German Noun Plurals in Simultaneous Bilingual vs. Successive Bilingual vs. Monolingual Kindergarten Children: The Role of Linguistic and Extralinguistic Variables

Abstract

(1) Background: The complex phenomenon of German noun plural inflection is investigated in three groups of German-speaking kindergarten children: (a) monolinguals (1L1), (b) simultaneous bilinguals (2L1) also acquiring Croatian, and (c) successive bilinguals (L2) acquiring Turkish as L1. Predictions of the usage-based schema model and of Natural Morphology concerning different linguistic variables are used to explore their impact on plural acquisition in the three groups of children. (2) Methods: A longitudinal study (from mean age 3;1 to 4;8) is conducted using two procedures (a formal plural test and spontaneous recordings in kindergarten), and the data are analyzed using generalized linear (mixed-effects) regression models in R. (3) Results: All children produce more errors in the metalinguistically challenging test compared to spontaneous speech, with L2 children being particularly disadvantaged. Socioeconomic status (henceforth SES) and teachers’ plural type frequency are most relevant for 1L1 children, and kindergarten exposure is more relevant for L2 children, while the linguistic variables are more important for 2L1 children. (4) Conclusions: The main predictions of the schema model and of Natural Morphology are largely confirmed. All of the linguistic variables investigated show significant effects in some analyses, but morphotactic transparency turns out to be the most relevant variable for all three groups of children.

1. Introduction

A frequently discussed problem of studies on language acquisition is that they are difficult (if not impossible) to replicate because no single individual has exactly the same language experience as another one due to the sheer amount of potentially influencing external and internal variables that may also interact with each other (see, e.g., Kidd and Donnelly 2020 for first language acquisition; Titone and Tiv 2022 for bilingualism; Yamashita and Neiriz 2024 for language teaching). Usage-based theoretical approaches are currently widely acknowledged as state of the art in language acquisition research (e.g., Ellis and Wulff 2020; Morgenstern 2022; Chee et al. 2023; Geeslin et al. 2023; Cole and Dumont 2023, among many others) as they account for these individual differences while trying to assess the impact of different types of variables (e.g., age, socioeconomic status/SES, gender, age of onset of L2 acquisition, amount and properties of input by parents, teachers, siblings, peers in each language, kindergarten exposure, etc.) that shape individual language experiences. While low-SES children are known to be disadvantaged in their acquisition of vocabulary and grammar due to the reduced quantity and quality of their parents’ language input (e.g., Weizman and Snow 2001; Gathercole and Hoff 2007), these SES effects are often less pronounced in bilingual children due to certain advantages in cognitive flexibility and working memory gained from bilingual language experiences (e.g., Bialystok and Viswanathan 2009; Blom et al. 2014). Concerning gender, only some studies find differences in early language acquisition, with slight advantages for girls and higher variability in boys (for an overview, see Rinaldi et al. 2023). Especially for bilingual children, kindergarten exposure is important to acquire the majority language (e.g., Blom et al. 2012), which is an important prerequisite for their later success at school (e.g., Paradis et al. 2017). Nevertheless, monolingual low-SES children may also benefit from more time spent in kindergarten (e.g., Ready 2010) as kindergarten teachers’ input tends to compensate for social inequalities (e.g., Templ et al. 2018).

In addition to the extralinguistic variables mentioned above, formal features of linguistic phenomena that may vary much between the languages to be acquired may also impact the acquisition process. The criteria for determining the ease of acquisition of such formal features are extensively discussed within the usage-based schema model (henceforth SM, see Köpcke 1993, 1998; Wecker 2016; Köpcke and Wecker 2017) and within Natural Morphology (henceforth NM, see Dressler 2003; Laaha et al. 2006; Korecky-Kröll 2011).

This study investigates the morphological phenomenon of noun plural inflection in children acquiring German in monolingual, simultaneous bilingual, and successive bilingual contexts. The other languages acquired by the two groups of bilingual children are Croatian and Turkish. Noun plural inflection is an important morphological category occurring in all three languages despite showing quite different acquisition patterns that may be partially attributed to typological differences between the languages, as will be explained in the following subsections.

There were some practical reasons why Turkish and Croatian were chosen: First, both languages are spoken by large groups of migrants in Austria. In addition, we received ample and long-term funding to study the acquisition of German by children with Turkish backgrounds growing up in Austria, who were at that period considered to be the group of migrant children most disadvantaged by their relatively poor German skills, resulting in rather low educational attainment. In contrast, children with Croatian backgrounds were often mentioned as “model migrants” in public discourse as they tended to acquire German quite early (often as simultaneous bilinguals) and to reach higher educational levels. Further, Marina Camber chose simultaneous acquisition of Croatian and German as a subject of her thesis (Camber 2020) because she could not find any successive bilinguals with this language background.

Overall, noun plurals constitute an appropriate testing ground for assessing children’s acquisition of morphology1.

1.1. German Noun Plurals

In this study, we focus on children’s German output produced in an experimental elicitation test as well as on their spontaneous output produced in teacher–child interaction in kindergarten for the following reasons: (1) These two procedures are comparable in all three groups of children, (2) German is the language spoken by all children investigated, and (3) among the three languages, German has the most challenging system of noun plural inflection.

German noun plural inflection has been a hotly debated issue since the 1970s (for an overview see, e.g., Korecky-Kröll et al. 2018): While it has been claimed that there is a difference in acquisition and processing between a regular (default) plural and irregular plurals in the case of generativist dual-route models (see, e.g., Clahsen 1999), usage-based and connectionist single route models (e.g., the SM by Köpcke 1993, 1998; Goebel and Indefrey 2000; Hahn and Nakisa 2000) emphasize the importance of type and token frequency, salience, and cue validity of forms and NM (e.g., Dressler 2003; Klampfer et al. 2001; Laaha et al. 2006; Korecky-Kröll 2011) argued in favor of additional formal features, such as constructional iconicity, morphotactic transparency, and productivity. Child language acquisition results have been used as important external evidence for the use of different models in most of these studies.

The acquisition of German noun plurals has often been described as a u-shaped curve of correct forms (e.g., Korecky-Kröll et al. 2018): After an early phase characterized by many correct rote-learned plural forms as well as by noun-external means, such as numerals or quantifiers to express plurality (e.g., Wecker 2016), children start overapplying the most frequent plural markers (usually +en and +e, see, e.g., McCurdy et al. 2020) to other forms, resulting in many overgeneralization errors (corresponding to the low point of the u-shaped curve) before finally arriving at productively used target-like forms. Although both L1 and early L2 acquisition follow these pathways, L2 children may have specific difficulties with some aspects of German noun plurals depending on their L1. For example, children with L1 Turkish have more problems in discovering the gender-based regularities of German plural marking because Turkish does not have gender (e.g., Wegener 2008).

As shown in

Table 1. The eight plural markers of Standard German with examples.

