is het niet voorgekomen. Sprekers uit gemeente Bernheze hebben deze constructie ook slechts vier keer geuit, dus het komt wellicht minder vaak voor in semi-spontane spraak. Ook dubbele negatie komt slechts zes keer voor in de gehele dataset van getranscribeerde zinnen, waardoor er geen nieuwe kennis is vergaard op dit gebied. Daarnaast is de voorspelling dat de sprekers uit gemeente Boxmeer minder vaak de morfo-syntactische constructies zouden gebruiken dan de sprekers uit gemeente Bernheze ook niet geheel uitgekomen, doordat de absolute resultaten van de sprekers uit gemeente Boxmeer hoger uit-


vallen dan de absolute resultaten van de sprekers uit gemeente Bernheze.

In tegenstelling tot bovengenoemde verschijnselen, is er wel nieuwe informatie over verbuiging van voornaamwoorden en lidwoorden. Het voornaamwoord ‘die’ werd door de sprekers uit Bernheze 89 procent verbogen in de mannelijke vorm ‘diejen’, terwijl de sprekers uit Boxmeer dit in 50 procent van de gevallen deden. Dit is een groot verschil. Het zelfstandige voorkomen van ‘enne(n)/inne(n)’ komt in gemeente Boxmeer twee keer zo vaak voor als in Bernheze: tien keer tegeno-ver vijf keer. In 43 procent van de gevallen worden de Brabantse onbepaalde lidwoorden in Bernheze gebruikt en in 65 procent bepaalde lidwoorden, tegeno-ver 36 en 50 procent van de Kleverlandse sprekers respectievelijk. Lidwoorden komen veel voor in de taal. Doordat op meer dan eenderde van de gevallen het onbepaald lidwoord de Brabantse verbuigingsregels worden toegepast door zowel sprekers uit gemeente Bernheze als sprekers uit gemeente Boxmeer, en op op zijn minst de helft van de gevallen van bepaalde lidwoorden de Brabantse ver-buigingsregels worden toegepast in beide gemeenten, zou dit kunnen verklaren waarom sprekers van het Kleverlands vinden dat zij met een Brabantse tongval spreken.

Een suggestie voor verder onderzoek is het analyseren van meerdere opnames uit het Land van Cuijk. De dorpen die op de grens van de Middenpeelweg liggen zouden andere resultaten hebben kunnen opgeleverd, omdat in die dorpen het taalcontact met het Brabantse dialect misschien groter is geweest. Deze dorpen zouden een overgangsgebied kunnen zijn wat betreft de morfosyntactische ei-genschappen. Bovendien zou zowel een diachroon als een synchroon longitudi-naal onderzoek een toegevoegde waarde zijn en inzicht bieden in het effect van taalcontact en de aanleg van de Middenpeelweg. Echter, deze methode kent be-perkingen vanwege het ontbreken van data en opnamemateriaal. De omliggende gebieden waar het Brabants (Noord-Meierijs) gesproken wordt, zou daarbij ook in de vergelijking meegenomen moeten worden.

Daarnaast is het ook nodig om meer onderzoek te verrichten naar deze verschijnselen op zichzelf, om uitspraken te kunnen doen over de vraag of het Kleverlands en het Noord-Meierijs berusten op dezelfde grammatica of niet. Hier-bij zou er bovendien onderzocht kunnen worden of de morfosyntactische verschi-jnselen ook voorkomen in het Kleverlands dat niet in Noord-Brabant gesproken wordt, of in andere niet-Brabantse dialecten.

5. ConclusieDe vijf morfo-syntactische verschijnselen ((1) verdubbeling van het subject; (2) dubbele negatie; (3) bijvoeglijke naamwoorden; (4) verbuiging van voornaam-woorden en (5) verbuiging van lidwoorden) komen niet allemaal voor in de op-names die in dit onderzoek zijn getranscribeerd. Zo heeft er geen verdubbeling van het subject plaatsgevonden in gemeente Boxmeer. Ook dubbele negatie is in gemeente Boxmeer slechts één keer geuit. Bijvoeglijke naamwoorden waarbij de

49Milou Leeijen

verbuigings-e die verloren gaat in de vrouwelijke en meervoudsvorm, zoals in het Brabants gebruikelijk is, komt drie keer vaker voor in gemeente Bernheze dan in gemeente Boxmeer. Daarnaast komen verbuigingen van voornaamwoorden en verbuigingen van lidwoorden volgens de Brabantse verbuigingsregels in beide dialecten veel voor. Met deze resultaten is er geen dialectgrens aangetoond op gebied van de vijf morfosyntactische verschijnselen. Het taalcontact dat na de aanlegging van de Middenpeelweg mogelijk werd, is een mogelijke oorzaak van de Brabantse morfo-syntactische eigenschappen die voorkomen in het niet-Bra-bantse dialect dat wel in Brabant gesproken wordt: het Kleverlands (Swanenberg, 2010). ■

Received April 2019; accepted September 2019.

ReferencesCornelissen, G. (1995). De dialecten in de Duits-Nederlandse Roerstreek-grensdialectolo-

gisch bekeken. Vereniging voor Limburgse Dialect- en Naamkunde. Giesbers, C. H. E. (2008). Dialecten op de grens van twee talen. een dialectologisch en

sociolinguïstisch onderzoek in het Kleverlands dialectgebied (Dissertatie). Groesbeek: Reijngoudt-Giesbers.

Giesbers, H. W. M. (1989). Code-switching tussen dialect en standaardtaal. Amsterdam: PJ Meertens-Instituut voor Dialectologie, Volkskunde en Naamkunde.

Meertaligheid en Taalstoornissen. (z.d.). Syntaxis dialecten. Geraadpleegd van https://meertaligheidentaalstoornissenvu.weebly.com/syntaxisdialecten.html#

Padovan, A., Tomaselli, A., Bergstra, M., Corver, N., Etxepare, R., Dold, S. (2016). Minority languages in language contact situations: three case studies on language change. Us Wurk. Tydskrift foar frisityk, 65(3-4), 146-174.

Roeper, T. (1999). Universal Bilingualism. Bilingualism: Language and Cognition 2(3), 169-186.

Swanenberg, J. (2010). Dialectgrenzen in Noord-Brabant, tussen 1937 en 2009. Taal en Tongval 62(1), 71-85.

Swanenberg, A. P. C., & Brok, H. (2008). Het Brabants beschreven. Alphen aan de Maas: Veerhuis.

Swanenberg, J., & Meulepas, J. (2011). Het nieuwe Brabants. Een onderzoek naar diver-siteit in taal en cultuur onder jongeren in Noordoost-Brabant. Taal en Tongval 63(2), 303-327.

Visser, P., Nortier, J., & Swanenberg, J. (2015). Dialect als stereotypering Regionale identi-teit in Noord-Brabant verbeeld en verwoord op Facebook. In Artikelen van de 8e Anéla Conferentie, 111-125. Utrecht: Uitgeverij Eburon.

Zeldenrust, D. (z.d.). Nederlandse Dialectenbank. Geraadpleegd van http://www.meer-tens.knaw.nl/ndb/gmaps.php?t=nl


Prepositions and language transferImproving recall in automatic error correction

D. (Dylan) BongaManuscript written during his RMA Linguistics, Utrecht University, Utrecht

KEYWORDSAutomatic error correctionlanguage transferprepositionssecond language acquisition

ABSTRACTIn second language learning, prepositions are often a source of mistakes. These mistakes can be analysed using automatic error correction, such as the Determiner And PrePosition Error Recogniser (De Felice & Pulman, 2008; 2009). However, these models often have a low recall level. This means that the model is not able to find the targeted data. I argue that the recall level can be improved by including possible language transfer from the native language (L1) to the model. To support this claim, I pro-pose to replicate the study done by Jarvis & Odlin (2000), who studied Swedish and Finnish learners of English as a second language. However, I propose a replication with Hungarian speakers and Hungarian speakers of German. If agglutinative L1s (Finnish and Hungarian) show a similar distribution, and a different distribution from non-agglutinative L1s (Swed-ish and German), this could be implemented to improve the recall of the preposition error correction models.

1. IntroductionWhen learning a second language, typical learners will make many mistakes. Some of these mistakes will be resolved early on. However, some aspects of grammar remain more difficult to learn, such as the correct use of prepositions. The findings of an early study on English as a Second Language (ESL) show that prepositional errors are found in 18% of sentences written by learners of English, from 15 differ-ent language backgrounds (Dalgish, 1985). Although it is hard to generalize over such a diverse group of participants, one can cautiously assume that prepositions are problematic in the context of second language acquisition (SLA), at least in lan-guage writing. This assumption is further backed by Bitchener, Young, & Cameron (2005), who performed an experiment with 53 participants from a great variety of language backgrounds. They reported the mistakes made in an informal writing task and found that around 30% of the errors involved prepositions.

One way of resolving the mistakes made with prepositions, is constructing an auto-matic preposition error recogniser and using it to correct the errors. Such a system could make correcting for teachers much easier, but more importantly, it could also support second language learners when they are writing in their L2. Such systems were constructed, for example, by Chodorow, Tetreault, & Han (2007), as well as De Felice & Pulman (2008, 2009). Both use a maximum entropy classifier. This type of algorithm is used because the features observed in the data set (for instance:

51Milou Leeijen

words) are not conditionally independent. Even if they were, the maximum entropy classifier works more reservedly, whereas the alternative (a naive Bayesian classi-fier) would be too greedy. Chodorow et al. (2007) also showed that a naive Bayes model performs worse than the maximum entropy classifier they used.

In order to see how successful a model is, there are different measures to report. Three important measures that are very often used are accuracy, precision, and re-call. Accuracy and precision are often seen as synonyms, which is incorrect. Table 1 gives the possible outcomes of a classifier. The observed values are the values that are considered the ‘correct’ values, whereas the predicted values are the values that the model assigns. Accuracy, precision, and recall are calculated by combining the outcomes from Table 1.

Table 1 Classification table of predicted and observed values.

Observed values

Predicted values True FalseTrue True positive False positiveFalse False negative True negative

The definitions of the quality measures mentioned above are given in (1-3), as functions of the outcomes in Table 1. The abbreviations in (1)-(3) resemble the outcomes in Table 1. Applied to preposition correction, precision would be the amount of true corrections divided by the total amount of corrections. Recall can be interpreted as the amount of true corrections divided by the amount of cases where correction was needed. Lastly, accuracy is the amount of truly analysed cases (so correction if needed and no correction if no correction needed), divided by all counted instances.

