2.1 Gender assignment and agreement in German

German has a three-way gender system with masculine, feminine and neuter. Gender is marked on determiners (definite and indefinite articles, demonstratives, possessives), attributive adjectives and relative pronouns. Masculine is the most frequent and neuter the least frequent gender, although the exact numbers vary depending on the type of corpus (e.g., spoken vs. written, child-directed vs. adult-directed, basic vocabulary vs. advanced vocabulary). Schiller & Caramazza (2003) report 43% masculine, 38% feminine and 19% neuter gender, based on lemmas in the CELEX database (4164 monomorphemic nouns). For child speech, Binanzer (2017) reports 42% masculine, 31% feminine and 27% neuter; for basic vocabulary Wegener (1995) finds 44% masculine, 33% feminine and 23% neuter nouns. In many accounts on German, masculine is considered to be the default for gender assignment in German. The reason is not only its comparatively higher frequency –masculine is only slightly more frequent than feminine– but rather the assigned gender when assignment rules are in competition or when no specific gender assignment rules apply (e.g., Steinmetz 2006). On the other hand, Opitz & Pechmann (2016) have provided evidence that masculine gender requires increased processing load, which would speak against its default status. According to Köpcke & Zubin (2009, pp. 148–150), it is in fact possible to construct an argument for all three genders as the default in German, depending on theoretical position. Based on examples of headless NPs such as das Schöne ‘the beautiful’ and reference to unknown objects (Was ist denn das? ‘What’s that?‘), they claim that Neuter gender could be considered the ‘system default’, though admitting that such examples could also be genderfree. Corbett & Fraser (2000: 69f) argue that such examples could be construed as showing the syntactic gender default in German. The syntactic default is generally assumed to be Neuter, based on presentative or expletive constructions (das ist meine Mutter ‘that is my mother’, es scheint … ‘it seems …’). Thus, there is no general consensus about the notion of default gender in German (see also Enger (2009) for a critical discussion of defaults, and Corteen (2019) for a discussion of defaults within Optimality Theory).

The most frequent elements on which gender marking is visible are articles, which are obligatory with singular count nouns. Gender interacts with number, definiteness and case, and there is considerable syncretism, which makes structural cues to nominal gender somewhat opaque. For example, gender is unambiguously marked on definite singular DPs in the nominative case, although the adjective is underspecified with respect to gender. In indefinite DPs in the nominative case, there is syncretism between neuter and masculine articles, but the adjective is unambiguously marked for gender (see Table 1). In the plural, there is no gender marking.

Besides structural cues to nominal gender (i.e., elements in syntactic agreement with the noun), there are morphological, phonological and semantic cues on the noun itself (Köpcke 1982; Köpcke & Zubin 1983). German generally follows the natural gender rule, in that nouns referring to male beings are masculine, and nouns referring to female beings are feminine.² Suffixes can be associated with particular genders, e.g., –chen and –lein are associated with neuter gender, even if there is a clash with natural gender, as illustrated by the examples Mädchen_N ‘girl’ and Männlein_N ‘little man’. There are numerous phonological cues to gender assignment, but many of them reflect probabilistic tendencies rather than rules. Amongst the most reliable ones for each gender are the following: Disyllabic nouns ending in [ə] are associated with feminine gender 90% of the time (e.g., Katze_f ‘cat’), monosyllabic nouns starting with the onset cluster /ʃ/+C_ are associated with masculine gender 81% of the time (e.g., Strand_M ‘beach’), and nouns ending in [ɛt] are associated with neuter 95% of the time (e.g., Bett_N ‘bed’) (see, e.g., Köpcke 1982).

2.2 The acquisition of gender by German monolingual children

Gender is most frequently marked on articles, which is why first gender marking is closely related to first article use. Mills (1986: 63) reports first uses of definite articles in spontaneous child data around age 1;10–2;2, and Stern & Stern (1928: 44, 86) around 1;2–1;10. Articles are typically not gender marked at this point but reduced or underspecified forms, e.g., de. Such forms are often considered to be proto-articles whose function is to mark a syntactic position, while the acquisition of morphology lags behind (e.g., Bottari et al. 1993/1994). Sporadic use of gender-marked articles is reported from age 2;0–2;4 (Mills 1986), 1;9 (Bittner 2006), and 1;5 (Szagun 2007), though with substantial individual variation. acquisition tends to be defined as the point when children cease to make gender-marking mistakes (90% accuracy). This happens around age 3;0 (Mills 1986; Szagun et al. 2007).

Based on a single case study, Bittner (2006) argued that children’s early articles do not have a gender-marking function at all, but are used to mark syntactic function instead, with der referring to subjects and den to objects. Der in nominative position constitutes more than 90% of the forms used before age 2;4 regardless of the noun’s gender; die is overused between 2;5 and 2;11, and das is hardly ever overused. According to Bittner the gender system only emerges at 2;11, but the dissociation between masculine and neuter only by the age of 4. However, since her data are primarily presented in terms of case and only non-target uses are quantified, the study leaves open what the rate of correct gender marking was. Further, the early acquisition of case in German is not generally confirmed (e.g., Tracy 1986).