No.	Plural Marker	Singular	Plural	English Plural
1	+s	Baby	Baby+s	babies
2	+(e)n	Katze	Katze+n	cats
3	+e	Hund	Hund+e	dogs
4	umlaut+e	Zug	Züg+e	trains
5	zero	Teller	Teller	plates
6	umlaut	Apfel	Äpfel	apples
7	+er [ɐ]	Bild	Bild+er	pictures
8	umlaut+er [ɐ]	Haus	Häus+er	houses

, Standard German has seven overt plural markers consisting of suffixes and/or umlaut (i.e., stem vowel change) and one zero morpheme. The three grammatical gender classes (masculine, feminine, and neuter) are mostly marked by definite and indefinite articles in the singular and only rarely by derivational suffixes on nouns (e.g., -in for female persons such as in Lehrerin ‘teacher-SG:FEM’). In addition, the article system exhibits a lot of syncretism, and although plural forms are gender-neutral (the nominative and accusative definite article is die for plural nouns of all genders, but also for feminine singulars), the gender of the noun is relevant for the selection of the plural marker. For example, nouns of feminine gender have a very high probability of taking an -(e)n plural, whereas -e is more probable for masculine and neuter nouns and -s is rare for nouns of all three genders, but particularly rare for feminines (see, e.g., McCurdy et al. 2020).

As a representative of the usage-based perspective, Köpcke (1993, pp. 81–91; 1998, p. 309) formulated an SM of Standard German number formation by identifying several features (the number of syllables, the word-final phonology of the noun, and form of the definite article as a more or less reliable indicator of the grammatical gender of the noun) as crucial for the classification as a more or less prototypical singular or plural according to their cue strength2 for singularity and plurality. In this model, German singular and plural nouns may thus be represented by the following continuum (Köpcke 1998, p. 309; see also Korecky-Kröll et al. 2018):

• Monosyllabic, final stop, article der or das (a prototypical singular of masculine or neuter gender, e.g., der Hund ‘the-MASC dog’);
• Polysyllabic, final -er (usually pronounced as [ɐ]), article der or das (probably a singular of masculine or neuter gender, e.g., der Teller ‘the-SG:MASC plate’, but rarely also a genitive plural ([der Stapel] der Teller ‘[the stack of] the-PL:GEN plates’);
• Polysyllabic, final -e, article die (with equal probability either a feminine singular or a plural of any gender, i.e., neither a prototypical singular nor a prototypical plural, e.g., the rhyming words die Schleuse ‘the-SG:FEM sluice’ vs. die Mäuse ‘the-PL mice’, but also the examples die Katze ‘the-SG:FEM cat’ vs. die Hunde ‘the-PL dogs’ in Table 1).
• Polysyllabic, final -er, article die (probably a plural of masculine or neuter gender, e.g., die Teller ‘the-PL plates’, and less probably a feminine singular, e.g., die Mauer ‘the-SG:FEM wall’);
• Polysyllabic, final -en, article die (a prototypical plural, e.g., die Katzen ‘the-PL cats’ from the feminine Singular die Katze ‘the cat’, but also die Daumen ‘the-PL thumbs’, a zero plural from the masculine singular der Daumen ‘the-SG:MASC thumb’).

In the case of zero plurals (No. 5 in Table 1), the plural is formally identical with the corresponding singular form and can be disambiguated for number only by the determiner (1a) or the lack thereof (as German—like English—uses no articles in indefinite plural constructions, such as in 1b) or by optional attributive adjectives (1c). If the zero plural is used in subject position, the inflected verb form agreeing with the subject acts as another cue for number disambiguation (1d). The examples in (1) are drawn from the spontaneous speech data of the children involved in this study:

(1)	a.	ich	mag	die	Fühler.	(L2_02Hf, 2;11)
		I	like-1SG	the-PL	feeler-0PL
		‘I like the feelers/antennae.’
	b.	mach~ma		noch	Anhänger?	(1L1_05Lm, 3;1)
		make~we		still	trailer-0PL
		‘Do we still make trailers?’
	c.	dick+e		blau+e	Streifen.	(1L1_25Lm, 3;0)
		thick-PL-INDEF		blue-PL-INDEF	stripe-0PL
		‘Thick blue stripes.’
	d	das		sind	Kuchen.	(1L1_22Lf, 2;11)
		this		be-3PL	cake-0PL
		‘These are cakes.’

Wecker (2016), as well as Köpcke and Wecker (2017, pp. 83–84), refined Köpcke’s original schema model by including s plurals and forms endings in el (which may be singulars or plurals) and proposed the following reliability scale of endings signaling plurality in German nouns that will also be used in our analyses (see column A of

Table 2. Scales of linguistic variables according to the SM and to NM.

A: Plural Reliability (SM)	B: Constructional Iconicity (NM)	C: Morphotactic Transparency (NM)	D: Productivity (NM)	E: Total Naturalness (NM): B+C+D
0: 0 (Hund)	0: Teller	0: Äpfel, Häuser	0: Äpfel, Häuser	1: Äpfel
1: -e (Katze, Hunde)	1: Äpfel	1: Bilder, Hunde	1: Züge	2: Häuser
2: -el (Stapel)	2: Züge, Häuser	2: Babys, Katzen,	2: Hunde, Teller	3: Züge
3: -er (Teller)	3: Babys, Katzen,	Teller	3: Babys, Katzen	4: Teller, Bilder
4: -s (Babys)	Bilder			5: Betten, Hunde
5: -en (Katzen)				6: Schuhe
				7: Stifte
				8: Babys, Katzen

, from most to least reliable): (e)n > s > er > el > e > 03.

Similarly, proponents of NM (e.g., Dressler 2003; Laaha et al. 2006; Korecky-Kröll 2011) argue for a slightly different scale according to three parameters of NM, namely constructional iconicity, morphotactic transparency, and productivity (see columns B, C, and D of Table 2). Following Peirce (1965), NM describes constructional iconicity as the relationship between form and meaning in the sense that the core meaning of plural, “more”, should also be reflected in a longer plural form compared to the corresponding singular (e.g., Mayerthaler 1981; Dressler 2000, pp. 288–90; Korecky-Kröll 2011, p. 142). Thus, constructional iconicity is highest in suffixed plurals as in German Kind+er ‘children’ or English kid+s, followed by plurals with suffixes and umlaut (Seifert 1988, p. 29), as in German Züg+e ‘trains’ or English brethr+en, lower in pure umlaut plurals, as in German Äpfel ‘apples’ or English feet, and non-existent in zero plurals4, as in German Teller ‘plates’ or English fish (Dressler 2000, p. 290). Subtractive plurals, which do not exist in Standard German, are anti-iconic.

In contrast, morphotactic transparency is highest in suffixed plurals, but also in “trivially transparent” zero plurals because they do not show any “opacifying obstruction” (Dressler 2000, p. 290), lower in plurals with final stop revoicing (e.g., singular Bild pronounced as [bɪlt]—plural Bilder [bɪldɐ] ‘pictures’, see Laaha and Dressler 2012) and lowest in plurals with umlaut (no matter whether they are suffixed or not).