(1) Precision: TP / (TP + FP)

(2) Recall: TP / (TP + FN)

(3) Accuracy: TP + TN / (TP + FP + FN + TN)

The system used in Chodorow et al. (2007) reached a precision of 80%. They re-port precision, instead of accuracy, because they stress that it is most important to have as little false positives as possible. They also report a recall level around 30%. This means that the system is unable to detect a great part of all made mis-takes. De Felice & Pulman (2009) reported an accuracy of 70%, and a recall of 35%.1 In other words, both systems process the data that they find fairly well, but they do not find much of the targeted data. Chodorow et al. (2007) reported that a low level of recall is typical for the domain of error detection.

1 Additionally, De Felice & Pulman (2008) report a high accuracy on determiner correction. However, this was only tested on L1 data and it is not relevant to this proposal.

Dylan Bonga


Alternatively, one could use unsupervised learning. This approach is taken by Is-lam & Inkpen (2010) for L1 data and Islam & Inkpen (2011) for L2 data. Islam & Inkpen (2011) reported that this method performs well, as both accuracy and re-call are quite high. However, this study does not only consider prepositional errors, making this model incomparable to the maximum entropy classifier. Therefore, this proposal will not consider unsupervised learning as an alternative approach.

In this proposal, I shall further discuss the systems made by Chodorow et al. (2007) and De Felice & Pulman (2008, 2009). I will describe the training data and the classification process in both algorithms. Further, I shall discuss the quality measures that both systems achieve on written L2 data and how they can be interpreted. Both systems display a decent level of accuracy or precision, but a low level of recall. I will propose a follow-up study that further examines the effects of language transfer on prepositional use in L2 writing. Additionally, I will argue that taking language transfer into account in the classifier model may improve the level of recall for the maximum entropy classifier. The main question that this pro-posal will focus on, is formulated as follows: to what extend can the level of recall be improved when using a maximum entropy classifier for automatic preposition error correction?

2. Models for automatic error correctionThis section looks into two models that are used for automatic error correction, namely Chodorow et al. (2007), in section 2.1, and De Felice and Pulman (2008, 2009), in section 2.2. Section 2.3 will briefly discuss how these models compare.

2.1 Classifier with rule-based filtersThe classifier used by Chodorow et al. (2007) is a maximum entropy classifier combined with rule-based filters. These filters function as predefined rules of errors that are easily recognisable, such as redundantly repeated prepositions (‘friends with with the same interests’). This classifier was trained on a part of the MetaMetrics corpus and issues of the Sant Jose Mercury newspaper. The Meta-Metrics corpus contains texts especially adapted to different reading levels. The part that was used in Chodorow et al. (2007) was aimed at 11th and 12th grade. This training data was tagged for Part-of-Speech (POS) and chunked. The test data was a random set of ESL essays. Here, the classifier had to assign a new value (an occurrence of a preposition) to one of 34 classes, with each class resembling one preposition. The assignment was done mainly on basis of the adjacent words, i.e., two or three words to the left and to the right of the target word. The classifier had a confidence parameter, which made the algorithm skip a correction if the top ranked candidates to replace the error were ranked too close to each other. For the testing phase, the threshold for the confidence parameter was set on 0.9. This means that if the algorithm is less than 90 percent certain about a correction, the error is skipped.

53Milou Leeijen

The test data were essays written by ESL students from a Chinese, Russian or Japanese background. Chodorow et al. (2007) reported several problems with the test data. Firstly, they found that there were many spelling errors and punctuation errors. This does not only influence the classifier, but also the POS-tagger, leading to a lower performance. The sentences that contained spelling and punctuation errors were omitted for that reason. The classifier was tested on a total of 2000 preposition contexts. This leads to a decent performance, with a precision meas-ure of around 70%. They added a naive Bayesian classifier, which performs worse when used by itself. This naive Bayesian classifier was allowed to veto a decision made by the maximum entropy classifier, which raised to precision to 88% and lowered the recall even further to about 16%.

2.2 DAPPER The maximum entropy classifier presented by De Felice & Pulman (2008, 2009) is DAPPER, or Determiner And PrePosition Error Recogniser. DAPPER is a classi-fier trained to find and evaluate the use of determiners and prepositions, for the purpose of error correction. For this classifier, the determiner error recogniser is a separate module that functions fully without the preposition error recogniser, and vice versa. Since this proposal only focusses on prepositions, the module that evaluates determiners is left aside.

De Felice & Pulman (2009) explained that their work is different than previous research, because they train the classifier on a wider range of syntactic and se-mantic components, which includes a full syntactic analysis of the context of a preposition. Instead of using a great number of different prepositions, like Cho-dorow et al. (2007), De Felice & Pulman (2009) restricted their search to the nine most frequent prepositions in the British National Corpus. They use about 9 million ‘contexts’, that contain one of these nine most frequent prepositions. All this data is processed with a POS-tagger, a stemmer, a lexicographer for verb and noun classes, a syntactic parser, and a named entity recogniser. Processing the data provides a feature vector, which is a representation of the context and all the extra information gained during processing. Furthermore, they only considered the contexts in which a preposition was needed.

De Felice & Pulman (2009) tested their model on a subset of the Cambridge Learner Corpus, which had been manually tagged for errors. They stripped the cor-pus of all error tags. The corpus contains data from learners from several differ-ent language backgrounds, but De Felice & Pulman (2008, 2009) did not report which backgrounds are represented in the test data. However, Nicholls (2003) reported that the complete corpus has data of 86 mother tongues and that about one third of the corpus is tagged for errors. Like in Chodorow et al. (2007), the cor-pus contains spelling errors. However, De Felice & Pulman (2009) did not omit the sentences containing these errors. The advantage of this approach is that there is more training data and more test data. They do explain that certain mistakes may lead to a worse performance, but also mention that this only concerns 3%

Dylan Bonga


of the data. They do adjust the final level of performance to balance out these, and other, errors. This leads to an accuracy of 69% where the given preposition was used correctly and an accuracy of 42% where it needed to be corrected. They report a recall of 35%.

2.3 ComparisonBoth systems mentioned above perform quite well on a difficult task. They both mention different types of difficulties and how they are coped with. One problem that can be resolved quite easily, is given by De Felice & Pulman (2009). While analysing the data, it came to their attention that the ESL students tend to rely on a small set of lexical chunks, which they overuse. Implementing the use of lexical chunks into the model would improve both precision and accuracy. However, one problem remains unsolved for both systems, which is the low level of recall. The source of this problem is unclear, although De Felice & Pulman (2009) did report that the recall levels vary greatly for the nine different prepositions they used. Even though this is not a priority for error correction algorithms, raising the recall would improve the applicability of the system.

3. Language transfer One possible solution that will boost the recall is to implement L1 influence. Much research has been done on the topic of crosslinguistic influence (CLI), or language transfer. This influence can go from the L1 to the L2, but also the other way around. If there are more languages involved, like an L3 or L4, there can be influence from all the languages spoken by the learner (De Angelis, 2007; Odlin, 2013). The transfer can be positive, which can reinforce the acquisition process, or negative. In this proposal, I will focus mainly on negative transfer from the L1 into the L2. This form of transfer can be seen as interference of the L1, when us-ing the L2.

CLI can be accounted for by most recent linguistic frameworks, such as Feature (re)assembly (Lardiere, 2008; Domínguez, Arche, & Myles, 2011), Usage-Based theory (Tomasello, 2000), and Dynamic System theory (De Bot, Lowie, & Verspoor, 2007). These frameworks make similar predictions on when language transfer happens. Both Feature (re)assembly and Usage-based theory predict that lan-guage transfer happens when the language learner has not fully acquired the new language and is still reassembling features (according to Feature (re)assembly) or adjusting the language mapping (according to Usage-based theory). This does imply that the learner should make no more mistakes when the language is fully acquired (and vice versa, that the language is only fully acquired when the learner makes no more mistakes).

Dynamic System theory rather approaches language learning as a linear function. Thus, the language acquisition grows by 1 plus the level of available resources. This principle covers the difference between learners from different L1s by stating that the resources available (from the L1) are different between learners. It also

55Milou Leeijen

accounts for interlearner differences, since all learners have different resources and thus will learn differently.

As mentioned before, the use of prepositions remains a problem for L2 learn-ers (Dalgish, 1985; Bitchener et al. 2005). Unsurprisingly, language transfer also happens in the use of prepositions, as is shown by Jarvis & Odlin (2000). They show that ESL learners with a Finnish L1 experience more difficulties using spatial-reference prepositions than ESL learners with a Swedish L1. They reason that Finnish learners perform worse, because they agglutinate spatial-reference in their L1, rather than using prepositions like Swedish and English. Examples are given in (4) for Swedish and (5) for Finnish2. Jarvis & Odlin (2000) reported that similar observations have been made by Sulkala (1996) about Estonian learners of English. Estonian, like Finnish, is an agglutinative language and expresses spa-tial reference by means of bound morphology.

(4) i huset in house.specific ‘In the house’

(5) talo -ssa house-Inessive ‘In the house’

In their study, Jarvis & Odlin (2000) let the participants watch a silent film and afterwards made them write a narrative for that film. They tested almost 400 par-ticipants, which results in a great amount of data. Because the writing task does not contain many restrictions for the participants, the resulting data is hard to compare. This is why they analysed the usage of prepositions in these narratives by focussing only on the verbs take and sit. Jarvis & Odlin (2000) explained that Finnish ESL learners prefer to use different prepositions than Swedish ESL learn-ers. Furthermore, the Finnish learners omit prepositions where it is not allowed to omit, while the Swedish learners do not.

Summarized, the study by Jarvis & Odlin (2000) shows that the use of preposi-tions of Finnish and Swedish ESL learners follows a different distribution. This could be implemented into the maximum entropy classifier. Implementing the L1 transfer into a classifier would raise the possibility to focus on a contexts in which a speaker of a certain L1 would make the same mistake, where a speaker of a dif-ferent L1 would not. More concretely, if a Finnish learner would omit prepositions where it is not allowed, where a Swedish learner would (almost) never do this, this could help the classifier analyse these contexts more thoroughly. In this way the classifier could detect more mistakes, thus increasing the level of recall.