Based on longitudinal data, Mills (1986: 64) observed that the feminine article eine was the most common early form, also with nouns of masculine or neuter gender. When indefinite articles became frequent, ein and eine fluctuated, sometimes in combination with the same noun. Mills speculated that the incorrect use of the indefinite article may be due to “the high frequency of the ending –e in the article and adjective paradigms […] leading to an overgeneralization of the –e ending on all prenominal units.” Definites appeared later than indefinites. While Mills observed an association of monosyllabic nouns with der (p. 83) and of the schwa-ending with die (p. 85), errors suggested that not all assignment was cue based. Errors involved overuse of die, even with nouns containing masculine cues, e.g., Bagger (-er is a cue for masculine) and Hund (monosyllabic nouns tend to be masculine). Thus, overuse of die does not result from early acquisition of the feminine cue alone. Article omission was still frequent between the ages of two and three years. Since the children produced far more definite articles with feminine nouns than with nouns of the other two genders, Mills concluded that children are more confident with the feminine gender. Her account is largely in line with Stern & Stern (1928), who reported a distinction between masculine and feminine forms by 2;6 as well as gender errors resulting from overuse of die. Mills (1986) also investigated older children (ages 3;2–6;3) in an experimental situation, and results showed few incorrect productions of articles (5%), with none of the 6-year-olds producing any errors. There were comparatively fewer errors with feminine, and when errors were made with masculine and feminine, children rarely resorted to neuter. Feminine and neuter were rarely confused with one another.

Overall, the literature review indicates considerable individual variation, but some overall tendencies: (i) despite the challenges posed by the system, children’s development of gender in German appears to be comparatively fast (mostly before age 3), (ii) feminine articles emerge early and neuter articles late, (iii) there is some sensitivity to phonological noun-internal cues, and (iv) article omission or production of underspecified articles may be considered children’s strategies to avoid morphological markings of gender and case.

2.3 The acquisition of gender by German bilingual children

Studies on the acquisition of German by early bilinguals can be distinguished in terms of whether German is acquired simultaneously as an L1 or sequentially as an early L2. Another relevant distinction is whether German is acquired along with a language that has gender and, if so, how transparent gender marking in that language is.

In Müller’s (1999) study of three German-French simultaneous bilingual children, gender was marked from around 1;10–2;4. Initially, the children mostly omitted determiners, except for French un and German ein, which seemed to be used as numerals and irrespectively of the gender of the noun they preceded. Subsequently, they started using determiners productively and assigned gender based on phonological and semantic criteria. Gender assignment with definite articles was unproblematic, except for neuter, which occurred only from around age 2;4–2;8. Assignment and overgeneralization errors were determined by phonological and semantic properties. The children used feminine definite articles with nouns ending in schwa, even if these nouns were masculine, i.e., overusing the schwa-regularity. For instance, the children produced die_F Hase_M ‘rabbit’ or die_F Affe_M ‘monkey’; alternatively they used the correct gender but adjusted the noun ending to make assignment more rule-conform, as in der *Aff ‘the monkey’ (monosyllabic nouns being associated with masculine). One of the children also overused der with monosyllabic nouns (der_M Haus_N ‘house’, der_M Geld_N ‘money’, der_M Bett_N ‘bed’), indicating sensitivity to the tendency for monosyllabic nouns to be masculine. Neuter remained a problem until the age of 5, and neuter nouns were initially classified as masculine (Müller 1999: 393).

Eichler, Jansen & Müller (2013) investigated gender marking in naturalistic data from 17 simultaneous bilingual children acquiring a Romance language (French, Italian or Spanish) along with German or another Romance language. They compared gender errors across all languages. The children were followed longitudinally between 2 and 3;6 years and gender marking was investigated based on article production. The bilingual children were more successful in their Romance languages than in German (and children acquiring Italian and Spanish had advantages over children acquiring French), although most of the children grew up in Germany. Knowledge of cues was not explicitly discussed, but the children were aware of the schwa-rule, as evident from overgeneralizations such as die_F Junge_M ‘the boy’ (p. 565). Neuter was acquired late.

Also based on naturalistic data, Wegener (1995) studied the acquisition of gender by L1-Turkish, Polish, and Russian children acquiring German as an early L2. The children started out using bare nominals; their first articles had the function of marking indefinite (nonspecific) as opposed to definite reference. Ein was the most frequent indefinite article and die the most frequent definite one, arguably due to the high token frequency of these forms in the input. Subsequently, the children used articles for case-marking or argument-marking functions: r-forms for subjects, s-forms for direct objects and e-forms for plurals, somewhat similar to what is claimed by Bittner (2006) for monolingual German. The children eventually used articles successfully, according to Wegener primarily due to lexical learning (p. 15), but they also showed sensitivity to formal rules. For example, after only 13–15 months of exposure the L1-Polish and L1-Russian children assigned masculine gender to monosyllabic nouns and nouns in –er and feminine to nouns in schwa (pp. 16–17). The L1-Turkish children in grade 4 performed at chance, possibly due to the absence of gender in Turkish.

Dieser (2009) presented longitudinal naturalistic data (ages 1–6;6) and cross-sectional semi-naturalistic and experimental data (ages 3–10) from German-Russian children. She investigated both languages and compared simultaneous and sequential bilinguals. All learners, independently of type of bilingualism, passed through similar acquisition sequences in both languages, starting with a two-way gender system, feminine and masculine for Russian, feminine and non-feminine for German (see also Hopp 2011). Gender differentiation started with indefinite articles, before being extended to definite articles. Dieser concluded that monolinguals and bilinguals make use of lexical and rule-based learning in both languages, with a higher amount of rule-based learning in Russian. The German rules included the schwa-rule.