In inflection, productivity refers to the ability of a morphological marker to form new potential inflectional word forms (e.g., when integrating loanwords, see Dressler 2003; Laaha et al. 2006). Productivity is assumed to be gradual, depending on the number of potentially competing productive patterns for a singular form according to its gender and its word-final phonology (for details see Laaha et al. 2006, pp. 279–84, partially refined in Korecky-Kröll 2011, p. 208): Thus, full productivity is assumed for forms with no potential competitor, i.e., -n plurals of feminine (e.g., Katze-n ‘cats’) and animate masculine nouns ending in e schwa (e.g., Hase-n ‘hares’) and for -s plurals of nouns of all genders ending in unstressed full vowels other than -a (e.g., Babys ‘babies’), whereas forms with one competitor (e.g., -en or -s plurals for feminine nouns ending in -a or consonant, zero or -s plurals for masculine nouns ending in -en or -er) are considered productive. A lower degree of productivity is assigned to weakly productive plurals with two potentially competing forms (e.g., -e, -umlaut + e and -s plurals competing for monosyllabic masculine nouns with umlautable stem vowels such as Zug ‘train’), while pure umlaut plurals and -er plurals with and without umlaut are classified as unproductive, as they cannot be applied to recent loanwords. Constructional iconicity, morphotactic transparency and productivity may be summarized as a composite variable of “Total naturalness” (see column E of Table 2).

1.2. Croatian Noun Plurals

In contrast to the German noun plural system, the Croatian system is morphologically richer (insofar as it has more case forms), more regular and more transparent than the German system (Camber 2020, p. 131). It has the same three genders as German (i.e., masculine, feminine, and neuter), but no articles (Camber 2020, p. 80) and no zero plural. Instead, Croatian nouns have clearly gender-dependent suffixes, each gender class has its own prototypical declension, and the majority of nouns belong to one of the three prototypical declension classes named after their genitive singular endings (Kovačević et al. 2009, p. 155). Thus, prototypical feminine nouns ending in -a belong to the e-declension (i.e., substitute -a with +e), while prototypical masculine and neuter nouns constitute the a-declension (Kovačević et al. 2009, p. 155, see the three left columns of

Table 3. Plural marking according to declension classes in Croatian vs. plural marking according to vowel harmony in Turkish.

Prototypical Declension Classes in Croatian			Plural Marking in Turkish
Gender	SG	PL	Vowel harmony	SG	PL
feminine	-a	+e	front vowel	–	+ler
(e-declension)	djevojčica ‘girl’	djevojčice ‘girls’		kedi ‘cat’	kediler ‘cats’
masculine	consonant, -o	+a	back vowel	–	+lar
(a-declension)	auto ‘car’	auta ‘cars’		araba ‘car’	arabalar ‘cars’
neuter	-o, -e	+a
(a-declension)	drvo ‘tree’	drva ‘trees’

).

A few additional rules concern morphonologically conditioned palatalization or vowel insertion (for details see Kovačević et al. 2009, pp. 155–58), but overall, the selection of the plural marker is much more predictable than in German. Apart from a few errors related to pluralia tantum, which were inflected like singular forms, Kovačević et al. (2009) did not find any problems concerning the acquisition of noun plurals in their case study on the monolingual acquisition of Croatian.

However, in their contrastive analysis of German and Croatian noun plurals produced by the bilingual children that are also investigated in this study, Camber (2020), as well as Camber and Dressler (2022), discovered a mutual influence of both languages in terms of a preference for +e suffixation (the only homophonous one) in both languages (in correct plurals as well as in overgeneralizations) as -e is the most important plural marker for feminine nouns in Croatian and for masculine and neuter nouns in German. As the only homophonous plural suffix it is clearly preferred by the bilingual children, and also overgeneralized to nouns of other genders (particularly to Croatian masculine nouns and to German feminine nouns where -e plurals are unproductive).

It is important to note that German plurals were acquired according to Croatian input conditions and vice versa. Thus, German productive zero plurals also result in ungrammatical Croatian zero plurals, whereas the bilingual children produced fewer zero plurals than monolingual children. These developments represent blind alley developments (Dressler et al. 2020, 2023), which children give up very soon due to the constant contrasting child-directed speech. They can be interpreted as an overlay of target structures in terms of a superposition, a concept adapted from quantum physics by Mattiello and Dressler (2022).

1.3. Turkish Noun Plurals

Turkish noun plural formation is even more regular and transparent. In contrast to Croatian and German, Turkish has no gender and no adjective inflection and—unlike German—no definite articles, but bir ‘one’ is also used as an indefinite article (see, e.g., Kornfilt 1997, p. 106; Ketrez 2023). The plural suffix -lAr follows the rules of vowel harmony (i.e., words with front vowels take the -ler allomorph, which also contains a front vowel, e.g., kedi+ler ‘cat+PL’, whereas words with back vowels take the -lar allomorph, e.g., araba+lar ‘car+PL’, see the three right columns of Table 3). Vowel harmony has been shown to be acquired early and without any problems in first language acquisition of Turkish (for production by age 2, see, e.g., Aksu-Koç and Slobin 1985; Ketrez and Aksu-Koç 2009; for comprehension already around 10 months, see, e.g., Altan et al. 2016), and therefore, the selection of the correct plural allomorph is also unproblematic for Turkish-speaking children (Ketrez and Aksu-Koç 2009).

In contrast to German and Croatian, there is no number agreement between plural numerals or other plural quantifiers and the noun (Ketrez and Aksu-Koç 2009, p. 20), i.e., the noun always appears in the singular, as the NP is already sufficiently marked for plurality by the quantifier (e.g., iki kedi ‘lit. two cat’), and a plural marking on the noun would be ungrammatical in this context. This makes contexts of obligatory overt plural marking on the noun in German NPs with quantifiers particularly challenging for bilingual children with L1 Turkish (e.g., Korecky-Kröll et al. 2018).

1.4. Research Questions

Our main research questions are as follows: Which linguistic variables predicted by the two different theories (SM and NM) best explain the German nouns plurals produced by our three diverse groups of children, i.e., monolingual German-acquiring children, simultaneous bilingual children acquiring Croatian and German and successive bilingual children acquiring first Turkish and then German? Which other (sociodemographic, input, or procedure) variables may additionally be involved?

2. Methodology

2.1. Participants

As mentioned in Section 1, we compare three groups of kindergarten children with a mean age of 3;1 at the beginning of the study who were growing up in the same city (Vienna/Austria) but in different acquisition settings: (1) monolingual children (1L1) acquiring only German (N = 29), (2) successive bilingual children (L2) acquiring first Turkish as their family language and German as their early L2 in kindergarten (N = 27), and (3) simultaneous bilingual children (2L1) acquiring both Croatian and German from birth (N = 4), i.e., 60 children in total (31 from higher SES (HSES), 29 from lower SES (LSES) families, 31 boys and 29 girls, see

Table 4. Numbers of children per group according to their background variables.