2 Special thanks to Walther Glödstaf for helping with these examples.

Dylan Bonga


4. Proposal Even though the conclusions from Jarvis & Odlin (2000) and Sulkala (1996) sug-gest that speakers of an agglutinative language behave similarly in respect to prepositional errors in ESL, there is need for more solid evidence before imple-menting this abstract concept of language transfer into the classifier. For that pur-pose, I propose to replicate the study done by Jarvis & Odlin (2000), but this time using speakers of Hungarian, which is an agglutinative language like, and also closely related to, Finnish. In order to have a comparable experimental group that lives in very similar conditions, like in Jarvis & Odlin (2000), I propose to also test Hungarian participants that have German as their L1, since German is a minority language in Hungary. German, like Swedish and English, uses prepositions and should thus behave differently than Hungarian in this test. Jarvis & Odlin (2000) only focussed on spatial-reference prepositions. Likewise, this paper will also fo-cus on this type of prepositions. Including other types of prepositions would go beyond the scope of this proposal.

As mentioned before, Jarvis & Odlin (2000) predominantly looked at only two dif-ferent verbs in both Swedish and Finnish. Ideally, more verbs would be analysed. However this would go beyond the scope of this proposal, since the analysis is a time consuming process. Thus, the replication study will consider the same verbs. Like in Jarvis & Odlin (2000) the participants will be in their third, fifth, or seventh year of learning English, both for the Hungarian speakers and for the German speakers. The participants will be shown the same 8-minute silent film (Modern Times) and will be asked to write a narrative immediately after, just like in Jarvis & Odlin (2000). Furthermore, a group of adult speakers of English will be compiled to function as a control group. The study is replicated as precisely as possible in order to make the results better comparable. This is necessary, since language transfer is a phenomenon that shows a high level of variation. In order to isolate language transfer, other possible influences (like age of the participants and the task at hand) should be accounted for by exactly replicating Jarvis & Odlin (2000).

5. ImplicationsNot only can the results of the proposed study strengthen the claims Jarvis & Odlin (2000) made, the collected data will also be very useful for comparison. The maximum entropy classifier could be used on the newly gathered data and on the data from Jarvis & Odlin (2000). The results could be helpful in generalizing the results for agglutinative languages.

The experiment I proposed can have multiple outcomes, of which I will discuss three. Firstly, if the results are alike for all four L1s (Finnish, Swedish, Hungarian, and German), or not similar at all, then language transfer might not be traceable (yet) in this data, or at least not with this method. Secondly, it could be possible that the languages used in Jarvis & Odlin (2000) show similar patterns, and the two languages from the replication study do the same. This would imply that the socio-topological environment (Finland versus Hungary) affects the use of prepo-

57Milou Leeijen

sitions in ESL. However, this seems highly unlikely, since Jarvis & Odlin (2000) al-ready showed that Swedish and Finnish learners use prepositions differently and I expect the same results for German and Hungarian learners. Lastly, if preposi-tional errors are indeed influenced by language transfer, the results of the experi-ment would show that the agglutinative languages pattern together and the non-agglutinative languages pattern together. This last outcome would make language transfer a valuable attribute that can improve the maximum entropy classifier.

If indeed I find the third possible outcome, mentioned above, then language trans-fer should be implemented into the maximum entropy classifier, at least to make a difference between agglutinative languages and non-agglutinative languages. This should help increase the recall, as was argued before, because the clas-sifier would have additional conditions or rules that can be applied to analyse a certain context and evaluate it. Importantly, it should be stressed that Jarvis & Odlin (2000), as well as this replication study, only focus on spatial-reference prepositions. Including different prepositions would help improve the classifier even further.

Further developing the maximum entropy classifier has several benefits, both for the teacher and the learner in ESL settings. One way of making the classifier espe-cially helpful for the learner is suggested by De Felice & Pulman (2009): the clas-sifier could mark the error and then rank several possible alternatives. However, they note that this is mostly useful for the more advanced students. Beyond help-ing the learners, the classifier could also be directly applied in everyday situations. For instance, it could be a helpful tool for non-native speakers that need to write an application letter. It could even be added to the grammar correcting software currently used in word processing programs.

Lastly, if the classifier is further improved, it can also be helpful to analyse prepo-sitional errors in ESL corpora. As mentioned above in this proposal, different lan-guages might pattern alike, according to similar language background. Not only could the classifier be used to identify other languages that show similar patterns as Finnish and Hungarian, or Swedish and German, it could also be used to iden-tify new patterns. For instance, there are languages, such as Polish, that some-times use prepositions and sometimes agglutinize. However, further research has to be done in order to permit an analysis of these languages. ■

Received April 2019; accepted April 2019.

ReferencesAngelis, G. D. (2007). Third or additional language acquisition. Multilingual Matters.Bitchener, J., Young, S., & Cameron, D. (2005). The effect of different types of corrective

feedback on ESL student writing. Journal of Second Language Writing, 14(3), 191–205. Chodorow, M., Tetreault, J., & Han, N. R. (2007). Detection of grammatical errors involving

prepositions. Proceedings of the Fourth ACL-SIGSEM Workshop on Prepositions, 25–30. Dalgish, G. (1985). Computer-assisted ESL research and courseware development. Com-

Dylan Bonga


puters and Composition, 2(4), 45-62.De Bot, K., Lowie, W., & Verspoor, M. (2007). A dynamic systems theory approach to second

language acquisition. Bilingualism: Language and Cognition, 10(01), 7–20.De Felice, R., & Pulman, S. (2008). A classifier-based approach to preposition and deter-

miner error correction in L2 English. In Proceedings of the 22nd International Confer-ence on Computational Linguistics - COLING ’08, 1, 169–176.

De Felice, R., & Pulman, S. (2009). Automatic detection of preposition errors in learner writ-ing. CALICO Journal, 26(3), 512–528.

Dominguez, L., Arche, M. J., & Myles, F. (2011). Testing the predictions of the feature-as-sembly hypothesis: evidence from the L2 acquisition of Spanish aspect morphology. In BUCLD 35: Proceedings of the 35th Annual Boston University Conference on Language Development, 1, 183-196.

Islam, A., & Inkpen, D. (2010). An unsupervised approach to preposition error correction. In Proceedings of the 6th International Conference on Natural Language Processing and Knowledge Engineering, 1–4.

Islam, A., & Inkpen, D. (2011). Correcting different types of errors in texts. In C. Butz & P. Lingras (Eds.), Advances in Artificial Intelligence (pp. 192–203). Springer Berlin Heidel-berg.

Jarvis, S., & Odlin, T. (2000). Morphological type, spatial reference, and language transfer. Studies in Second Language Acquisition, 22(4), 535–556.

Lardiere, D. (2008). Feature assembly in second language acquisition. The role of formal features in second language acquisition, 106-140.

Nicholls, D. 2003. The Cambridge learner corpus: Error coding and analysis for lexicography and ELT. In Proceedings of the Corpus Linguistics 2003 conference, pages 572–581.

Odlin, T. (2013). Crosslinguistic influence in second language acquisition. In C. A. Chapelle (Ed.), The Encyclopedia of Applied Linguistics, (pp. 436-486). Oxford, UK: Blackwell Pub-lishing Ltd.

Sulkala, H. (1996). Finnish as a second language for speakers of related languages. In M. Martin & P. Muikko-Werner (Eds.), Finnish and Estonian: New target languages (pp. 143–158). Jyväskylä, Finland: Center for Applied Language Studies, Universityof Jy-väskylä.

Tomasello, M. (2001). First steps toward a usage-based theory of language acquisition. Cognitive Linguistics, 11(1–2), 61-82.

59Milou Leeijen

A Natural Language Processing based text analysis of populist rhetoric in social media text messagesN.A. (Nizar) HirzallaMSc Artificial Intelligence, Vrije Universiteit, Amsterdam

KEYWORDSNatural Language Processingtext miningcomputationally detecting populist rhetoricsocial media

ABSTRACTThe political approach of populism is rising internationally while populist parties are gaining more followers. Populist parties tend to use social media as the primary method of spreading their ideas. Similarities have been detected in how these populist parties exactly spread the message on social media from a linguistic perspective, as these messages always contain a typical rhetoric and usage of words that can be described as populist rhetoric. Using definitions and characteristics of populism and populist rhetoric as coined by language and political experts an analysis based on Natural Language Processing (NLP) text analysis methods is done. The NLP text mining based methods of topic modelling and senti-ment analysis are used for this purpose. These methods provide insights into the parties’ rhetoric, the used vocabulary, what they deem the most important topics and what entities the parties perceive as good and bad. The Yellow Vest Moment, which has been a debating point regarding if it can be defined as populist movement, is used as an example case to seek if this type of analysis can classify this movement as populist. Two established Dutch populist parties, PVV and Identitair Verzet, are used as well for comparison.

1. IntroductionAn international rise of populism has been observed in the last decade (Algan, Guriev, Papaioannou, & Passari, 2017). This is a trend that has been met by dis-approval overall due to the negative connotations associated with populism, as well as been dubbed as a significant threat worldwide by several experts (Cox, 2017; Goodwin, 2011). Noticeable among populist parties is that they tend to use social media extensively to spread the core message. As social media is easily accessible, these parties potentially have a large outreach and as such can acquire new followers and spread their messages relatively easy (Engesser, Ernst, Esser, & Büchel, 2017). However, these types of messages have been found problematic as they usually antagonize a significant portion of the overall population (Gidron & Bonikowski, 2013). This can in turn lead to a divided nation or even the collapse of the democratic foundations of a country (Cox, 2017; Gidron & Bonikowski, 2013; Goodwin, 2011).

Nizar Hirzalla


Rhetorical and linguistic patterns have been detected in how these messages are usually constructed by the populist writer. Such patterns manifest themselves, for instance, by repetitive usage of certain words in different messages or certain types of rhetoric that function repetitively as the core element of the message (Engesser et al., 2017; Jagers & Walgrave, 2007). With these patterns established, this could be used for further purposes such as detecting the rise of upcoming po-pulist parties. This can be done through a computational analysis using different kinds of Natural Language Processing (NLP) techniques, the further implications being that new parties can computationally be detected using populist rhetoric and therefore can be approached as such. To illustrate this, an analysis based on a comparison between two popular and established populist parties in the Nether-lands, PVV and Identitair Verzet (IV), and an upcoming party that gained traction recently, The Yellow Vest Movement (YVM), will be made. Geert Wilders’ PVV (’Party for Freedom’) and IV (’Identitarian Resistance’) are both well-known right-wing po-pulist parties that have a significant number of followers.