Ruberg (2013) compared German monolingual children (ages 3–5) and sequential bilingual children with different heritage languages (Polish, Turkish, Russian; ages 4;0–6;3). Based on real and nonce words, his results show that by the age of 4;0 monolingual children use correct definite and indefinite articles. The sequential bilingual children did not show target-like gender marking after 30 months of contact with German. When making assignment errors, both the monolingual and the sequential bilingual children defaulted to masculine, in line with Steinmetz’s (2006) claim that masculine is the default gender in German. Ruberg’s results further showed that both monolingual and bilingual children made use of morphophonological cues, assigning gender more often correctly if the nouns ended in schwa (associated with feminine) or were monosyllabic (associated with masculine). When the noun did not exhibit a prominent gender cue, children mainly used the default strategy. As in Wegener’s study, the L1 Turkish children did not attend to cues to the same extent as the two other bilingual groups.

Finally, Lemmerth & Hopp (2018) investigated lexical and syntactic CLI in simultaneous and sequential bilingual German-Russian children (mean age 8), focusing on German. Elicited production data showed that the children assigned gender correctly independently of whether the nouns have congruent or different genders in the children’s two languages. By means of eye-tracking, the authors further investigated whether the children used articles and adjectives to anticipate an upcoming noun. The simultaneous bilingual children behaved like monolingual controls, while the sequential bilinguals made use of gender marking only when the gender of the noun was equivalent in the two languages. Lemmerth & Hopp (2018) speculated that sequential bilinguals assign gender in the L2 by first accessing the lexicon of their L1, while there is no such influence in simultaneous bilinguals. This latter claim raises the interesting question on what basis the children drew a parallel between feminine, masculine and neuter gender in German and Russian. Unless the children had been taught German and Russian grammar explicitly, we believe that there is no reason for them to match specific genders one-to-one in the two languages.

In summary, although there is substantial variation in the acquisition of German gender by bilinguals, there are many indications that bilingual children show sensitivity to the same cues as monolingual children. The variation seems to be determined by the language combination, specifically the presence/absence and transparency of gender marking-rules in the other language, but it remains unclear what exactly can be transferred across languages. Table 2 summarizes some relevant studies on German.

2.4 Grammatical gender in Russian and its acquisition

As mentioned in the Introduction, gender in Russian is mostly transparent (Corbett 1982; 1991), with three rules that predict gender assignment to the majority of nouns in the language: In the nominative singular, masculine nouns typically end in a consonant (e.g., dom_M ‘house’), most feminine nouns end in -a (e.g., lisá_F ‘fox’), and most neuter nouns end in -o (e.g., vedró_N ‘bucket’). There are also some phonological gender cues that are ambiguous, most notably affecting nouns that end in a palatalized consonant, which may be either masculine or feminine (e.g., gus’_M ‘goose’ or kost’_F ‘bone’), as well as nouns ending in schwa, which may be either feminine or neuter (e.g., parta_F [partə] ‘desk’ or sito_N [sitə] ‘sieve’). The majority of nouns ending in a schwa are feminine, leading to early overgeneralizations of schwa-final neuters to feminine (Ceitlin 2009; Mitrofanova et al. 2018). Russian has no articles, but gender is marked as agreement on adjectives, demonstratives, quantifiers, some numerals and verbs in the past tense. According to Corbett (1991), masculine and feminine genders are almost equally frequent (46% and 41% respectively), with masculine being considered the default for gender assignment, while neuter is by far less frequent (13%).

Table 3 summarizes (morpho-)phonological gender cues in Russian, comparing them to those we introduced above for German. We remind the reader that there are more possible noun endings in both German and Russian, which are often predictive of gender, but we restrict ourselves to those relevant for the present study.

The current study is a continuation of similar studies on gender assignment in bilingual populations acquiring two gender languages that differ with respect to transparency. Rodina & Westergaard (2017) elicited real words to investigate the gender system of Norwegian-Russian bilinguals and found that transparency plays a major role, as gender was acquired faster in Russian than in Norwegian, even though Russian was the children’s heritage language. There was no difference between the bilinguals and monolingual controls in the majority language (Norwegian), but the children’s performance in Russian was dependent on the amount of input. Unlike Polinsky (2008), who found that some (adult) heritage speakers of Russian in the US were developing a reduced gender system with only masculine and feminine, Rodina & Westergaard found that the children generally had a three-gender system and that only the children with the least amount of input defaulted to masculine, thus seemingly developing a Russian variety without gender. Mitrofanova, Rodina, Urek & Westergaard’s (2018) follow-up study used nonce words and showed that the heritage children were generally sensitive to the transparent phonological cues in Russian, and defaulting to masculine decraesed with a higher Cumulative Length of Exposure to Russian.

The German-Russian children in the current study have also been investigated in their heritage language Russian and compared with monolingual controls (Mitrofanova et al. 2021), with the same three experiments as in the present study: real words, nonce words, and nonce words with mismatched cues (i.e., noun endings and adjectival morphology cueing for different genders). The findings show sensitivity to both transparent and ambiguous cues in Russian for both monolingual and bilingual children, but bilingual children used the masculine default to a considerably larger extent. For the ambiguous cues, all children displayed a preference for feminine with nouns ending in a schwa and sensitivity to more fine-grained cues for the nouns ending in palatals, such as the type of palatal consonant. In the mismatch experiment, the importance of the syntactic (agreement) cue increased significantly with age, while that of the noun-internal phonological cue decreased, echoing findings from French by Karmiloff-Smith (1979).