Group	Acquisition Setting	Languages	SES	Gender	Total
(1)	1L1	German	high	7 boys, 8 girls	15
	1L1	German	low	8 boys, 6 girls	14
(2)	L2	Turkish, German	high	6 boys, 8 girls	14
	L2	Turkish, German	low	7 boys, 6 girls	13
(3)	2L1	Croatian, German	high	2 boys	2
	2L1	Croatian, German	low	1 boy, 1 girl	2
	Total				60

). We are aware that the sample size of group (3) is much smaller than that of the two other groups because it was investigated within a dissertation project with fewer resources, whereas the two other groups were part of the larger third-party financed INPUT project. However, this does not negatively affect the results as we use generalized linear and generalized linear mixed-effects models, which are robust against small sample sizes (see, e.g., Schielzeth et al. 2020). Nevertheless, the results from group (3) should be considered with more caution than the results of the two other groups.

Simultaneous bilingual acquisition means that such children start and continue to acquire both languages from birth, or at least during early infancy, in a relatively balanced way (typically from their parents), whereas successive bilingual acquisition corresponds to monolingual acquisition of the family language from birth followed by early acquisition of the second language (typically via daycare from about age two onwards). This results in quite different acquisition paths, which are obviously also different from monolingual acquisition (e.g., De Houwer 2021).

The primary criterion for the assessment of SES was the highest level of education of the parental main caregiver according to the International Standard Classification of Education 1997 (ISCED-97, see OECD 1999). The LSES group comprised ISCED-97 levels 1-3b (i.e., from compulsory schools to apprenticeships and vocational schools without high school diploma), whereas the HSES group had ISCED-97 levels 3a-6 (i.e., from high school diploma to PhD). As a secondary criterion, the prestige of the main parental caregiver’s profession according to the International Socioeconomic Index of Occupational Status (ISEI, see Ganzeboom and Treiman 1996) was also assessed. Only if a parent was classified as LSES according to the educational level but had a profession that was significantly more prestigious (at least 10 points above the median of our sample, see Czinglar et al. 2015) was the child “upgraded” to the HSES group. However, for most parents, the educational level and the ISEI were highly correlated and, thus, only three out of sixty children (one 1L1 and two L2 children) received this upgrade (see also Korecky-Kröll et al. 2019).

All of the information on acquisition settings, SES, and kindergarten exposure (in hours, since kindergarten start) was obtained from interviews with the children’s parental main caregivers (see Appendix A for the relevant questions).

2.2. Methods

We used two different methods to investigate children’s acquisition of noun plurals over a time span of 1 ½ years:

• A plural elicitation test in German (conducted twice by the researchers with all children) using a procedure similar to Berko’s (1958) famous wug test was administered; however, real words were used instead of pseudowords because the results should be comparable to actual plurals produced in spontaneous speech (for details on the procedure and the materials see Appendix A);
• One-hour-long spontaneous speech audio and video recordings of teacher–child interaction in kindergarten were conducted and the “best” 30 minutes, i.e., the ones with the most direct verbal interaction between the teacher5 and the target child, were selected by the researchers to be transcribed by advanced students. For German kindergarten recordings, we analyzed all recordings available (4 recordings of 30 minutes for 58 children, and 3 recordings of 30 minutes for one LSES L2 girl who entered kindergarten only before the second datapoint), while for one HSES L1 girl, we had no permission to make recordings in her kindergarten, and thus, only her plural test was included.

The spontaneous speech recordings were carried out by the researchers with a video camera and a small audio recorder that the kindergarten teacher wore around their neck. They took place at children’s mean ages, 3;1, 3;4, 4;4, and 4;8, whereas the plural test elicitation tests were conducted only at mean ages of 3;4 and 4;8, as several other tests required for the project had to be conducted at the other datapoints. The data were transcribed in CLAN (MacWhinney 2000) and tagged for parts of speech and morphology using CLAN and MS Excel. Each target plural, as well as each actual plural, was then tagged according to the preference scales shown in Table 2. Finally, we conducted generalized linear mixed-effects regression models and generalized linear models to investigate the effects of different independent sociodemographic factors (language background, SES, and gender), of the procedure (test vs. spontaneous speech), input variables (log-normalized cumulative teacher plural types and tokens, i.e., decimal logarithms of all plurals used by the kindergarten teachers in the current recording and all previous recordings6, log-normalized total estimated hours of kindergarten exposure according to the interviews with parents), as well as linguistic variables represented by the preference scales on children’s production of noun plurals (see Table 2). The main dependent variables were all binomial (coded as 1 and 0): (1) correct vs. incorrect plurals (2) incorrect zero plurals (which are often repetitions of the singular stimulus forms used for elicitation in the plural test and usually evidence of less advanced plural development) vs. correct plurals and overt overgeneralizations, and (3) overt overgeneralizations (i.e., the use of another overt plural marker instead of the standard German marker) vs. correct plurals and incorrect zero plurals. Overt overgeneralizations are generally expected to be found in more advanced younger children who have already discovered some patterns of German plural formation that they may now apply to other forms and thus create potential, although non-existent plurals (see, e.g., Korecky-Kröll et al. 2018). A fourth category of “other non-target plurals” comprised a few lexical replacements (e.g., Ratten ‘rats’ instead of Mäuse ‘mice’), other morphological forms such as diminutives (e.g., Vogis ‘bird-DIM-PL’ instead of Vögel ‘birds’), truncations7 (Zu instead of Zug or Züge ‘train(s)’), reduplications (Zuzu ‘train(s)’), and three answers in Croatian. Due to the heterogeneity of answers, the category of “other non-target plurals” was excluded from further analyses. Raw numbers of all four categories are listed in

Table A2. Raw numbers of plural tokens produced by the three groups of children in the German plural test and in spontaneous speech in kindergarten.

Procedure		(1) 1L1	(2) 2L1	(3) L2	Total Tokens
Test	Correct plurals	605	49	186	840
	Incorrect zero plurals	320	95	608	1023
	Overt overgeneralizations	205	15	81	301
	Other non-target plurals	56	10	72	138
Spontaneous	Correct plurals	682	114	288	1084
	Incorrect zero plurals	22	2	14	38
	Overt overgeneralizations	26	2	13	41
	Other non-target plurals	6		1	7
Total tokens		1954	286	1450	3690

in Appendix B. For the first three categories, models were compared according to their Akaike information criterion (AIC) values, with lower AIC values indicating better model fits. Of the 730 models conducted, only the best models with the lowest AIC values, as well as the simplest models containing only single independent variables, are reported here. The following significance levels were adopted: p < 0.001 (***), p < 0.01 (**), and p < 0.05 (*); non-significant trend: p < 0.1 (.).

3. Results

The results section will focus on the differences between the three groups of children. In Section 3.1, we will provide a brief analysis of all children and show that it is more appropriate to investigate both procedures (plural test and spontaneous speech) separately. Consequently, Section 3.2 and Section 3.3. deal with the results of the plural test and of the spontaneous speech recordings.