The YVM mainly combats economic inequality and has been named a populist movement but not officially yet seen as such due to it being new, making it an inte-resting example case (Hankla, 2018; Nossiter, 2018). The results of the analysis will determine if the YVM can potentially be classified as a populist movement through computational means and substantiate the populist label it received by the media.

However, the goal of this paper is not to define ’populism’ or to classify parties as populist but rather to construct a prototype for a universal method to determine whether text messages match the criteria of populist rhetoric.

2. Related workLiterature from two domains are especially relevant to this paper. The first domain constitutes studies related to defining populism and populist rhetoric. The second domain relates to technical studies in which different NLP techniques are used.

2.1 Defining populism and populist rhetoricThere does not seem to be a universal consensus on the definition of populism (Mudde, 2013). However, a definition coined by Cas Mudde in 2004, based on a set of studies, has been dubbed as an influential and iconic definition due to its expressiveness and universal usage (in other studies but also among the general public) (Gidron & Bonikowski, 2013). Mudde defines populism as follows:

"[populism is] a thin-centered ideology that considers society to be ulti-mately separated into two homogeneous and antagonistic groups, ‘the pure people’ versus ‘the corrupt elite’ and which argues that politics should be an expression of the volonté generalé (general will) of the people. "

(Mudde, 2004,p.543)

Guest paper

61Milou Leeijen

Due to its influential nature, this definition will be used for this paper. Some in-trinsic properties of populism can also be defined to get a better understanding of what typical populist rhetoric encompasses. ‘Typical’ patterns that emerge in populist rhetoric, as identified by political linguist Szilagyi (Szilagyi, 2017), are the following and are in accordance with a 2017 study (Engesser et al., 2017):

(i) Constantly referring and advocating to/for ‘the people’ (ii) Usage of personal pronouns to connect with the audience, with the intended purpose to create a ‘us versus them’ mentality (iii) Calling out and criticizing the elite (iv) Repeated use of metonymies (using alternative words to describe what is factually meant) (v) Emphasizing victimhood of the audience

These patterns do not necessarily always indicate whether a person is a populist or not. The major factor here is consistency: if these patterns form the central component of your rhetoric and are used consistently, then the person using this rhetoric can be identified as a populist (Szilagyi, 2017).

2.2 NLP techniquesNLP is a well-established domain within Artificial Intelligence and is concerned with processing and analyzing large amounts of natural language data, usually with the intent to enhance human-computer interaction or to computationally under-stand ‘human language’ (Collobert et al., 2011). Text mining is a technique often part of NLP-based analyses, as it can play an essential part in obtaining relevant information from large amounts of text, for example, through deriving patterns (such as the established patterns in populist-driven texts) (Kao, 2007).

Going more in depth, text mining can be used for many intents and purposes once a dataset or a corpus has been established. Such a purpose would be to detect how certain topics are perceived by the examined population, which is also known as a sentiment analysis. There are also NLP-techniques for summarizing purposes. Making a dictionary of a defined number of text posts with the intent to model topics to showcase the most discussed/important topics is one of such techniques (topic modelling).

For this paper, the NLP-techniques of text mining, topic modelling and a sentiment analysis will prove to be useful. These techniques can be performed to obtain re-levant data from the text source, find the most important topics and usage of words in that data and how they are perceived sentimentally (Liu, Tang, Dong, Yao, & Zhou, 2016; Maier et al., 2018; Pak & Paroubek, 2010; Westerlund, Leminen, & Rajahon-ka, 2018). Therefore, this paper will seek if the usage of these techniques combined provides a good universal method for finding populist rhetoric in text messages.

Nizar Hirzalla


3. MethodsThe steps that will be taken for text mining, topic modelling and sentiment analy-sis will be thoroughly explained in this section. In addition, the steps for obtaining the data as well as pre-processing the data is elaborated upon.

3.1 Text miningThe data used for this paper consists of posts from the official PVV and IV Facebook-pa-ges (approximately 300000 and 1000 followers). Due to YVM not having an official page, the three most active YVM Facebook pages (combined approximately 5000 followers) are instead used (‘Gele-Hesjes-NL’, ‘GeleHesjesVlaanderen’ and ‘DeGeleHesjes’). Upon visual inspection, the YVM based in Flanders (Belgium) did not have any significant dif-ferences with those based in the Netherlands and as such can be grouped as one.

For obtaining and text mining posts on Facebook, Netvizz (version 1.6)1 is used. Net-vizz is a text mining tool with multiple functions for analyzing text posts on Facebook (Rieder, 2013). In Netvizz, the 'Page posts' module is chosen as this focuses on re-trieving posts. For all three parties, 999 posts are selected for retrieval ( the maxi-mum number that can be given as input for obtaining posts). For PVV and IV the last 999 posts were obtained this way. For YVM this netted 274 posts. This lower number of posts can be attributed to the fact that YVM is a recent movement and therefore logically will have a lower number of posts in comparison to the esta blished parties. The obtained text was cleaned up and prepared for topic modelling by doing the following:

(i) All strings that solely contain a picture, video or a shared link are removed, as these cannot be interpreted for text analysis purposes. (ii) A standard Dutch and English list of stop words2 is used to filter stop words from the obtained text posts. (iii) Specific stop words, not in the standard lists, used by the different parties are manually picked out and filtered out as well. (iv) Single characters, concatenations of punctuation marks, smileys, and other insignificant fragments of text are filtered out as well. (v) Strings that are empty as a result of the previous filters, are not included.

This produced the results as shown in Table 1. Table 1 shows the number of posts and the average length of a post from the three parties before cleaning up the data and after cleaning up the data.

1 https://tools.digitalmethods.net/netvizz/facebook/netvizz/index.php2 Stop words are commonly used words that do not tell us much about a topic on their own (Rajaraman & Ullman, 2011). For example, words like: which, here, else, over and the.

Guest paper

63Milou Leeijen

Table 1 Number of text messages and the average length of a text message (in words) respec-tively before and after cleaning the different datasets.

Political party Before clean-up After clean-up

PVV 999 - 22 754 - 8

IV 999 - 26 892 - 13

Gele hesjes 274 - 58 272 - 27

3.2 Topic modellingThe process of topic modelling itself consists of multiple phases, which we will go through one by one in this section.

3.2.1 LemmatisationWith the typical NLP-task of lemmatisation, all words are derived to their base form. This makes it possible to classify, and thus interpret, different forms of the same word as one. This gives us better results, as different forms of the same word will not be seen as different words. After preparing the texts, lemmatisation is done with Frog (version 0.15)

3, which makes it possible to lemmatise large quanti-

ties of text posts quickly. (Van den Bosch, Busser, Canisius, & Daelemans, 2007)

3.2.2 Building a viable corpusAfter lemmatisation, a dictionary of the text is made and converted to a corpus based on the bag-of-words model. This model represents texts, which in the case of the used datasets for this paper ranged from a single sentence towards small monologues, as a multiset of words (an unordered collection of words where each word can appear a finite number of times). This multiset of words disregards gram-mar but instead focuses on the multiplicity and salience of the different words (Zhang, Jin, & Zhou, 2010).

3.2.3 Using latent Dirichlet allocation to model topics. With Python module Gensim (version 3.7.1)

4 using latent Dirichlet allocation (LDA), topic modelling is performed (Rehurek & Sojka, 2010). LDA is a type of sta-tistical classification that uses a generative model to classify similar data in a corpus (Blei, Ng, & Jordan, 2003). A generative model is based on the joint prob-ability distribution that seeks if different variables can be classified in fitting any particular range or a set of values specified for that variable. In the case of LDA, it looks for connections between different words and seeks to explain why some parts of the data is similar by discriminating. It operates by viewing the input data as a mixture of various topics and whether these topics can be characterized by the words in the dataset. In the context of this paper, this means that it computes

3 https://languagemachines.github.io/frog/4 https://pypi.org/project/gensim/

Nizar Hirzalla


probabilities for topics that come forward prominently based on how certain words seem to relate to each other, which can then be linked to populist rhetoric. The purpose of this is to find the most important talking points of the party that pub-lished the texts.

To illustrate how LDA would work in practice, an example will be given. For this example the topic threshold will be set at 5. The case for this example will be a political party that frequently talks about immigration, climate change, leaving the EU, decreasing taxes and education on social media. LDA’s generative model sta-tistically classifies that terms related to these topics are consistently used in the same context and same messages, indicating a cluster.

The more the same terms occur in the same context in different text posts about the topics, the stronger the cluster will be and the stronger the probability will be of the most occurring terms in that cluster. Eventually this will produce an output of the mentioned five topics with the most occurring terms that define these topics.

3.2.4 Computing and visualizing the most relevant and salient termsUsing Python module pyLDAvis (version 2.1.2) (Mabey, 2018), the most relevant words and most salient terms within the topics are computed. pyLDAvis extracts information from a fitted LDA topic to create an interactive web-based visualization for easy interpretation of the topic modelling results. For calculating the relevancy of words within topics and saliency, in combination with the LDA results, the following formulas from a 2012 study (Chuang, Manning, & Heer, 2012) and a 2014 study (Sievert & Shirley, 2014), respectively, are used. These formulas are the following:

(i) saliency(termw) = frequency(w) * [sumtp(t|w) * log(p(t|w)/p(t))]

(Chuang et al., 2012)

(ii) relevance(termw|topict) = λ * p(w|t) + (1 − λ) * p(w|t)/p(w)

(Sievert & Shirley, 2014)

In these formulas w and t respectively stand for 'word' and 'topic'. Saliency is a mea-sure for indicating how noticeable and important a term is, and in this case indicates how much it tells about the topic. Relevancy in this case is a weighted average of the probability of the word in the context of the given topic and the pro bability of the word in the context of the given topic normalized by the probability of the topic.

Furthermore, pyLDAvis shows how different topics from the LDA results are linked and as such whether there are similarities between the topics. The Python script for the procedures of (1) pre-processing text, (2) LDA topic modelling, and (3) visuali-zing the models has been published on Github

5.

