When teaching works and time helps: Noun modification in L2 English school children

3.1 The participants

The data in the present study have been elicited longitudinally for a period of 2 years from 95 L1 Spanish children learning English as primary school students in a bilingual school in the region of Castile and León (Spain). The school follows the CLIL (Content and Language Integrated Learning) methodology and so, other than the English as an L2 subject, two or three more subjects are also taught in English, depending on the school year level. As in Gris-Roca’s (2017) policy of inclusion, a combination of teaching methods is used: the communicative language teaching, the CLIL methodology in the content subjects, and task-based language teaching. As per the language background questionnaire filled out by the children’s parents, children have never been exposed to English in a naturalistic context, that is, they have not spent long periods abroad in an English-speaking country nor have been in contact with native speakers of the language out of the school context. Informed consent from parents was obtained and the school collaborated for the whole process of data collection. Ethical approval from the University of Valladolid was obtained before data collection as part of the UVALAL (University of Valladolid Language Acquisition Lab) research activities (protocol approval ref. PI 19-1461).

Table 1 provides information on the four experimental groups, comprising different participants each: the non-NN instructed groups (younger and older) and the NN instructed groups (younger and older).

The two instruction groups (i.e., non-NN instructed and NN instructed) differ in terms of whether they have followed or not the pedagogical intervention program (see Section 3.2).

In each instructed group, two subgroups were considered so that length of exposure could be addressed. In this case, length of exposure is reflected in the years that participants have been exposed to L2 English in a school context. That is, participants that have been exposed to English for 2 years constitute the younger groups and participants that have been exposed to English for 4 years are referred to as the older groups. As for the children’s age, the range goes from 7 to 10 years old.

As for the amount of exposure to L2 English in the school context, children received on average 6.5 hours per week (35 weeks in a normal academic year). This amounts to the following ranges, as in Table 1: while the younger groups have received between 341.25 and 568.75 hours of English input, the older groups have received between 796.25 and 1023.75. This counting takes into account that testing was implemented half-way through the academic year. That is, for instance, in the case of the younger groups, whose L2 exposure amounts to 341.25, 227.5 hours corresponding to the first year of exposure have been counted; and 113.75 hours corresponding to half of the second year of exposure have been added to those of the first year, because this is when they were tested. Out of school exposure was not controlled for, but none of the children were involved in English immersion contexts of any kind and, in this respect, the participants’ exposure to English is considered to be quite homogenous.

The four participant groups have been tested in two consecutive years (i.e., year 2 and year 3). The longitudinal nature of this study is captured in Table 2.

Table 2

Longitudinal study: length of exposure and NN instruction

This experimental design, as reflected in Table 2, allows us to actually target the role of instruction (i.e., non-NN instructed and NN instructed groups) and its interaction with length of exposure (i.e., younger and older groups; year 2 and year 3).

Regarding the proficiency of the participants, Cambridge Young Learners Exams of the corresponding age were used in order to set homogeneous groups. The results were analyzed by means of two independent sample t-tests with Welch correction for unequal variances. For the two younger learners’ groups, the test showed that there was no statistically significant difference in the proficiency scores between groups (t(42.832) = −1.71, p = 0.09, d = 0.4, 95% CI [−1.80, 0.12]). For the two older learners’ groups, no statistically significant difference was found either (t(42.414) = 0.07, p = 0.93, d = 0.02, 95% CI [−0.87, 0.94]). These results ensure that the groups are comparable in terms of age and length of exposure and that, within each age group, differences only appear in the type of instruction received. This addresses one of the drawbacks pointed out by Muñoz (2011), when referring to how typical studies that aim at measuring L2 ultimate attainment are carried out: “learners have not been matched in terms of experience with the L2, and so input across participants has not been comparable” (Muñoz 2011, 114). This drawback has been controlled in the present study: all the experimental participants have been carefully selected in order to have the most homogenous and similar groups possible, except for the length of exposure and the NN instruction variables. At the time of the last testing (year 3), and during the two previous years (years 1 and 2), all participants have been enrolled in the same school, taught by the same teacher and with the same textbooks for all the subjects taught in English. Therefore, there are reasonable grounds to assert that, as far as the school context is concerned, all the participants received the same amount of instructional input, so they are comparable: the age-matched groups (younger non-NN instructed and younger NN instructed, on the one hand; and older non-NN instructed and older NN instructed, on the other), and the instruction type groups (younger non-NN instructed and older non-NN instructed; and younger NN instructed and older NN instructed).

Therefore, combining the general exposure to L2 English (younger versus older groups, year 2 versus year 3) with the specific NN intervention program implemented for two years, or lack thereof (NN instructed versus non-NN instructed groups), the role played by length of exposure and by the specific instruction received could be explored and teased apart.