4.1 Participants

53 bilingual children (ages 3–11, mean 6,2) were recruited from a Russian language school in Singen, Hohentwiel (South Germany), a Russian daycare centre in Stuttgart, and four in Berlin. All children were heritage speakers of Russian growing up in Germany. These children spoke no other languages than German and Russian. 18 children came from German-Russian families (ages 3–11; mean 6.22; sd = 2.42), i.e., families in which one parent spoke Russian and the other German (henceforth ‘mixed’ families) and 35 from families in which both parents spoke Russian (ages 4–10, mean 6,22; sd = 1.92). Family type coincides with type of bilingualism, i.e., children from mixed families were simultaneous bilinguals and children from only Russian-speaking families were sequential bilinguals. Age of onset to German varied (range 0–6 years, mean 2.1). For most children first exposure to German was at the age of 3 years when entering kindergarten.³ Additionally, we tested 24 monolingual German children (ages 3–9; mean 4.92; sd = 1.43).

4.3 Analysis

For a first overview, we classified all utterances in terms of whether the elicited gender target (definite or indefinite article or adjective) was realized or not. The gender targets were indefinite articles and adjectives (indefinite condition) or a definite article (definite condition). If the elicited target was provided by the child, the realization was coded as M, F or N. If the child did not provide the elicited gender target, we distinguished replacements of indefinite by definite articles and vice versa, article omissions and non-classifiable responses. The responses were not classifiable if the noun was changed or if no gender target was provided. An overview of the responses for the monolingual and bilingual groups is provided in Table 4. Note that masculine and neuter responses cannot be distinguished in the indefinite condition due to syncretism.

For further analysis, gender assignment was coded based on the definite article, but some adjustments were necessary because the children did not always produce the elicited target.

• If no article was produced, the adjective in the indefinite condition was used as an indicator of assigned gender, but only if it was clearly marked as masculine (e.g., roter) or as neuter (e.g., rotes). Potentially feminine marked adjectives (e.g., rote) were not considered, because the form can represent all genders in definite DPs (cf. Table 1). Further, Table 2 suggests that the bilingual children still have problems with definiteness markings.⁴ Note that this procedure potentially biased the analysis against feminine marked DPs.

• If neither an article (or other determiner) nor an adjective marked as neuter or masculine was produced, we refrained from coding any assigned gender.

• Responses in which the children used a different noun than the intended one were excluded, unless the noun contained the same gender assignment cue as the target word. For instance, Flagge_F instead of Fahne_F ‘flag’ was included because both contain the schwa-cue but Bodensee_M ‘Lake of Constance’ instead of Strand_M ‘beach’ was excluded.

• If children could not name the depicted item themselves, the answer was not included. The rationale for this is that an unknown word is like a nonce word for the child, and we wanted to restrict the results of experiment 1 to words that the children knew.

• We included non-target-like productions in which the children added an ‘s’ or ‘r’ to the adjective to make the adjective rhyme with the definite determiner in the neuter and masculine conditions, e.g., das_N *rote-s Bett_N (‘the red bed’), der_M *grosse-r Schrank_M (‘the big closet’). Although the adjective form is not target-like, it can be taken to indicate the assigned gender.

• We coded inconsistently marked gender across the three elicited responses and agreement mismatches within the DP, e.g., eine_F grosses_N Bett_N (‘a big bed’). In both cases, it is not clear which gender the children have assigned, and these responses were not included when determining which gender the children had assigned.

Finally, we excluded data from nine children who completed less than 50% of the task or provided responses without determiners or any gender-marked adjective more than 50% of the time, two from mixed families and several from Russian families (see Table 3).

4.4 Results

The remaining dataset consisted of 44 bilingual German-Russian children, 16 from German-Russian families (age 3–11; mean 6,62; sd = 2.40) and 28 from Russian families (ages 4–10, mean 6.51; sd = 1,78). Table 5 shows that the children from Russian families omitted determiners most often. Gender switches across DPs were infrequent in all groups.

As a next step, we excluded children who defaulted to one particular gender. Defaulting was defined as using one gender 75% to 100% of the time. There were no monolingual defaulters, but among the bilinguals five defaulted to feminine and one to neuter. Discarding defaulters left us with 23 monolinguals and 38 bilinguals (23 from RR families and 15 from mixed families). The results of the Real word task are illustrated in Figure 2 for the feminine condition (left), the masculine condition (middle) and the neuter condition (right) for each of the three groups of children. Arrows indicate where we find significant differences.

Figure 2

Gender assignment with real nouns in German: monolingual children (GG), children from German-Russian (GR) and children from Russian-Russian (RR) families. Significant contrasts between the groups are marked by arrows.

We applied generalized linear mixed effects logistic regression analysis to statistically model the probability of an accurate response on the real word task depending on the three main predictors (Family type, Condition and Age) and their interactions. Model comparison (Anova) indicated that Age (continuous) as a main effect and the interaction of Family and Age could be dropped. The final model included the main effects of Family type and Condition, their interaction, as well as an interaction of Age and Condition, and a three-way interaction of Condition, Family and Age. Participants and Items were included as random intercepts. All models were fit using R version 4.0.3 (release 2020-10-10). The analysis revealed significant interactions of Age and Condition (masculine p = .03, and neuter p = .02), indicating that all groups of participants become significantly more accurate with age with respect to masculine and neuter gender as compared to the feminine. The output of the final model is summarized in Table A1 in the Appendix. For a more detailed comparison of the groups within conditions, we conducted post-hoc pairwise comparisons using the R package emmeans (Lenth et al. 2019). These revealed that the monolingual children performed significantly better than the children from Russian (RR) and mixed families (GR) on the masculine (p < .001 in both cases) and neuter conditions (p = .003 and p < .001, respectively). No other contrasts were significant (see Tables A2 in Appendix 1).