5 https://github.com/NizarH/PopulismTopicModelling

Guest paper

65Milou Leeijen

3.3 Sentiment analysisWith sentiment estimating tool SentiStrength (version 2.3)6, the sentiment analysis is performed. SentiStrength rates input data, in the form of sentences or small texts, on a scale from -4 (negative) to 4 (positive) using pre-defined data as a refe rence point (Thel-wall, Buckley, Paltoglou, Cai, & Kappas, 2010), in this case a Dutch dataset of words7. The main list of words used for reference is a standard list of all Dutch words correlated with positive and negative sentiment. To get more accurate results, other lists are also included: ‘booster words’, negating words, slang words, question words, English words that get used often in Dutch texts, emoticons and derision terms. Booster words consist out of words that amplify a certain word, for instance the word ‘very’. For example, in the case of ‘that is very good’ the positive score of good increases. Negating words do the opposite; if good, is preluded with ‘not’, the positive score becomes 0. Derision terms reduce negative scores slightly by looking at certain irony-associated words like haha or lol. With this reference data SentiStrength is applied on all of the cleaned datasets.

4. ResultsFor topic modelling, a threshold of five topics is chosen as this gives the best results. Fewer topics would cluster some terms too broadly, while more topics would return terms with too loose of a connection to the topic

4.1 Topic Modelling ResultsThe computed topics seem to range from promoting voting for the PVV in elections (topic 1, 5) to reiterating typical PVV talking points (topic 3, 4) (see Appendix A8). How these topics are related to each other can be seen in Figure 1.

The most salient terms show that there are some terms with a relatively high fre-quency: ‘Nederland’, ‘PVV’ and ‘StemPVV’ (translates to: ‘vote for PVV’) . The most salient terms seem to overlap with the most relevant terms in the five classified topics (see Appendix B).

Figure 1 shows no overlap between topics. Topics 1, 2, 3 and 5 are relatively close to each other (in the same hemisphere) while topic 4 is on the complete other side, indicating a more distinct topic in comparison to the other topics.

For IV, the topic modelling results show many similarities with PVV. There is no overlap between the IV topics and the spread of the topics is the same as with the PVV topics: topic 1, 2, 3, 5 are in the same hemisphere and close to each other while topic 4 is far-off, as can be seen in Figure 2.

6 http://sentistrength.wlv.ac.uk/index.htmlDownload7 https://github.com/NizarH/PopulismTopicModelling/tree/master/SentiStrengthDutch/dutch8 Find the online version of this article, including appendices, at linguujournal.nl.

Nizar Hirzalla


Figure 1 . Intertopic distance map. A representation of the relations and connectivity between the different LDA produced topics of the PVV dataset.

Guest paper

67Milou Leeijen

Figure 2 . A representation of the relations and connectivity between the different LDA produced topics of the IV dataset.

As IV is an activist group, unlike political party PVV, there are no topics asking for votes. The topics all indicate to be centered thematically around political talking points. Terms such as ‘activist’ seem to indicate that some of these topics may also be centered around promoting activism (see Appendix A). For IV, the salient terms seem to have a higher frequency than the PVV salient terms, which can be explained due to a broader range of talking points (see Appendix B). There are some terms that stand out due to their very significant presence, such as 'Ned-erland', 'land', 'nieuw', 'groot' and 'verzet'. These terms translate, respectively, to 'the Netherlands', 'country', 'new', 'big' and 'resistance'. For the YVM, thematically, the topics all seem to promote some form of activism, but are also addres sing the readers and reiterating their role as 'the people' (see Appendix A). The relations between these topics can be seen in Figure 3.

Nizar Hirzalla


Figure 3 . A representation of the relations and connectivity between the different LDA produced topics of the YVM dataset.

Figure 3 shows an overlap between topic 3 and 5. The structure of having four top-ics in the same hemisphere (relatively close to each other) with a distinct topic, as was also observed with the other two parties, seems to be returning with the YVM, this time topic 1 being the far-off, distinct topic. We can also see that most of the salient topics have a relatively low frequency on average, but there are also very noticeable high frequency salient terms such as 'mens', 'volk', 'geld', 'bank', 'alleen' and 'denken' (see Appendix B). These terms translate, respectively, to 'per-son', 'people', 'money', 'bank', 'alone', 'to think'.

4.2 Sentiment AnalysisFor the sentiment analysis, terms that had the most positive scores and the terms with the most negative scores were considered and analysed. These were terms with a score ranging from 3 to 4 and -3 to -4, respectively. Scores from -1 to 1 were

Guest paper

69Milou Leeijen

considered neutral and scores -2 or 2 were considered too marginal to give a good indication of how different terms are perceived. The strongly rated posts were also selected because they are the best interpretable and therefore give us the most insight. The results of the sentiment analysis for all three parties can be seen in Table 2.

Table 2 shows some similarities between the parties and how they perceive diffe-rent topics. Terms such as 'Rutte', 'de overheid' ('the government'), 'kabinet' ('cabi-net') are being perceived as a negative associated terms, for all three parties. Also when looking at positive associated terms there seems to be overlap with terms such as 'volk', 'wij', 'ons' (translates to 'the people', 'we' and 'us'). The PVV and IV seem to practically have the same perceptions thematically and content-wise, considering they share many similarities in both positive and negative associated terms.

5. DiscussionThe results of topic modelling provide insight into the rhetoric used by the parties as well as their most discussed topics. These range from promoting activism/voting towards talking about topics that concern the parties. It becomes clear that social media is used for multiple purposes: (1) to promote (physical) activity in support of the party or the party’s ideology, (2) to inform the readers of topics that concern the party, and (3) to form a narrative in which an (emotional) reaction tends to be instigated from the readers.

Some of the produced topics were clear in what they exactly constitute, e.g. PVV topic 1, 4, IV topic 1, 5, and YVM topic 1 and 4. For example, a topic referencing to the ’good people’ and their supposed victimhood can be found in topic 4 of the YVM topics. Sometimes it was not clear what the topic exactly constitutes because the clustering of terms is too broadly, e.g. IV topic 2, 4 and YVM topic 3, 5. The intertopic distance charts showed us that the topics do in fact have much variance indicating a multitude of concerns for the parties. There was hardly any overlap with the different topics, except for one case in the YVM dataset. Some topics were clustered far away from others (topic 4 for PVV and IV, topic 1 of YVM), this is most likely caused due to the fact that these topics include combination of terms that differ significantly from the other four topics per the Intertopic Distance metric.

The sentiment analysis gained insights in populist rhetoric and the diffe rent par-ties’ perceptions on different topics. It helped providing some needed context for the topic modelling results. The sentiment analysis showcases a perceived ‘bad’ common foe for all three parties due to the sheer amount it is talked about and forms the backbone of the parties‘ rhetoric. This is also illustrated by a clear di-chotomy in what is consistently perceived as negative versus what is consistently perceived as positive.

Nizar Hirzalla


Table 2 The results of the sentiment analysis for the (1) PVV, (2) IV, and (3) The YVM (Gele Hesjes), showcasing the most positive and negative associated terms

Party The most positive associated terms (at least a score of 3 or more)

The most negative associated terms (at least a score of -3 or less)

PVV Nederland, Nederlandse cultuur, Burgers or Volk, Ons or Wij, PVV (PVV members but also people who vote PVV) Also a combination of people, groups or parties who speak out against the negative associated terms

Rutte, VVD, Kabinet + Parlement, EU, Islam (and also associations with Islamic beliefs, figures, culture or countries), Terreuraanslag/Terreur in combination with Islam, Geweld or Misdrijf in combination with Islam/muslims, morroccans, or migrants, Gelukszoekers or Asielzoekers (usually to refer to refugees or migrants)

Identitair Verzet

Nederland, Nederlanders, Ons, Wij or Onze x where x pertains to Dutch people, institutions, culture or achievements, people associated with Identitair Verzet (members, activists, protesters, supporters), Afrikaners

Rutte, VVD, Dutch left-wing parties, Extreemlinks, EU, Europa, Islam (and also associations with Islamic beliefs, figures, culture or countries), Salafisten or Jihadisten, Cultuurmarxisme, Asielzoekers or Illegalen, Migranten

Gele Hesjes

Volk, Wij, Gele Hesjes, De normale burger or mensen, De arme burger (and terms showcasing/emphasizing victimhood of 'the people'), terms associated with standing up for the 'victimized people', terms associated with improving the climate or milieu with collective behaviour change, terms associated with protesting or rebelling against the rulers or elite, terms associated with economic equality

Rutte, De overheid, Banken, Verdeeldheid or terms pertaining to economic inequality, System, Europa, Belasting, terms pertaining to making profit by 'the elite' at the expense of 'the people', Klimaatmaatregelen or terms pertaining to combat climat change with increasing ecological costs for 'the people'

Guest paper

71Milou Leeijen

For PVV and IV the negative perceived entities can be defined into two groups: The elite or ‘the corrupt rulers’ (VVD, Rutte, the government, EU) on one hand and an existential threat on the other hand (Islam, migrants). For the YVM, the same division can be found, with the former being Rutte, the government and Europe and the latter banks and economic inequality. They see these threats operating in real time and have a sense of being collectively oppressed and economically exploited by both the banks as well as the elite (Marijnissen, 2018).

The PVV and IV parties seem to be more aligned with nationalist and right-wing senti-ments, while the YVM aligns more with socialist and left-wing sentiments (Otjes & Louwerse, 2013; van Elsas, Hakhverdian, & van der Brug, 2016). Content and main concerns of the parties aside, it becomes clear that all parties have the earlier defined ‘us versus them’ dichotomy forming their rhetoric. For instance, all parties constantly refer to ’the people’ or other synonyms with the same effect. Additionally, emphasi-zing victimhood of ‘the people’ and the parties themselves happens constantly, most apparent in the YVM dataset. Often, personal pronouns such as ’we’ or ’us’ are used in order to connect with the audience. Furthermore, the usage of metonymies, such as PVV referring to all migrants and refugees as ‘fortune seekers’ is also a recurring theme. Lastly, calling out and criticising the elite seems to be the most used rhetoric as this is constantly done and overall seems to form the biggest priority for the parties. These outcomes combined seem to give a full match with the defined characteristics of populist rhetoric in the Related Work section, for all three parties.

5.1 Limitations of this paper and the used methodsWhile performing the different NLP-analyses, some issues and limitations became ap-parent. The first one being that the YVM dataset is significantly smaller than the PVV and IV dataset, possibly yielding less representative results. This was partially dealt with by using multiple Facebook pages to gain a larger YVM dataset.