3.1. All Children: Effects of Language Background, SES and the Procedure on German Plural Production

As shown in

Table 5. All children: language background * SES * procedure.

Fixed Effects	Estimate	Standard Error	z Value	p Value	Significance
(Intercept)	2.8354	0.2363	12.000	<0.001	***
Lgbg2L1	0.6717	1.0822	0.621	0.534812
LgbgL2	−0.6449	0.3437	−1.877	0.060571	.
SESLSES	−0.5587	0.3240	−1.725	0.084609	.
ProcTest	−2.3773	0.2299	−10.340	<0.001	***
Lgbg2L1:SESLSES	0.3400	1.2819	0.265	0.790829
LgbgL2:SESLSES	1.1446	0.6324	1.810	0.070331	.
Lgbg2L1:ProcTest	−2.0106	1.0730	−1.874	0.060953	.
LgbgL2:ProcTest	−1.1643	0.3408	−3.416	0.000634	***
LSES:ProcTest	−0.3189	0.3154	−1.011	0.311941
Lgbg2L1:SESLSES:ProcTest	0.4797	1.2686	0.378	0.705368
LgbgL2:SESLSES:ProcTest	−0.4806	0.6315	−0.761	0.446619
Significance levels:	*** 0.001	** 0.01	* 0.05	. 0.1 (trend)

and Figure 1, the best overall generalized linear mixed-effects model for the three groups of children and both procedures indicates significant effects of the procedure (Proc, with the levels spontaneous speech and test) and of the language background (Lgbg, with the levels 1L1, 2L1, and L2), including mutual interaction effects as well as trends involving SES (levels: HSES and LSES).

The effect of the procedure is highly significant. Children produce fewer correct plurals in the test than in spontaneous speech (z = −10.340 ***), and there is also a significant interaction effect between language background and the procedure indicating a particularly strong disadvantage for L2 children in the test (z = −3.416 ***), whereas there is only a trend for 2L1 children (z = −1.874.) to produce fewer correct plurals in the test than 1L1 children. Overall, L2 children show a slightly lower performance than both groups of L1 children (z = −1.877.). There is also an overall SES trend indicating fewer correct plurals in LSES children (z = −1.725.). However, an opposite interaction trend is found for L2 children (z = 1.810, p = 0.070.), indicating that HSES and LSES L2 children show a very similar overall performance, with a minor advantage for HSES L2 children in the test and a minor advantage for LSES L2 children in terms of spontaneous speech (see Figure 1). As the patterns of the three groups of children differ considerably according to the procedure, the results will be described separately for both procedures (for the test in Section 3.2 and for spontaneous speech in Section 3.3).

3.2. German Plural Test Results in the Three Groups of Children

3.2.1. Correct Plurals in the Plural Test

The best generalized linear models for correct plural tokens produced by the three groups of children are shown in

Table 6. Correct plural tokens in the plural test (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	SES * Datapoint + Productivity + cumulative teacher plural type frequency (log)	Datapoint + Morphotactic transparency	Datapoint + Morphotactic transparency
Variables and variants of the best models (z values and significance levels)	SESLSES (−3.664 ***), Datapoint 4 (3.495 ***) Productivity (10.635 ***), cumulative teacher PL type frequency (log) (1.397) SESLSES * Datapoint 4 (−1.966 *)	Datapoint 4 (4.317 ***), Morphotactic transparency (5.471 ***)	Datapoint 4 (2.901 **), Morphotactic transparency (8.994 ***)

. All groups show a significant effect of datapoint (1L1: z = 3.495 ***; 2L1: z = 4.317 ***; L2: z = 2.901 **) indicating a significant increase in correct plural tokens from the mean age of 3;4 to 4;8 in the plural test. Despite the fact that SES yields a significant main effect (z = −3.664 ***) as well as an interaction effect with the datapoint (z = −1.966 *) in 1L1 children, it is not significant in the two groups of bilingual children. Concerning the linguistic variables, both groups of bilingual children show a particularly strong preference for transparent plurals (2L1: z = 5.471 ***; L2: 8.994 ***), whereas productivity is apparently more important for the monolingual children8.

To illustrate the effects of other potential variables on the production of correct plurals in the three groups of children,

Table 7. Correct plural tokens in the plural test (one independent variable models).

Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
Datapoint 4	5.252 ***	3.774 ***	2.588 **
SES LSES	−7.120 ***	−0.170	−1.40
Gender male (m)	−0.734	−1.461	−0.856
Kindergarten exposure hours (log)	3.124 **	4.188 ***	2.413 *
Cumulative teacher PL type frequency (log)	4.553 ***	3.765 ***	1.144
Cumulative teacher PL token frequency (log)	4.158 ***	3.921 ***	2.005 *
PL schema reliability	3.207 **	3.196 **	7.026 ***
Constructional iconicity	0.952	−3.096 **	−12.064 ***
Morphotactic transparency	9.728 ***	5.197 ***	8.948 ***
Productivity	10.256 ***	5.059 ***	9.757 ***
Total PL naturalness	8.945 ***	3.623 ***	4.184 ***

shows the effects of the different independent variables investigated within simple models containing only the respective independent variable. We find many similarities in the three groups of children. The datapoint, the kindergarten exposure in hours, and the cumulative plural token frequencies of the kindergarten teachers show significant main effects, such as the linguistic variables of plural schema reliability, morphotactic transparency, productivity, and total plural naturalness, although to different extents. However, the actual teacher input is most important for the 1L1 children, followed by the 2L1 children, and kindergarten exposure shows the strongest effect in 2L1 children, followed by the 1L1 children—but both of these input variables show only weak effects in L2 children. In contrast, schema reliability is most important for the L2 children. However, both groups of bilingual children show a dispreference for iconic forms (i.e., they prefer zero plurals over suffixed plurals), whereas 1L1 children do not show any preference in this respect.

3.2.2. Incorrect Zero Plurals in the Plural Test

As mentioned in Section 2, incorrect zero plurals are often considered to be evidence of less advanced language development. In the plural test, they correspond to repetitions of the given stimulus form, but as zero plurals are one of the main plural markers of German (see Table 1), some of them may nevertheless be considered potential (although not actual) plural forms.

Table 8. Incorrect zero plural tokens in the plural test (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	SES + Actual PL schema reliability + cumulative teacher PL type frequency (log)	SES * Datapoint + Actual PL productivity	SES * Actual PL total naturalness
Variables and variants of the best models (z values and significance levels)	SESLSES (6.167 ***), Actual PL schema reliability (−13.933 ***), cumulative teacher PL type frequency (log) (−2.881 **)	SESLSES (−0.687), Datapoint 4 (−4.275 ***), Actual PL productivity (−5.730 ***) SESLSES * Datapoint 4 (3.043 **)	SESLSES (2.943 **), Actual PL total naturalness (−9.744 ***) SESLSES * Actual PL total naturalness (−2.804 **)

shows the best generalized linear models for incorrect zero plural tokens in the three groups of children. SES yields a significant main effect as well as an interaction effect with datapoint in the 1L1 and in the L2 children. Compared to their HSES peers, both LSES 1L1 children and LSES L2 children produce more incorrect zero plurals in the plural test. However, both groups of bilingual children show interaction effects of SES with other variables: LSES 2L1 children produce more incorrect zero plurals at the later datapoint, while the overall number of incorrect zero plurals decreases with age in the group of 2L1 children, as shown by the significant main effect of datapoint. The input variable of cumulative teacher plural type frequency is significant only for the 1L1 children. Regarding the linguistic variables, the most significant effects are found for actually produced incorrect zero plurals, but they are all negative. We find particularly little schema reliability in 1L1 children, particularly little productivity in 2L1 children and particularly little total naturalness in L2 children.