Overall, the results indicate that children in all groups use all three genders productively (above 60% accuracy with all genders), although nouns in the feminine condition seem to be the least difficult.

5.1 Participants

The participants were the same as in Experiment 1, but more children took this task. The initial dataset consisted of 60 bilingual German-Russian children, 20 from German-Russian families (age 3–11; mean 6.67; sd = 2.47) and 40 from Russian families (age 4–10 years, mean 6.50; sd = 1.89). We tested the same 23 monolingual German children as in Experiment 1.

5.3 Analysis

Tables 6 provides an overview of the responses for the monolingual and bilingual groups. Again, masculine and neuter responses cannot be distinguished in the indefinite condition due to syncretism.

The coding of responses followed the same algorithm as in Experiment 1 (see above) with some adjustments for nonce nouns. We discarded the following types of items:

• Cases in which the child did not (re)produce the target word, unless the cue was kept constant (e.g., Klapett_(N) for Lapett_(N) was included, but Fretts_(?) for Fruxe_(F) was excluded);

• Instances in which it was impossible to determine the assigned gender, (i) because the gender was incongruent between the determiner and the noun, e.g., *eine_F blaues_N Plett, (ii) because the gender was underspecified, e.g., blau Plett (no article, no inflection on the adjective), (iii) because the children switched genders across the elicited DPs.

As before, we included only children who completed more than 50% of the task. The resulting dataset consisted of 23 monolingual German-speaking children, 44 bilingual German-Russian children: 18 from German-Russian families and 26 from Russian families.

5.4 Results

Similar to the Real word task, the groups differed significantly in the probability of providing an article-less response: children from Russian families did so 32% of the time, while children from German-Russian families did so only 15%. Again, the children rarely switched (only 4% of the time with no differences between children from the two family types). The monolinguals failed to provide an article or another gender-marked element only 6% of the time, while gender switches constituted 1% of the data. When coding the data, we proceeded as follows, in analogy to Experiment 1.

• The assigned gender was coded based on the definite article.

• If no article was produced in the definite condition, we used the indefinite condition as an indicator of the assigned gender: adjectives marked as masculine (e.g., roter) or neuter (e.g., rotes) were taken to be indicators of the assigned gender. Recall that in the absence of an article, feminine marked adjectives (e.g., rote) were not considered, because the form can represent all genders in definite DPs (cf. Table 1 and footnote 4).

• We included non-target-like productions in which the children added an ‘s’ or ‘r’ to the adjective to make the adjective rhyme with the definite determiner in the neuter and masculine conditions, e.g., das_N *rote-s Plett_N (‘the red bed’).

• If neither an article (or other determiner) nor an adjective marked as neuter or masculine was produced, we refrained from coding any assigned gender.

• We excluded cases in which gender was marked inconsistently across the three elicited responses and cases of agreement mismatches within the DP, e.g., eine_F grosses_N Plett_N (‘a big Plett’).

As a result of the first cleaning, we excluded cases where the child did not respond, produced a noun that was different from the prompted noun, a bare noun and a bare adjective, an incomprehensible response or one that couldn’t be classified as either M, F or N. The remaining data were classified into M, F and N responses following the procedure described in 5.3 above. Figure 3 illustrates the distribution of responses in the Nonce word experiment.

Figure 3

Gender assignment with nonce nouns in German: monolingual children (GG), children from German-Russian (GR) and children from Russian-Russian (RR) families. Significant contrasts between groups are marked by arrows.

To statistically model the probability of following the phonological gender cue in the Nonce word task, we fit a series of generalized linear mixed effects logistic regressions with different sets of predictors. The best model included three main effects: Condition, Family type, and Age (continuous), two two-way interactions (Condition and Family, and Condition and Age), and a three-way interaction of Condition, Family and Age. An Anova model comparison suggested that the interaction of Family and Age could be dropped. Participants and Items were included as random intercepts. The model revealed a significant effect of Family (RR: p = .007), Condition (Masculine: p = .008, Neuter: p < .001) and Age (p = .002), as well as a significant interaction of Condition (N) and Family (DR: p = .03; RR: p < .001), a significant interaction of Condition (N) and Age (p < .001) and four significant three-way interactions of Condition (F), Age and Family (DR: p = .004; RR: p < .001), and Condition (N), Age and Family (DR: p = .003; RR: p = .001). The output of the model is summarized in Table B1 in the Appendix.

Post-hoc pairwise comparisons of groups within conditions revealed three significant differences. First, the GR and RR children were significantly more likely than the GG children to use Feminine gender in the Feminine condition (p < .001). GR children were marginally but not significantly different from the RR children (p = .07). Second, RR children were significantly more likely to use Masculine in the Masculine condition than GG children (p = .01). GR children were marginally different from the GG children (p = .07). Third, GG children were significantly more likely to use Neuter in the Neuter condition than both the GR and the RR children (p < .001 in both cases). No other differences were significant. The results of the post-hoc comparisons are summarized in Table B2 in the Appendix.

It should be noted that many children chose one particular gender (almost) throughout the entire task. We had a closer look at these defaulters (classified as children who used one particular gender 75% of the time or more) and non-defaulters separately. Among the 23 monolinguals, 11 children used neuter 75% or more of the time, and one child (3 years old) used feminine 100% of the time. The bilingual data varies more in terms of default strategies. Among the 44 bilinguals, 13 defaulted to neuter and 8 to feminine. The latter are distributed equally over German-Russian and Russian families. Interestingly, defaulting correlates with age, as younger children default less to neuter (all 3-year-olds are non-neuter-defaulters, while most 6-year-olds and all 7-year-olds are neuter-defaulters). Table 7 provides an overview of defaulting strategies comparing mono- and bilinguals. The distribution of (non-)defaulters and their demographics is presented in Table 8, indicating that as many as half of the children in each group chose a default strategy in their responses. Interestingly, the vast majority of monolingual defaulters used Neuter gender across the board, while bilingual defaulters fall into two groups: children that used Feminine across the board and participants that defaulted to Neuter. There were no Masculine defaulters in the dataset.