From the topic modelling results, it was not always clear how to interpret the topics. Possibly, using larger datasets for all parties could cluster some terms with more sig-nificance which then could be interpreted more easily. Doing a sentiment analysis did in turn provide some needed context for the topic modelling results.

As for the sentiment analysis, there is a limitation in how the text messages are interpreted. The perception of what can be regarded as negative or positive has a subjective connotation. For instance, people with different characters can per-ceive the negativity or positively of a certain text messages differently, and as such the used simple linear scoring might not reflect reality adequately.

Additionally, SentiStrength has a hard time detecting sarcasm, potentially mak-ing some of the produced scores inaccurate. The mentioned ’derision terms’ list of words as referenced in Section 3.3 can act as a simplified workaround for this, but this in itself poses problems. Some of these words might instead not be used for the purpose of irony, but then still get counted as such this is hardcoded into

Nizar Hirzalla


SentiStrength. Properly detecting sarcasm or irony cannot be done at this moment, as detecting sarcasm is still a topic of research with difficulties on its own in NLP-research (González-Ibáñez, Muresan, & Wacholder, 2011). To somewhat counteract this, a randomly selected group of posts from all parties were manually inspected for sarcasm, although this way sarcasm was only sporadically detected. Further-more, SentiStrength has difficulties seeing ‘the bigger picture’, for instance when a text post consists of multiple sentences with half of these sentences being positive and the other half being negative. This would net an overall neutral score, but this does not represent the content of the text post in a useful way for the purpose of this paper. The latter issue was dealt with by manually checking some of the longer text posts that received a neutral rating.

Furthermore, SentiStrength cannot detect specific jargon or ‘invented’ words of the populist parties, such as ‘kansenparel’ as used by IV, which is used to describe non-western migrants or refugees in a derogatory way. These types of words have to be added manually in the reference data, which is less efficient.

6. Conclusion and future workMultiple NLP techniques have provided new insights in the YVM’s use of rhetoric.This paper has shown that YVM adheres to the definition coined by Cas Mudde, engulfing what populism exactly constitutes (Mudde, 2004). It becomes clear from the sentiment analysis in combination with the topic modelling results that the 'us vs them' mentality where 'us' is defined as pure, good, victimized people and 'them' the corrupt, bad elite and/or an existential threat is very present in YVM’s rhetoric. This dichotomy becomes especially clear when looking at Table 2, which illustrates the clear division between the positively perceived ‘us’ and related concepts and the negatively perceived ‘them’ and related concepts. Next to this, it also became clear that the YVM linguistic style of writing and rhetoric also consistently adheres to all five of the populist characteristics, as defined by language and political ex-perts (Engesser et al., 2017; Szilagyi, 2017).

The same, on both fronts, was also true for the established populist parties, PVV and IV, as was expected. In comparison with these parties the YVM tends to be the same regarding using populist rhetoric and linguistic writing style, although there was difference found in the rhetoric content-wise. The YVM leans much more to left wing populism and socialism while PVV and IV adhere more towards right wing populism and nationalism.

In the greater context, this paper has also provided a prototype for a method to sufficiently find if new groups and/or parties can be computationally classified as populist. In our experiments, we created a corpus and applied automatic text analy-sis methods in the form of topic modelling and sentiment analysis to obtain insights into the linguistic characteristics of populist rhetoric.

Guest paper

73Milou Leeijen

For the future, this method could be improved by including a technique that accu-rately can detect sarcasm in text consistently. Sarcasm can twist the meaning of a text post significantly and as such should be considered when doing a sentiment analysis. Therefore, this paper recommends further research in the field of NLP regarding detecting sarcasm. Also, for future research this paper re commends to apply this method to more and bigger datasets. This can be applied to other up-coming parties to see if it can consistently classify parties as populist accurately. To add on to that, for future research this paper also recommends to seek if the method in this paper can also be accurately used for other languages (non-Dutch), other types of populist parties and for parties that do not post on social media but rather use more traditional ways of spreading the message. Finally, this paper also recommends to consider non-populist parties and to seek whether the method can confirm them as such.

AcknowledgementI want to thank my supervisors Marieke van Erp and Victor de Boer for providing the means to carry out the research, their assistance and supervision. I also want to thank political experts Ernst van den Hemel and Sakina Loukili for their advice and feedback. ■

Find the online version of this article, including appendices, at linguujournal.nl.

Received April 2019; accepted August 2019.

ReferencesAlgan, Y., Guriev, S., Papaioannou, E., & Passari, E. (2017, November). The European Trust

Crisis and the Rise of Populism (tech. rep. No. 12444). C.E.P.R. Discussion Papers. Re-trieved from https://ideas.repec.org/p/cpr/ceprdp/12444.html

Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet Allocation. J. Mach. Learn. Res. 3, 993–1022. Retrieved from http://dl.acm.org/citation.cfm?id=944919.944937

Chuang, J., Manning, C. D., & Heer, J. (2012). Termite: Visualization Techniques for Asses-sing Textual Topic Models. In Advanced visual interfaces. Retrieved from http://vis.stan-ford.edu/papers/termite

Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., & Kuksa, P. (2011). Natu-ral language processing (almost) from scratch. Journal of machine learning research, 12(Aug), 2493–2537.

Cox, M. (2017, January). The Rise of Populism and the Crisis of Globalisation: Brexit, Trump and Beyond.Engesser, S., Ernst, N., Esser, F., & Büchel, F. (2017, August). Populism and social media: how politi-

cians spread a fragmented ideology. Information, Communication & Society, 20(8), 1109–1126.Gidron, N. & Bonikowski, B. (2013). Varieties of Populism: Literature Review and Research

Agenda. Weatherhead Working Paper Series, No. 13-0004. Weatherhead Working Pa-per Series, No. 13-0004.

González-Ibáñez, R. I., Muresan, S., & Wacholder, N. (2011). Identifying Sarcasm in Twitter: A Closer Look. In Acl.

Goodwin, M. J. (2011, September). Right response: Understanding and countering populist extremism. Chatham House. London: The Royal Institute of International Af-fairs, Chatham House. Retrieved from https://kar.kent.ac.uk/54401/

Hankla, C. (2018, December). What French populists from the ‘50s can teach us about the ‘yellow vests’ roiling Paris today. Retrieved from http://archive.is/jBikW

Nizar Hirzalla


Jagers, J. & Walgrave, S. (2007, May). Populism as political communication style: An empirical study of political parties’ discourse in Belgium. European Journal of Political Research, 46(3), 319–345.

Kao, S., A; Poteet. (2007). Natural Language Processing and Text Mining. SpringerInternational Publishing.Liu, L., Tang, L., Dong, W., Yao, S., & Zhou, W. (2016, September). An overview of topic mo-

deling and its current applications in bioinformatics. SpringerPlus, 5(1), 1608. Mabey, B. (2018). pyLDAvis 2.1.2. Retrieved from https://pyldavis.readthedocs.io/en/la test/Maier, D., Waldherr, A., Miltner, P., Wiedemann, G., Niekler, A., Keinert, A., Adam, S. (2018).

Applying LDA Topic Modeling in Communication Research: Toward a Valid and Reliable Methodology. Communication Methods and Measures, 12(2-3), 93–118.

Marijnissen, H. (2018, December). De gele hesjes zijn het zat! Maar wat? Retrieved from https://www.trouw.nl/home/de-gele-hesjes-zijn-het-zat-maar-wat-~a184703a/

Mudde, C. (2004, September). The Populist Zeitgeist. Government and Opposition, 39(4), 541–563. Mudde, C. (2013). Are populists friends or foes of constitutionalism? The Foundation for Law,

Justice and Society.Nossiter, A. (2018, December). How France‘s ‘Yellow Vests’ Differ From Populist Move-

ments Elsewhere. Paris. Retrieved from https://www.nytimes.com/2018/12/05/world/europe/yellow-vests-france.html

Otjes, S. & Louwerse, T. (2013, November). Populists in Parliament: ComparingLeft-Wing and Right-Wing Populism in the Netherlands. Political Studies, 63(1), 60–79.Pak, A. & Paroubek, P. (2010). Twitter as a corpus for sentiment analysis and opinion min-

ing. In Lrec (Vol. 10, 2010, pp. 1320–1326).Rajaraman, A. & Ullman, J. D. (2011). Data Mining. In Mining of massive datasets (pp.

1–17). Cambridge University Press. Rehurek, R. & Sojka, P. (2010). Software Framework for Topic Modelling with Large Corpora.

In Proceedings of the lrec 2010 workshop on new challenges for nlp frameworks (pp. 45–50). Valletta, Malta: ELRA.

Rieder, B. (2013). Studying Facebook via Data Extraction: The Netvizz Application. In Pro-ceedings of the 5th annual acm web science conference (pp. 346–355). WebSci ’13. New York, NY, USA: ACM.

Sievert, C. & Shirley, K. (2014). LDAvis: A method for visualizing and interpreting topics. In Proceedings of the workshop on interactive language learning, visualization, and inter-faces (pp. 63–70). Association for Computational Linguistics.

Szilagyi, A. (2017). How to talk like a populist. VICE News. Retrieved from https://news.vice.com/en_us/article/gyd98y/how-to-talk-like-a-populist

Thelwall, M., Buckley, K., Paltoglou, G., Cai, D., & Kappas, A. (2010, December). Sentiment Strength Detection in Short Informal Text. J. Am. Soc. Inf. Sci. Technol. 61(12), 2544–2558.

van den Bosch, A., Busser, B., Canisius, S., & Daelemans, W. (2007, January). An efficient memory-based morphosyntactic tagger and parser for Dutch.

van Elsas, E. J., Hakhverdian, A., & van der Brug, W. (2016, November). United against a common foe? The nature and origins of Euroscepticism among left-wing and right-wing citizens. West European Politics, 39(6), 1181–1204.

Westerlund, M., Leminen, S., & Rajahonka, M. (2018). A Topic Modelling Analysis of Living Labs Research. Technology Innovation Management Review, 8, 40–51.

Zhang, Y., Jin, R., & Zhou, Z.-H. (2010, December). Understanding bag-of-words model: A statistical framework.