Table 9. Incorrect zero plural tokens in the plural test (one independent variable models).

Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
Datapoint 4	−3.422 ***	−4.692 ***	−1.217
SES LSES	7.769 ***	1.707 .	3.548 ***
Gender m	−0.048	0.628	−0.251
Kindergarten exposure hours (log)	−1.869 .	−4.863 ***	−3.127 **
Cumulative teacher PL type frequency (log)	−3.292 ***	−4.105 ***	0.414
Cumulative teacher PL token frequency (log)	−2.402 *	−4.393 ***	−1.291
Target PL schema reliability	−1.441	−0.414	−2.926 **
Actual PL schema reliability	−14.443 ***	−6.682 ***	−15.21 ***
Target PL constructional iconicity	2.854 **	2.994 **	10.913 ***
Actual PL constructional iconicity	−0.035	−0.014	−0.026
Target PL morphotactic transparency	−6.507 **	−3.628 ***	−7.679 ***
Actual PL morphotactic transparency	0.039	0.019	0.036
Target PL productivity	−4.111 ***	−3.165 **	−5.322 ***
Actual PL productivity	−13.761 ***	−7.093 ***	−16.31 ***
Target PL total naturalness	−3.252 **	−1.632	−0.655
Actual PL total naturalness	−13.78 ***	−0.006	−12.26 ***

gives some additional insights into the effects of the different independent variables on the production of incorrect zero plurals (again investigated within the models exclusively examining the respective variable). This time, we find more differences than similarities. While both groups of L1 children make significant progress from age 3;4 to 4;8 in producing fewer incorrect zero plurals at the later datapoint, the progress made by the L2 children is non-significant. SES has the strongest impact on 1L1 children, followed by L2 children, whereas there is only a trend in 2L1 children to produce more incorrect zero plurals. While less kindergarten exposure is more associated with higher numbers of incorrect zero plurals in both groups of bilingual children, the actual input plural types and tokens produced by the teachers significantly impact only both groups of L1 children. The linguistic variables are not particularly revealing for this category: incorrect zero plurals appear more frequently when the target plural is less transparent, less productive and more iconic (in all groups of children), less reliable (in L2 children), and less natural (in 1L1 children). However, the actual incorrect zero plural forms are generally less reliable, less productive and less natural than their corresponding correct target forms and they also show no effects of constructional iconicity or morphotactic transparency. Thus, it seems that incorrect zero plurals are used as a sort of emergency default plural (Korecky-Kröll 2011, pp. 245–48) when children are unable to decide in favor of an overtly marked plural form.

3.2.3. Overt Overgeneralizations in the Plural Test

The last category to be investigated in the plural tests are overt overgeneralizations, i.e., overapplications of plural suffixes and/or umlaut to other plural forms that do not take these markers in Standard German. As mentioned in Section 2, overt overgeneralizations are often found in younger, more advanced children who have acquired some, but not all patterns of German noun plural formation. The results for overt overgeneralizations produced by the three groups of children are shown in

Table 10. Overt overgeneralization tokens in the plural test (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	Actual PL constructional iconicity	SES * Datapoint + Target PL productivity	SES * Kindergarten exposure hours (log) * Actual PL total naturalness
Variables and variants of the best models (z values and significance levels)	Actual PL constructional iconicity (8.924 ***)	SESLSES (0.011), Datapoint 4 (0.012), Target PL productivity (−2.791 **), SESLSES * Datapoint 4 (−0.013)	SESLSES (2.387 *), Kindergarten exposure hours (log) (3.273 **), Actual PL total naturalness (3.286 **) SESLSES * Kindergarten exposure hours (log) (−2.582 *), SESLSES * Actual PL total naturalness (−2.395 *), SESLSES * Kindergarten exposure hours (log) * Actual PL total naturalness (2.542 *)

. While both groups of L1 children show only significant effects of two different linguistic variables (constructional iconicity of the actual plural form produced in 1L1 children: z = 8.924 **; productivity of the target plural form in 2L1 children: z = −2.791 **), the L2 children show significant main effects of SES (z = 2.387 *), kindergarten exposure (3.273 **) and the linguistic variable of total naturalness of the actual plural form (z = −2.395 *). In addition, there is a complex three-way interaction between SES, kindergarten exposure and total naturalness in the L2 children (see Figure 2): overt plural overgeneralizations of LSES L2 children increase as a function of their hours of kindergarten exposure, indicating some intermediate progress in the acquisition of German plural patterns, whereas HSES L2 children’s overgeneralizations were more frequent at an earlier point in time (when the children had still less kindergarten exposure) but tend to decrease in favor of correct plurals later on.

Additional simple models are again shown in

Table 11. Overt overgeneralization tokens in the plural test (one independent variable models).

(1) Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
(2) Datapoint 4	−1.928.	2.596 **	1.397
(3) SES LSES	−1.605	−2.260 *	−3.619 ***
(4) Gender m	0.022	1.066	−0.826
(5) Kindergarten exposure hours (log)	−1.591	2.714 **	3.671 ***
(6) Cumulative teacher PL type frequency (log)	−1.153	1.844.	1.462
(7) Cumulative teacher PL token frequency (log)	−1.356	2.230 *	2.596 **
(8) Target PL schema reliability	−0.935	−1.817.	−3.789 ***
(9) Actual PL schema reliability	5.314 ***	2.857 **	7.562 ***
(10) Target PL constructional iconicity	−3.160 **	0.000	0.596
(11) Actual PL constructional iconicity	8.924 ***	0.009	7.636 ***
(12) Target PL morphotactic transparency	−3.39 ***	−2.218 *	−3.638 ***
(13) Actual PL morphotactic transparency	1.575	−3.332 ***	−6.989 ***
(14) Target PL productivity	−7.247 ***	−2.563 *	−4.533 ***
(15) Actual PL productivity	−0.74	−0.880	2.747 **
(16) Target PL total naturalness	−6.032 ***	−2.133 *	−3.341 ***
(17) Actual PL total naturalness	8.675 ***	3.204 **	10.85 ***

. Despite the fact that kindergarten exposure, cumulative teacher plural token frequency, and high SES are positively associated with overt plural overgeneralizations in both groups of bilingual children, we find a significant increase in these overgeneralizations between datapoints only in 2L1 children, but a decreasing trend in L1 children (the apparently complex developmental difference between HSES and LSES L2 children is visible in Figure 2 but not reflected in Table 10).