Figure 4 presents a comparison of responses by condition in the monolinguals and the two groups of bilinguals (excluding defaulters).

Figure 4

Gender response per condition: Monolinguals (GR) vs. Bilinguals from German-Russian (GR) and Russian (RR) families (excluding defaulters). Significant contrasts between the groups are marked by arrows.

We used a similar analysis as described above to assess the effect of the main predictors (Family type, Condition and Age) on the probability of following the gender cue provided in the Nonce word for a subset of children who did not default to one gender throughout the experiment. The winning model included the same predictors as the winning model for the whole Nonce word dataset, i.e., the three main effects: Condition, Family type, and Age, two two-way interactions (Condition and Family, and Condition and Age), and a three-way interaction of Condition, Family and Age. Participants and Items were included as random intercepts. The model revealed a significant effect of Condition (Masculine: p = .001 and Neuter: p = .004), as well as a significant interaction of Condition (Neuter) and Age (p = .002) and two significant three-way interaction of Condition (Neuter), Age and Family (GR: p = .04 and RR: p = .02). The output of the model is summarized in Table B3 in the Appendix.

Post-hoc pairwise comparisons of groups within conditions revealed the following significant differences. First, the GR and RR children were significantly more likely than the GG children to use Feminine gender in the Feminine condition (p < .001 in both cases). Second, RR children were significantly more likely to use Masculine in the Masculine condition than the GG children (p = .03). Third, the GG and GR children were significantly more likely to use Neuter in the Neuter condition than the RR children (p = .005 and p = .03, respectively). No other differences were significant. The results of the post-hoc comparisons are summarized in Table B4 in the Appendix.

The results indicate that after the exclusion of the defaulters both bilingual groups exhibit a significantly higher sensitivity to the feminine gender cue than the monolinguals, and the RR children exhibit a higher sensitivity to the masculine cue. In the neuter condition, the monolinguals and children from mixed families produce more neuter responses than children from Russian families. The children from Russian families use more masculine in the neuter condition, and it is possible that they are delayed in separating the masculine vs. neuter categories, as has been reported for monolinguals in the literature (cf. section 2.2).

5.1 Participants

We tested 20 monolinguals (ages 3–9, mean 5) and 38 bilinguals (age 4–10, mean 7), 11 from GR families and 27 from RR families. The bilingual children were a subset of the children who participated in Experiments 1 and 2.

6.3 Analysis

We followed a similar cleaning procedure as in Experiments 1 and 2 and discarded responses where the child did not produce the target word or where it was impossible to determine the assigned gender due to incongruent gender markings or underspecified gender. Unlike in Experiment 2, where no structural gender cues were provided, we did not observe any strong defaulting patterns. Only one bilingual child used feminine 100% of the time, and there were no neuter or masculine defaulters. Five further children (all from Russian families) were excluded from the analysis, because they produced less than 50% codable responses. The resulting dataset included 20 German monolingual children, 11 children from German-Russian families and 21 children from Russian families. No significant differences were observed between the bilingual children from German-Russian and Russian families, which motivated analyzing their results together. Unlike in the other experiments, children switched genders more often between their two responses. Monolinguals switched genders 37% of the time, children from German-Russian families 25% of the time and children from Russian families 28% of the time. As much as 80% (72/89) of all switches in the monolingual group were from M or F to N (consistent with the defaulting strategy). Bilinguals switched to N less than 50% of the time. The higher number of switches in this experiment was expected because the children were provided with conflicting cues.

6.4 Results

When the cues on the two targets were consistent with those on the noun (match condition), both mono- and bilingual children mostly produced a definite DP with the same gender as the one provided in the prompt, in fact as much as 85% of the time. Figure 5 summarizes the proportions of responses where children used the same gender as the one provided by the experimenter in the Match condition.⁵

Figure 5

Proportion of responses following the prompted (structural) gender cues in the Match condition by age.

As evident from Figure 5, both groups of children exhibited high sensitivity to the prompted gender cues. At the same time, younger children (especially bilinguals) produced a higher proportion of responses that did not follow the gender cue provided by the experimenter.

The main research question of Experiment 3 was whether we would observe any effect of group or age in the probability of following either the structural cue (on the determiner and adjective) or the phonological gender cue (on the noun ending) provided in the Mismatch condition. Figures 6 and 7 illustrate the proportions of responses that followed the adjectival or the phonological cue by individual participants from the two groups.

Figure 6

Proportion of responses following the structural cue in the Mismatch condition by age.

Figure 7

Proportion of responses following the phonological cue in the Mismatch condition by age.

Figures 6 and 7 indicate that monolinguals have a higher probability of following the structural gender cue in their responses than the bilinguals. In both groups, the probability of following the structural (and not the phonological cue) increases with age, and this increase is more visible in the bilingual group.