Guest paper

75Janna van Egmond

Taal en psychiatrie Een stage in combinatie met schrijven van een bachelorscriptie

An internship in combination with writing a Bachelor's thesis

J. (Janna) van EgmondStage gelopen tijdens haar BA Taalwetenschap, Universiteit Utrecht, Utrecht aan het UMC Utrecht MA Taal- en Spraakpathologie, Radboud Universiteit, Nijmegen

Janna interned at the UMC where she worked on data collection for a research on speech in various psychiatric disorders. During this time she also worked on her own Bachelor's thesis about language in schizophrenic patients.

“Ik zie het als heel waardevol dat ik een keer van dichtbij heb gezien hoe wetenschappelijk onderzoek gaat en alle stappen ook echt zelf heb uitgevoerd.”

AanleidingIn het studiejaar 2018-2019 heb ik stagegelopen aan het Universitair Medisch Centrum in Utrecht. Dit was in het vierde jaar van de bachelor Taalwetenschappen aan de UU. Ik heb mijn stage gecombineerd met het schrijven van mijn scriptie. Het lopen van een stage is bij de bachelor niet verplicht, maar toch had ik het gevoel dat dit een aanvulling zou zijn op het reguliere programma van de bachelor en had ik door een half jaar uitloop hier ook de ruimte voor. Bij het vak Klinische Linguïstiek van Frank Wijnen en Tessel Boerma kwam het onderwerp 'taal en psy-chiatrie' kort voorbij en dit trok gelijk mijn aandacht. Via een studiegenootje ben ik in contact gekomen met een onderzoeksgroep in het UMC die daar onderzoek naar doet.

Het onderzoekHet onderzoek, genaamd PRAAT, richt zich op taal en spraak bij verschillende psychiatrische aandoeningen, waaronder schizofrenie, depressies en de ziekte van Parkinson. Op dit moment is het stellen van de diagnose bij sommige psy-chiatrische aandoeningen gebaseerd op relatief subjectieve waarnemingen. Het gebruik van biologische markers, in dit geval taal en spraak, zou dit kun-

76 LingUU | 3.2 | 2019 Intern Extern

nen veranderen. Het onderzoek van PRAAT is gebaseerd op de aanname dat verschillende aspecten van taalbegrip en taalproductie bij mensen met - onder andere - schizofrenie is aangedaan. Door te analyseren welke aspecten er precies zijn aangedaan, kan er wellicht eerder een psychose gesignaleerd worden. Bij het PRAAT-onderzoek wordt er door interviews met patiënten en controleparticipanten af te nemen zoveel mogelijk spontane taaldata verzameld.

Mijn eigen onderzoek ging over coherentierelaties tussen zinnen of zinsdelen uit-gedrukt door voegwoorden bij schizofreniepatiënten, en gezonde controles. Het is onduidelijk of de taalproblemen bij schizofreniepatiënten komen door een gebrek-kig taalsysteem, of een gebrek in cognitieve processen die nodig zijn voor taal. Uit eerder onderzoek is gebleken dat bepaalde coherentierelaties minder complex zijn dan andere. Ook is uit eerder onderzoek gebleken dat cognitie is aangedaan bij schizofreniepatiënten. In mijn onderzoek heb ik gekeken of het gebruik van coherentierelaties verschilt tussen schizofreniepatiënten en gezonde con-troles. Ik verwachtte dat schizofreniepatiënten meer gebruik zouden maken van minder complexe coherentierelaties uitgedrukt door bijbehorende voegwoorden in vergelijking met gezonde controles, te wijten aan een verschil in cognitie. Conclu-derend heb ik kunnen zeggen dat binnen mijn sample een verschil bestaat in het gebruik van coherentierelaties tussen de twee groepen, waarbij sommige minder complexe coherentierelaties meer werden gebruikt door de schizofreniepatiënten.

De werkzaamheden voor het lopende onderzoekTijdens mijn stage hielp ik mee met onder andere de dataverzameling van het lopende onderzoek. Daarnaast deed ik met al bestaande data mijn eigen onder-zoek voor mijn scriptie. Een groot voordeel van het schrijven van mijn scriptie in combinatie met een stage vond ik dat ik een dataset kon gebruiken waar ik op een andere manier nooit bij was gekomen, en ik zo onderzoek kon doen naar een kli-nische populatie die ik erg interessant vind. Door mee te helpen aan het lopende onderzoek heb ik ook geleerd hoe je interviews moet afnemen bij klinische popu-laties, hoe je om moet gaan met onderzoeksdata, en hoe je deze moet verwerken. Doordat het PRAAT-onderzoek onderdeel was van een grotere onderzoeksgroep waren er ook wekelijkse, interessante meetings waar ik aan kon deelnemen en ook veel van geleerd heb.

Voordelen en eventuele struikelblokkenOf ik door wil in het onderzoek weet ik nog niet, maar ik zie het als heel waardevol dat ik een keer van dichtbij heb gezien hoe wetenschappelijk onderzoek gaat en dat ik alle stappen ook echt zelf heb uitgevoerd. Kortom, het is zeker aan te raden om je bachelorscriptie te schrijven in combinatie met een stage! Omdat het lopen van een stage, zeker in combinatie met het schrijven van je scriptie, niet gewoon-lijk is bij de bachelor, moet er wel een hoop geregeld worden en moet je je er wel bewust van zijn dat je zelf achter dingen aan moet. ■

77Roos Kerrebijn

Een desillusie voor de ouder, maar een ode aan het kindR. (Roos) KerrebijnBA Nederlandse Taal en Cultuur & BA Taalwetenschap, Utrecht University, Utrecht

The Infinite GiftCharles Yang

ISBN 9781451612998

Door verschillende hoogleraren werd mij The Infinite Gift van Charles Yang aangeraden. Ik – mijn taalkundig hart sneller kloppend – deelde mijn boekentip met een vakgenoot. Zij bleek sceptisch: een taalwetenschapper die is afgestu-deerd in de computerwetenschappen en ook nog banden heeft met de psycholo-gie, was dat niet gedoemd te mislukken? Het verleden in ogenschouw nemend, kon ik haar geen ongelijk geven. De pogingen om computers en taal te verenigen, hadden tenslotte weinig opgeleverd. Behalve een belangrijke les: alleen de mens bleek in staat de code van het ingenieuze taalsysteem te kraken.

De eerlijkheid gebiedt mij te zeggen dat ik doorgaans ook niet warmloop voor de psychologische visie op taalverwerving. Ik moet altijd op mijn tong bijten om niet aan ‘iedereen die vindt dat hij ook wat over taal mag zeggen’ te verkondigen wat de zienswijze van taalkundigen is. Zoals Peter Coopmans eens tijdens college zei: “Laat de studie van taalverwerving niet zomaar aan de psychologen over. Daar heb je toch echt ook taalkundigen voor nodig.”

Desondanks bleef het wetenschappelijke werk van Charles Yang mij ook niet on-opgemerkt. Zo schreef hij het interessante artikel Universal Grammar, statistics or both? (2004) en het boek The Price of Linguistic Productivity (2016). Toen ik na wat online research stuitte op de breed gedragen Chomskyaanse basis van zijn werk, was ik hoopvol gestemd.

78 LingUU | 3.2 | 2019 Book review

Taalontwikkeling: de basisaannames Het wordt meteen duidelijk: Charles Yang heeft veel in huis. Op pakkende wijze vertelt hij over het logische probleem van taalverwerving, de armoede van de stimulus en over hoe knap het is dat het kind zo’n systeem onder de knie krijgt. Het is deze intellectuele megaprestatie die we als te vanzelfsprekend beschou-wen, waar elk kind een beloning voor zou moeten krijgen. Yang beschrijft hoe abstract en complex ons taalsysteem in wezen is en het moordende tempo waarin het kind dit leert – nog voordat het zijn veters kan strikken! Bovendien zijn kin-deren innovators: zij bezitten allemaal een creatief vermogen als het taal betreft. De essentie van het taalverwervingsproces wordt in dit eerste hoofdstuk op aan-gename wijze gepresenteerd, zowel voor de bekende als onbekende in het vak-gebied. De indeling van de rest van het boek is als volgt: het begint met fonologie en gaat vervolgens naar woordenschatopbouw en via semantiek naar syntaxis. Geen basisconcept blijft onbesproken. De lezer wordt meegenomen in prenatale akoestiek, coarticulatie, het Gavagai-probleem en structuurafhankelijkheid.

Voor de onwetende ouderIndien je als ouder onbekend bent met taalverwervingstheorieën, dan kunnen sommige boodschappen uit het boek rauw op je dak vallen. Denk je dat je kind met het eerste 'papa' of 'mama' jou bedoelt? Dan heb je het mooi mis. Ben je in de veronderstelling dat de negatieve evidentie die je het kind geeft – ‘liep’ in plaats van ‘loopte’ – nut heeft? Helaas, dat blijkt een zinloze escapade. Een schrale troost geeft Yang trouwens wel: taalwetenschappers doen het ook. Niet omdat het zinvol is, nee, maar omdat zij zich dan betere ouders voelen. Heb je als ouder graag de touwtjes in handen? Misschien kun je die wat laten vieren. Op het vlak van het vroege vocabulaire gebruikt Yang de metafoor van kiezels op een strand: een nomen hier, een verbum daar. Je zult verrast worden. Voor de niet-taalkundige onderlegde ouder kan het kortom op een lichte teleurstel-ling uitlopen. Je bent als ouder wellicht een tikkie minder invloedrijk in het proces als geheel dan je vooraf had gedacht. Het eindproduct is meer een verdienste van je kind, dan van jou als ouder.

Voor de breed geïnteresseerdeVoor de breed geïnteresseerde is The Infinite Gift perfect leesvoer. Yang haalt filosofische stromingen, Bijbelverhalen, Engelse literatuur en meer aan: van Shakespeare-Engels tot Alice in Wonderland en van Cookiemonster tot Bob Dylan. De geweldige voorbeelden onderscheidt The Infinite Gift van andere boeken over taalverwerving. Bovendien beschrijft Yang alle klassieke experimenten die je zou moeten kennen, zoals het Ba/Pa-experiment (van het McGurk-effect) en de High Amplitude Sucking-methode. Elk deelgebied van de taalwetenschap wordt met verve beschreven, ook al staan sommige gebieden verder van Yang af als weten-schapper. Desalniettemin is het boek doorklonken van de nature-visie op taal. Dit wordt mede duidelijk doordat Yang de bètagebieden niet schuwt. Sterker nog: een groot deel van zijn boek bouwt voort op Darwins werk en de link met taal is snel gelegd. Yang roept biologen en wiskundigen op om zich, net als taalwetenschap-pers, over het vakgebied te buigen.