The linguistic variables also show some interesting effects. While opaque, unproductive, and unnatural target plural forms are generally dispreferred by all three groups of children (see rows 12, 14, and 16 in Table 11), less reliable plural forms (row 8) are significantly less preferred by L2 children, but non-iconic forms (row 10) are mostly avoided by 1L1 children. However, children’s dispreference of certain forms leads to overgeneralizations of higher preference patterns only in some cases: all three groups of children prefer the production of more reliable and more natural overgeneralizations (rows 9 and 17), but only 1L1 and L2 children prefer more iconic forms, only both bilingual groups prefer more transparent forms, and only L2 children prefer more productive forms.

3.3. German Spontaneous Speech Results in the Three Groups of Children

The same models as for the plural tests were also conducted to investigate the plural acquisition patterns of the three groups of children in the spontaneous speech data collected at four datapoints in kindergarten.

3.3.1. Correct Plurals in Spontaneous Speech

As shown in

Table 12. Correct plural tokens in spontaneous speech (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	SES + Morphotactic transparency	Total PL naturalness	Constructional iconicity
Variables and variants of the best models (z values and significance levels)	SESLSES (−1.969 *), Morphotactic transparency (3.932 ***)	Total PL naturalness (2.165 *)	Constructional iconicity (−1.592)

, the best models for correct plural tokens in spontaneous speech show fewer significant effects than the corresponding models of the plural tests (see Table 5). SES is again significant in the 1L1 group only, indicating that monolingual LSES children produce fewer correct plurals than their HSES peers (z = −1.969 *). However, this SES effect is much weaker than the one found for the plural test (z = −3.664 ***). In contrast to the plural test, there is also no effect of datapoint, i.e., no significant progress in the production of correct plurals produced in kindergarten can be found for the three groups of children, apparently because most plural forms produced at the first datapoint were already correct. Furthermore, in general, the three groups of children show different linguistic preferences in spontaneous speech compared to the test. While 1L1 children exhibit a particularly strong preference for productive forms in the test, they produce mainly morphotactically transparent forms in spontaneous speech (z = 3.932 ***), but a certain preference for transparency is, nevertheless, apparent in the test. The two groups of bilingual children show a strong preference for transparent forms in the test, but in spontaneous speech, total naturalness is stronger in the 2L1 children (z = 2.165 *), whereas a clear preference cannot be discovered in the L2 children, although they seem to have a slight, but non-significant dispreference for iconic forms.

The additional simple models of

Table 13. Correct plural tokens in spontaneous speech (one independent variable models).

Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
Datapoint 2	0.933	−0.007	−0.500
Datapoint 3	1.893.	−0.006	−0.012
Datapoint 4	1.103	−0.006	0.041
SES LSES	−2.16 *	−0.312	1.083
Gender m	−1.848.	0.642	0.056
Kindergarten exposure hours (log)	0.643	−0.004	−0.936
Cumulative teacher PL type frequency (log)	1.176	−0.454	1.257
Cumulative teacher PL token frequency (log)	1.114	−0.776	1.706.
PL schema reliability	1.563	1.408	0.829
Constructional iconicity	0.968	1.721.	−1.592
Morphotactic transparency	4.045 ***	0.006	−0.291
Productivity	2.384 *	1.522	−1.169
Total PL naturalness	3.118 **	2.165 *	−1.073

show further preferences of total naturalness and productivity in 1L1 children, but no other effects for the two groups of bilingual children.

3.3.2. Incorrect Zero Plurals in Spontaneous Speech

Incorrect zero plurals are very rare in spontaneous speech (N = 38 for the entire corpus, see Table A2 in Appendix B). However, a few significant effects of linguistic variables can still be found in the best models (see

Table 14. Incorrect zero plural tokens in spontaneous speech (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	Actual PL total naturalness	Target PL total naturalness	Actual PL total naturalness
Variables and variants of the best models (z values and significance levels)	Actual PL total naturalness (−4.97 ***)	Target PL total naturalness (−1.533)	Actual PL total naturalness (−3.998 ***)

), although their interpretation is challenging.

A dispreference for total naturalness, as shown in the actual incorrect zero plural forms produced by the 1L1 and the L2 children (1L1: z = −4.97 ***; L2: −3.998 ***) points again to a zero-emergency default plural strategy that is used when there is no other form available for them (Korecky-Kröll 2011, pp. 245–48). No significant preference or dispreference is found in the 2L1 children as they only use two tokens of incorrect zero plurals.

The other simple models shown in

Table 15. Incorrect zero plural tokens in spontaneous speech (one independent variable models).

Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
Datapoint 2	−0.427	0.003	−0.017
Datapoint 3	−1.397	0.003	0.283
Datapoint 4	−0.755	0.000	−0.116
SES LSES	1.253	−0.531	0.172
Gender m	1.121	−0.450	−0.588
Kindergarten exposure hours (log)	−2.125 *	0.100	0.783
Cumulative teacher PL type frequency (log)	−1.826.	−0.031	−1.871.
Cumulative teacher PL token frequency (log)	−1.567	0.334	−1.870.
Target PL schema reliability	0.783	−0.989	−0.125
Actual PL schema reliability	−3.909 ***	−0.947	−4.606 ***
Target PL constructional iconicity	1.764.	−1.515	1.375
Actual PL constructional iconicity	−0.011	−0.005	−0.007
Target PL morphotactic transparency	0.147	−0.005	0.264
Actual PL morphotactic transparency	0.017	0.005	0.010
Target PL productivity	1.732.	−1.244	2.168 *
Actual PL productivity	−3.983 ***	−1.245	−0.010
Target PL total naturalness	1.820.	−1.533	2.172 *
Actual PL total naturalness	−4.97 ***	−1.523	−3.998 ***

are also not particularly revealing: incorrect zero plurals used in spontaneous speech are, in general, less preferable in terms of linguistic variables than their corresponding overtly marked target forms. A more detailed analysis of the complete syntactic contexts of these forms might shed light on additional variables involved, but for reasons of space, this must be postponed to a future paper.

3.3.3. Overt Overgeneralizations in Spontaneous Speech

Overt plural overgeneralizations (see

Table 16. Overt overgeneralization tokens in spontaneous speech (best model for each group of children).

Models and Variables	(1) 1L1	(2) 2L1	(3) L2
Best overall model (significant variables in bold, asterisk in bold for significant interaction effects)	SES + Actual PL productivity	Target PL total naturalness	SES + Actual PL productivity
Variables and variants of the best models (z values and significance levels)	SESLSES (1.471), Actual PL productivity (−4.611 ***)	Target PL total naturalness (−1.426)	SESLSES (−1.539), Actual PL productivity (−2.266 *)

) are almost as rare as incorrect zero plurals in spontaneous speech (N = 41 for all three groups of children, see Table A2 in Appendix B). Thus, we find again no significant effect in 2L1 children, who produce only two tokens of such forms. In contrast, the groups of 1L1 and L2 children show a dispreference for productive forms (1L1: z = −4.611 ***; L2: z = −2.266 *). However, the corresponding correct target forms are even less reliable, less iconic, less transparent, and less natural than the actually produced forms (see

Table 17. Overt overgeneralization tokens in spontaneous speech (one independent variable models).