To analyze the results statistically, we applied generalized linear mixed effects logistic regression analysis where the probability of following the structural cue or the phonological cue was modelled based on the effects of Group (Bilingual vs Monolingual), Condition (Match vs Mismatch) and Age (continuous), as well as their interactions. An Anova model comparison was applied to choose the best model for each dependent variable. The final model predicting the probability of following the structural cue included three main fixed effects, an interaction of Group and Age, and a three-way interaction of all predictors. Participants and Items were included as random intercepts. The model revealed a marginally significant effect of group (p = .05), and a significant effect of age (p < .001), indicating that both groups become more likely to follow the structural gender cue with age, and that the monolinguals are marginally more likely to follow this cue than the bilinguals. No other effects were significant. The output of the model is summarized in Table C1 in the Appendix. Post-hoc pairwise comparisons of groups within conditions indicated that the monolinguals were significantly different from the bilinguals in the mismatch condition (p = .02), while there were no differences in the match condition (see Table C2 in the Appendix).

The model that predicted the probability of following the phonological cue included three main fixed effects, two interactions: Group and Condition, and Group and Age, and a three-way interaction of all predictors. Participants and Items were included as random intercepts. The model revealed a significant effect of Group (p = .03), and a significant effect of Age (p < .001), indicating that both groups become less likely to follow the phonological gender cue with age, and that the bilinguals were generally more likely to follow the phonological gender cue than the monolinguals. Furthermore, the model revealed a significant interaction of Group and Condition (p = .007), and a significant three-way-interaction of Group, Condition and Age (p < .001). No other effects were significant. The output of the model is summarized in Table C3 in the Appendix. Post-hoc pairwise comparisons of groups within conditions indicated that the monolinguals were again significantly different from the bilinguals in the mismatch condition (p = .02), while there were no differences in the match condition.

7.1 Sensitivity to formal cues in German vs. lexical learning

Experiment 1 showed that with familiar nouns, mono- and bilingual children acquiring German assign gender successfully in all three gender categories. Both groups of children were most accurate with the feminine nouns and least accurate with the masculine nouns. Given that most neuter nouns in our experiment had a very low frequency too, this suggests that factors beyond the frequency of noun items are at play. We hypothesize that form-based regularities connected to the phonological endings of the nouns are implicated: children showed at-ceiling accuracy with the feminine schwa-final nouns, where cue reliability is the highest, followed by the neuter nouns ending in -ett. The reliability of the masculine cue was very low, and accordingly, the children’s accuracy on this condition was the lowest. There were quantitative differences between the monolingual and the bilingual children, which can be expected given that the bilinguals, especially those from Russian families, had had less exposure to German.

Despite a general sensitivity to cues, the comparison between Experiments 1 and 2 shows that both groups are considerably less successful when assigning gender to novel words where only a phonological cue on the noun is provided. This finding is in line with the proposal that there exist two routes of grammatical gender processing: the lexical route, which retrieves grammatical gender features of the noun directly from the lexicon, and a form-based route that relies on form-based properties of the grammatical gender classes (see Caffarra et al. 2014). While the gender features of real words are stored in the lexicon and are deterministic, the form-based cues only reflect probabilistic tendencies and may differ in their reliability. The results of our Nonce word task may further reflect the underlying differences in reliability between different types of form-based gender cues. The feminine gender cue (final -e) was by far the most reliable cue for the participants, followed by the neuter cue, and finally the masculine cue. This mirrors the results for the adults when we piloted the nonce words.

Overall, phonological gender cues seem to have a relatively low reliability in German. In both groups, many children resort to a default gender in Experiment 2, choosing one particular gender 75% of the time or more. At the same time, children who do not default show a certain sensitivity to the respective gender cues. Those monolingual children who do not default to neuter assign neuter to nouns with the neuter cue, use feminine gender most often when the phonological cue is feminine, and masculine gender when the phonological cue is masculine. This means that defaulting and sensitivity to cues and regularities do not exclude one another, as has also been found by Ruberg (2013; see Enger 2009 for similar observations in adult systems). The picture for the bilingual children is less clear because some children default to neuter and others to feminine. More specifically, children from Russian families with less German input are more likely to default to feminine than children from mixed families with more exposure to German (the latter, by defaulting to neuter, behave more akin to German monolinguals). The non-defaulters use all three genders, with feminine being most frequent in the feminine condition, masculine in the masculine condition, and neuter in the neuter condition; the division between masculine vs. neuter is less clear, as shown in many previous studies on German (e.g., Stern & Stern 1928; Mills 1986; Müller 1999; Dieser 2009; Hopp 2011).

Finally, a comparison of Experiment 2 and the matching cue condition in Experiment 3 shows that mono- and bilingual children are more successful at assigning gender if there is a structural cue on a gender target (article or adjective) in addition to a phonological cue on the noun. Indeed, the defaulting patterns observed in Experiment 2 were not replicated in Experiment 3, where the children received structural cues in addition to phonological cues. This may be taken to indicate that children are sensitive to the relevant cues, but that phonological cues in German are weaker than in other languages.

7.2 Differences between monolinguals and bilinguals

As outlined above, both mono- and bilingual children have been found to be sensitive to phonological and structural gender cues in German. Based on previous research (Kaltsa et al. 2019; Egger et al. 2018; van Dijk et al. 2021) we hypothesized that bilingual German-Russian children may experience ‘acceleration’ (or facilitation) in the acquisition of gender in German due CLI from their heritage language Russian. More specifically, bilinguals were expected to be more sensitive to gender regularities in German, due to the high transparency of gender assignment in Russian. This was indeed borne out, as the bilinguals performed better than the monolinguals on both the feminine and the masculine cue in the nonce word task, i.e., nouns ending in schwa and nouns ending in a consonant (Experiment 2, see Figure 4). However, they did not perform better than the monolinguals on the neuter cue (which also ends in a consonant). This could be because they treat all consonant-final nouns as belonging to the same noun class (as in Russian) and thus use more masculine gender in this context. Alternatively (or additionally), the high performance of the monolinguals on the neuter cue is not due to sensitivity to the cue, but rather the tendency of the monolinguals to default to neuter.