79Roos Kerrebijn

Voor de avonturier Behalve door de tijd, reis je met Yang mee over de wereld: een echte taalsafari. Van elk continent komen talen aan bod, zoals Warlpiri uit Australië, Swahili uit Afrika en Navajo uit Amerika. Ook het Eskimo passeert de revue, waarbij de woordens-chatmythe wordt ontkracht. Volgens deze mythe hebben de Inuit een groot aantal woorden voor sneeuw. Dat grote aantal blijkt mee te vallen: zij hebben in werkeli-jkheid twee woorden voor sneeuw, precies zoveel als het Engels. Hoe die ruis dan is ontstaan? Yang beschrijft dat het Eskimo tot de polysynthetische talen behoort. In een polysynthetische taal als het Eskimo, gedragen categorieën als preposities zich als pre- en suffixen in het Engels. Dit resulteert in een sequentie, die foutief is geclassificeerd als woord. En een mythe is snel verspreid. Dat ook talen zich verspreiden, wordt niet uitgebreid besproken. Wel benadrukt Yang de noodzaak om te focussen op talen die onderbestudeerd en -gedocumenteerd zijn. De reden hiervoor is simpel: taalwetenschappers komen zelf nu eenmaal vaak uit Europa of Azië, waardoor er nog een berg aan onderzoek te verrichten valt. Er is dus altijd een goede reden voor een wereldreis. Mocht dit niet lukken, wees niet getreurd: door The Infinite Gift pik je binnen mum van tijd wat mee van ieder continent.

Een stilistisch hoogstandjeHet is de rijkdom aan taalgebruik in combinatie met de veelzijdigheid van voor-beelden, die het boek tot een stilistisch hoogstandje maken. Yang toont zich een meester op vele vlakken. Het zijn alleen al de prachtige metaforen in dit boek die je zullen bijblijven. En hoe handig is het, om een beeld paraat te hebben, als je vrienden vertelt over je studie? Dat je kunt zeggen: “het fysieke mechanisme van spreken, verschilt niet zo veel van spelen van saxofoon”. Of: “pre- en suffixen zijn parasieten; ze hebben iets anders nodig om te bestaan”. Hoe fijn zou dat zijn, als je weer eens moet komen opdraven op een familiefeestje. Zulke metaforen doen het vakgebied eer aan, in plaats van de eindeloze discussies over spelling. Mooi is ook de parallel die Yang laat zien tussen kinderen en volwassenen. Zoals de verschijnselen pronomen drop en complexificatie voorkomen in kindertaal, zo ook in volwassentaal. Zeg nu eens, spreek jij koala uit als “koala”, of voeg je eigenlijk een w toe (“kowala”)?

Voor leek of expert?Het enige wat ik me tijdens het lezen van The Infinite Gift heb afgevraagd, is wie de doelgroep precies is. Voor experts zal er inhoudelijk weinig vernieuwends in staan. Dit neemt niet weg dat het een ‘leuk leesboek’ is, bijvoorbeeld voor ’s avonds op de bank. Voor generatief taalkundigen zal het smullen zijn – en bovendien herkenbaar. Volgens Yang is dit specifiek soort taalkundigen geobsedeerd door zinnen die om principiële redenen ongrammaticaal zijn. Zodoende is de frequent gebruikte zin tijdens een conferentie: “Geweldige lezing! Maar in het Baskisch/Warlpiri/Hindi/Swahili…”. De expert op de andere kant van het spectrum, zoals de sociolinguïst, zal niet staan te juichen om Universal Grammar als one-size-fits-all. Ook betwijfel ik of de doorsnee leek op boomstructuren zit te wachten, al dan niet basaal. "Ja, natuurlijk", zou ik bijna vanuit mijn voorliefde voor syntaxis roepen.

80 LingUU | 3.2 | 2019 Book review

Maar de ervaring leert dat de euforie over structurele componenten niet iedereen gegeven is. Wel legt Yang elk concept stapsgewijs en toegankelijk uit. Dat zal de geïnteresseerde leek zeker waarderen.

Een geschenk en een uitnodigingHet boek betaalt de prijs voor zijn veelzijdigheid: een licht nadeel voor elke groep. Toch raad ik het boek aan alle lezers aan, vanwege de geweldige schrijfstijl en Yangs mateloze kennis over zijn vakgebied én omringende vakgebieden. The In-finite Gift is een krachtige oproep aan taalwetenschappers, wiskundigen en bi-ologen om de cirkel rond te maken. Daarnaast bevat het een boodschap aan zijn lezer: geef het kind – en niet de ouder – de credits voor de knappe prestatie die het levert. The Infinite Gift is hiermee misschien een teleurstelling voor de ouder, maar een ode aan elk kind. ■

Charles Yang (1973) is afgestudeerd in computerwetenschappen aan de MIT. Momenteel is hij verbonden aan de Universiteit van Pennsylvania, waar hij taal- en computerwetenschap en psychologie doceert. Yangs werk wordt uitgevoerd binnen het kader van Chomsky’s theorieën. Naast wetenschappelijke artikelen heeft hij verschillende boeken op zijn naam staan, waaronder The Infinite Gift (2006). Zijn meest recente boek is The Price of Linguistic Productivity (2016).

ReferencesYang, C. (2004). Universal Grammar, statistics or both? Elsevier, 8(10), pp. 451-456.Yang, C. (2006). The Infinite Gift: How Children Learn and Unlearn the Languages of the

World. New York: Scribner.Yang, C. (2016). The Price of Linguistic Productivity: How Children Learn to Break the Rules

of Language. Cambridge: The MIT Press.

The Daily Linguist 81

waarin hij regelmatig aansluit op ac-tualiteiten. Van Oostendorp kan lek-ker kritisch zijn op maatschappelijke ontwikkelingen of pas verschenen onderzoek. Net zoals Paulien Cor-nelisse en Wim Daniëls observeert hij uiteenlopende taalfenomenen en be-spreekt hij ze met een portie humor - kijk bijvoorbeeld eens zijn vlog over de bijzondere spreekstijl van vloggers. Omdat hij taalkundige is, kan hij taal-fenomenen vaak goed verklaren. Hij draagt ook allerlei Nederlandse gedi-chten voor. Een toffe manier om ken-nis te maken met dichters van vroeger en nu! Verder laat hij andere taalkun-digen en Neerlandici aan het woord over hun onderzoek of hun favoriete woorden. Van harte aangeraden!

Volgende radio-editie van Groot Dictee der Nederlandse Taal

Taalenthousiasten, neerlandici en piet- lutten opgelet: de volgende editie van het Groot Dictee der Nederlandse Taal vindt dit jaar op zaterdagmorgen 2 no-vember 2019 plaats! Oud-Kamervoor-zitter Gerdi Verbeet zal de tekst, welke is opgesteld door schrijver en taal-kundige Wim Daniëls, tussen 11.00 en 13.00 voorlezen in Frits Spits’ De Taalstaat op NPO Radio 1. Wie besla-gen ten ijs wil komen kan zich boven-dien alvast oefenen op de website van Onze Taal (https://onzetaal.nl/nieu-ws-en-dossiers/weblog/zomerschool-groot-dictee/). Succes!

Een mooie mengelmoes (Nicoline van der Sijs & Marc van Oostendorp)

Hoera! Een mooie mengelmoes van Nicoline van der Sijs & Marc van Oost-endorp is sinds juni 2019 gratis te be-wonderen als PDF! In het werk wordt het eerste onderzoek naar het Neder-lands in een schat van zeventiende-eeuwse kranten gepresenteerd. De uitgave is te vinden via Neerlandis-tiek (https://www.neerlandistiek.nl/2019/06/een-mooie-mengel-moes/).

Still Alice (2014)Still Alice (2014) is a gripping film on the life of Alice Howland, professor of linguistics at Columbia University, who is diagnosed with early-onset Alz-heimer’s disease just after her fiftieth birthday. Based on the novel by the same name, written by neurosurgeon Lisa Genova. Still Alice is currently available on Dutch streaming plat-form Videoland.

Dai y inguistL

I Th

e

Books, movies and documentaries

Not your average influencer: Marc van Oostendorp op YouTube

Op YouTube (youtube.com/user/fonologo) plaatst hoogleraar Marc van Oostendorp vlogs en andere filmpjes,

Media

LingUU | 3.2 | 2019 The Daily Linguist82

Documenting every language in the world, one video at a time

Have you every wondered what Yid-dish, Tok Pisin or Kenyan Sign Lan-guage sound or look like? Wiktongues is a non-profit organization which encourages volunteers from all over the world to document the languages around them, and uses these record-ings to create an online database on the Wikitongues YouTube channel (htt-ps://www.youtube.com/user/wiki-tongues). The result is a vast amount of videos covering both well-known and more obscure languages. Learn more here: https://wikitongues.org/.

Practice your Dutch over koffie en thee!

For the internationals among us: are you keen on practicing your Dutch? You can do so over tea and biscuits during the weekly taalcafés, which are organized by Querido! Their taal-café has just moved to a new location in The Colour Kitchen (Zuilen). You can pop by every Saturday between 15h and 17h. Participation is free. For more information, visit: https://www.queridon.nl/nieuw-in-zuilen-que-ridons-taalcafe/.

Meet RU:ts! Students from the Linguistics depart-ment at Radboud University Nijmegen were inspired by LingUU Journal, and decided to make their very own stu-dent-led, open access journal: RU:ts!

Research Master Linguistics stu-dents Iris Faber and Myrthe Reuver are leading the editorial board on the

Other

journey to their very first issue, com-ing November, 2019. The first issue will (amongst others) have articles on phonology and memes, the language of depression, and multimodal ex-periments with ideophones. Informa-tion on the launch date, and a launch party, will come through the RU:ts Fa-cebook page (https://www.facebook.com/RUtsJournal/) and their website (http://ruts-journal.ruhosting.nl/). ■

Documents

Volume 1 | Issue 1 | Volume 3 | Issue 2| November 2019 · 2(b) and 2(d) show this for necessives, impersonal passives and possessives. All of these latter verbs take different subject