Variables and Variants	(1) 1L1	(2) 2L1	(3) L2
Datapoint 2	−1.350	0.003	0.623
Datapoint 3	−1.663.	0.000	−0.662
Datapoint 4	−1.209	0.003	0.054
SES LSES	2.063 *	0.006	−1.440
Gender m	1.328	−0.450	0.774
Kindergarten exposure hours (log)	0.612	−0.094	0.765
Cumulative teacher PL type frequency (log)	0.183	0.655	0.778
Cumulative teacher PL token frequency (log)	0.097	0.738	0.183
Target PL schema reliability	−2.515 *	−0.004	−0.941
Actual PL schema reliability	0.627	0.004	0.278
Target PL constructional iconicity	−1.965 *	−0.863	1.174
Actual PL constructional iconicity	0.805	−0.040	1.368
Target PL morphotactic transparency	−5.243 ***	−0.005	0.000
Actual PL morphotactic transparency	−3.509 ***	−1.455	−0.060
Target PL productivity	−4.267 ***	−0.826	−0.572
Actual PL productivity	−4.673 ***	−0.266	−2.091 *
Target PL total naturalness	−4.845 ***	−1.426	−0.457
Actual PL total naturalness	−3.915 ***	−0.761	−1.083

).

4. Discussion

This study investigated the effects of different types of variables (sociodemographic, procedure, input, and linguistic features) on the acquisition of German noun plural morphology in three groups of kindergarten children: monolinguals/1L1 acquiring only German, simultaneous bilinguals/2L1 acquiring German and Croatian, and successive bilinguals/L2 acquiring Turkish as L1 and German as early L2. A formal test conducted twice and spontaneous speech recordings conducted four times in kindergarten yielded different results: while all children produced more plural errors in the test than in spontaneous speech because the test corresponds to a formal examination requiring much more metalinguistic awareness, this effect was particularly strong in the most disadvantaged group of L2 children (see also Korecky-Kröll et al. 2018).

In the test, but not in spontaneous speech, a significant increase in correct plurals during the investigation period was found for all children, whereas in spontaneous speech, correct plurals were already very frequent even at the first datapoint. While both groups of bilingual children showed a particular preference for morphotactically transparent forms in the test, SES (in interaction with datapoint) and productivity were more significant in the monolingual children. SES was also a significant factor in correct plurals in the spontaneous speech of monolingual children, but here, morphotactic transparency was more important for the monolingual group, whereas the correct plurals in the spontaneous speech of 2L1 children showed some evidence for a preference for total naturalness.

Incorrect zero plurals generally point to less advanced plural development. In the test, they may be interpreted as repetitions of the given singular stimulus forms, and they are indeed more frequently produced by LSES children (particularly by 1L1 and L2 children, but there is also an interaction with datapoint in 2L1 children indicating less progress in this bilingual group). High numbers of incorrect zero plurals are also associated with low cumulative teacher plural type frequency in monolingual children. The linguistic effects on incorrect zero plurals are rather inconclusive and point to a sort of emergency default status of zero plurals: when no other plural form is available for a child, they tend to produce the singular form, particularly in the test, whereas such forms are rare in spontaneous speech. L2 children with L1 Turkish are especially disadvantaged in the test, as their L1 requires a singular form in contexts following plural quantifiers, but their L2 German requires a plural target form—thus, they produce particularly high numbers of incorrect zero plurals in the test (see also Appendix B, Table A2). This is apparently an example of negative transfer from their L1 to their L2.

Overt overgeneralizations are considered the interesting low point in the middle of the frequently observed u-shaped curve from rote learning to the mastery of correct form: in the test, 1L1 children produce more iconic forms, whereas 2L1 children use fewer productive target forms, and L2 children show a complex interplay of SES, kindergarten exposure, and total naturalness of forms, which may be interpreted in terms of different timings within the u-shaped curve: while the more advanced HSES L2 children start using more natural plural overgeneralizations earlier and show a later decrease in favor of correct forms due to their increasing kindergarten exposure, their LSES peers show an increase in overgeneralizations only at a later point (i.e., with increasing kindergarten exposure). In spontaneous speech, overt plural overgeneralizations are nearly as rare as incorrect zero plurals, particularly in 2L1 children. Nevertheless, the 1L1 and L2 children show a dispreference for productive forms and a tendency to avoid particularly little reliable, little iconic and little natural target forms. Some evidence of transfer (or mutual influence of both languages in terms of a preference for +e suffixation) is found for the 2L1 children: it has a positive impact in the sense that correct +e plurals emerge relatively early, but also a negative impact that leads to a considerable number of +e overgeneralizations in the Croatian–German speaking children (see also Camber 2020; Camber and Dressler 2022). A similar tendency has been observed by Köpcke and Wecker (2017, p. 95) for nine-to-ten-year-old German L2 learners with L1 Turkish participating in a non-word experiment of German plurals: the high similarity of Turkish -ler to German -er may have motivated these older children to produce more -er plurals than the comparison groups with L1 Russian and L1 German. However, this transfer was only unidirectional from L1 to L2.

Overall, the procedure had the strongest impact on the results: all children produced more errors in the metalinguistically challenging test compared to spontaneous speech. However, SES and cumulative teacher plural type frequency were most significant in the 1L1 children, whereas kindergarten exposure turned out to be more important for the L2 children, and 2L1 children were more sensitive to the linguistic variables.

This study was innovative in the sense that it tested the predictions of two different (although not completely opposite) linguistic theories and their most important parameters based on data from three different groups of children acquiring three typologically different languages collected in two settings, which allowed the comparison of the impact of many different variables.

However, the study has also certain limitations: for example, the following problems should be tackled in future studies: (1) the lower number of children in the 2L1 group compared to the two other groups, (2) the languages that were spoken by only one group of bilingual children (with the consequence that no comparison of 2L1 and L2 Turkish or 2L1 Croatian and L2 Croatian was possible), or (3) the lack of information on home language exposure and frequencies of plural types and tokens of parental caregivers. In addition, the focus on the morphological plural forms might also be supplemented with additional morphosyntactic analyses of entire noun phrases to be able to identify the potential effects of number disambiguation within the sentence (as shown by Nickel and Werth 2022 for adult dialect speakers).

Nevertheless, the predictions of the SM and of NM were largely confirmed by our results. All linguistic variables showed significant effects in some analyses, but morphotactic transparency turned out to be the most relevant variable for all three groups of children. Overall, these results may be useful for educators and other professionals who are involved in developing language tests or support programs for diverse groups of children who acquire typologically different languages in different acquisition contexts.