One major difference between monolinguals and bilinguals concerns their defaulting strategies (N in monolinguals and F or N in bilinguals). First of all, the defaulting patterns observed in our study are somewhat unexpected given previous claims about masculine being the default for gender assignment in German (e.g., Steinmetz 2006; Ruberg 2013) and neuter being late acquired (Müller 1999; Eichler et al. 2013). Interestingly, among the monolingual children, the neuter defaulters represent the older children. It thus seems as if the F and M cues on the nouns were to some degree available to younger children, but as they grow older, the phonological cues get overridden by a more powerful neuter default. There are three possible explanations for this: (i) children identify the nonce objects with things, and the word for thing (Ding) is neuter in German (furthermore, according to Zubin & Köpcke 1986, nouns with a ‘vague’ reference, e.g., Objekt ‘object’, Zeug ‘stuff’, and nouns denoting superordinates, e.g., Instrument ‘instrument’, Möbel ‘furniture’, are generally Neuter in German, although some of these nouns may not be frequent in colloquial and child speech), (ii) the children were primed by the relative pronoun and question word was, because it ends in –s, which is a typical ending for neuter gender targets,⁶ viz. the definite article das and inflected adjectives, such as rotes ‘red’, or (iii) children default to neuter gender because it is in fact the default gender in German (cf. Köpcke & Zubin 2009, see section 2.1). Alternatively, the children are affected by the syntactic default, a possibility which gains plausibility given the challenges pertaining to German gender: if children postulate phonological rules, they must discover that there are many exceptions to them, while there are no exceptions to syntactic rules of gender agreement. On the other hand, there is no evidence, so far, that syntactic defaults are transferred to function as morphological default. Clearly, more research is needed to investigate the issue of grammatical gender default in German. We leave this question for future research.

We further suspect that the bilinguals’ defaulting to feminine has different reasons than the monolinguals’ defaulting to neuter, and that it could be motivated phonologically. The feminine cue is the strongest for the bilinguals, since it is the category with the highest percentage of target-like responses. This makes sense because the phonological cue for feminine is not only present in noun endings, but also on articles and adjectives, e.g., eine kleine Kerze ‘a small candle’, and the adjective form reappears with all genders in definite DPs (der_M grosse X, die_F grosse X, das_N grosse X ‘the big X’). Thus, children hear determiners, adjectives and nouns ending in schwa very frequently (see also Mills 1986), and by using more feminine overall, children might prime themselves. That is, the more they hear these forms, the more they produce them. An alternative (or additional) explanation is the existence of a similar cue in Russian, viz. nouns ending in schwa, which is mainly feminine (cf. section 2.3). While we think it is unlikely that the bilinguals link feminine gender in Russian to feminine gender in German, they may simply be more sensitive to this phonological cue in German, knowing that a similar cue is important (and highly predictive) in their heritage language.

Another difference between monolinguals and bilinguals is that the bilingual children seem to be more sensitive to the masculine cue. This is particularly evident in the group of children from Russian families. Given that consonantal endings in Russian are associated with masculine gender, it is not unlikely that children have used their knowledge of Russian cues and applied it to German, where our masculine nouns ended in a consonant. In fact, some children do not seem to distinguish clearly between masculine and neuter, and this can be expected because the neuter nouns in Experiment 2 ended in a consonant too. Note that transfer of cues, as we have speculated here, would not imply that the bilingual children make a connection between masculine, feminine, and neuter gender in German and the corresponding genders in Russian. Rather, since we know that they are highly sensitive to the consonantal noun-ending cue in Russian, they may simply assume that ending in a consonant is a transparent gender cue also in German, and thus expect that all nouns in -C will belong to the same gender. This means that it is not important which gender this is. The bilingual children may select the M gender forms simply because they are more frequent than N in German. Under this scenario, the ‘transfer’ of a cue would be based on the phonological feature and unrelated to a lexical feature.

In conclusion, mono- and bilingual children are sensitive to cues, but monolingual children default to neuter, while bilingual children default to neuter or feminine. It is possible that this is merely an effect of exposure and age, since the bilingual children have had less cumulative exposure to German, and it might be that the feminine defaulters would have picked a different strategy, had they been older. Admittedly, our study covers a large age range and the bilingual group is heterogeneous in terms of the learning profiles in German, specifically age of onset and exposure, which makes further research with larger participant groups desirable. At the same time, we were interested in the developmental patterns implicated in the acquisition of gender in German, thus looking at a wider age range was necessary in our study.

7.3 Phonological vs. structural cues

The results of Experiment 3 show that all groups of children rely more on structural cues than on phonological cues in a situation of cue mismatch. For the bilingual children we observed a significant developmental trend, with their reliance on phonological cues significantly decreasing and reliance on syntactic cues significantly increasing with age, in line with previous findings from languages with more transparent phonological gender rules (Karmiloff-Smith 1979, Lidz & Gagliardi 2015). The reason why we see relatively little sensitivity to phonological cues (unlike what Karmiloff-Smith (1979) has shown for French) might indeed be that these are less reliable in German, and that children start paying attention to structural cues from an earlier age, i.e., once they discover that there are numerous exceptions to phonological regularities.