3.2 The pedagogical intervention program

The taxonomy of the two different instructed groups (i.e., the non-NN instructed and NN instructed groups) derives from the specific pedagogical intervention program used with half the participants during three academic years. As indicated before, in the present study we focus on the last two years of this longitudinal study (year 2 and year 3). This explicit intervention program was implemented targeting one of the structures under analysis, NN compounds, since the other target construction in this study (i.e., AN strings) is already part of the curriculum in Spanish schools. Thus, AN strings are addressed via a set of varied activities and exercises included in the textbooks as well as explicit interventions on the part of the teacher. This AN instruction was exactly the same for all the participants in the different instructed groups and more importantly, no connection was ever established between AN strings and NN compounds. The NN instruction program, undergone by the NN instructed groups only, comprises a set of varied in and out-classroom activities aimed at overtly teaching the main properties and uses of English NN compounds (paying special attention to word order and productivity). It included different interpretation and production activities, planned beforehand (e.g., creating, naming, drawing, and explaining novel NN compounds, systematic interpretation of every compound appearing in the sessions), that were carried out and integrated in the regular school lessons in all the subjects taught in English. In these activities, the differences between English and Spanish compounds were highlighted. Besides the word order difference with Spanish as the most noticeable cross-linguistic difference, NN compounds, as previously mentioned, are highly productive in English, but not so much in the participants’ L1 (Snyder 2001, Gómez Garzarán and Fernández Fuertes 2020). Due to these two differences, NN compounds thus present an additional interest from a cross-linguistic point of view as well as a challenge for these learners.

The NN instruction program can be considered highly systematic and intrusive enough as it permeated all the L2 English curriculum. It is not possible to precisely quantify the number of hours this intervention was used given the different nature of the activities involved. As mentioned above, some activities were systematic and repetitive such as commenting each and every NN compound appearing in the textbooks. To illustrate this, the younger participants have been exposed to English for 341.25 hours in year 2 (Table 1) and, during this time, whenever a compound appeared, an activity that focused on word order was used. Other activities, however, were integrated in the didactic units and, for instance, involved making up novel realities by using NN compounds, or using songs to interpret NN compounds, etc. In this case, the younger participants in year 3 (568.75 hours of general exposure to English) performed seven different activities of this type and each was used at least three times. In either case, this pedagogical intervention took place systematically during the three years of the longitudinal study in the school context. Additionally, the teacher in charge of implementing the NN instruction program is a co-author of the present study, and he is a university-trained linguist and teacher. Furthermore, during the longitudinal study, he was the coordinator of the CLIL program at the school. As he was also the teacher for the non-NN instruction groups, all this ensures that teaching practices were highly controlled for in the four participant groups.

Such explicit instruction in NN compounds is not normally part of the usual classroom practices in the English as an L2 course in Spanish schools (Fernández Fuertes et al. 2008, Gómez Garzarán and Fernández Fuertes 2020). This situation reflects a contrast when NN compounds are compared to AN strings. In spite of sharing a similar underlying syntactic structure as far as word order is concerned (Gómez Garzarán and Fernández Fuertes 2020), NN compounds are not typically part of the curriculum while AN strings are typically included in the usual teaching practices of the Spanish school context.

3.3 The experimental task

An untimed acceptability judgment task has been designed, which includes 12 NN compounds and 12 AN strings randomized with 10 fillers. Fillers were disregarded for the analyses (e.g., copulative sentences with noun or adjective subject complements). In year 3 of testing, four extra experimental items of each type were added to the task so that, given that participants were already familiar with the task, more data were collected. Thus, the task comprises a total of 34 items and 42 items in year 2 and year 3, respectively. The task was preceded by a set of practice items, half of them grammatical and half of them ungrammatical and none involving the experimental structures (i.e., spelling errors, determiner–noun word order errors, copulative sentences with adjectives and determiner–noun phrases). Thus, two data sets were gathered, one in year 2 and one in year 3, for the four experimental groups (i.e., NN instructed and non-NN instructed groups and the corresponding two age groups in each case). This procedure follows a common practice in previous instructed L2 acquisition studies where participant groups are tested at two different moments in time using the same task (e.g., García Mayo 2003, Ortega and Iberri-Shea 2005).

The main criterion used to select the components of the two-word target structures was that all the lexical items were part of the vocabulary the participants are frequently exposed to. NN compounds were mostly taken either from the textbooks used in the different subjects or built up from nouns well known by the participants, including a few novel NN compounds such as “pirate bike” (grammatical NN compound) or “*umbrella frog” (ungrammatical NN compound given the picture referent provided). In the case of the adjectives in the AN strings, half of the adjectives referred to size or emotions, and the rest to colors. Nouns were balanced in terms of animacy, too (i.e., [±animate]).

In the task, the target structures were part of a question, as in Figure 1 (i.e., Would it be ok to say in English “a brown dog”/“a pirate bike” to describe this picture?). “a pirate bike” would be an example of a novel NN compound.

Figure 1

Grammatical experimental items: AN string and NN compound.

Participants were asked to provide an answer using a four-option Likert scale, for which they were asked to act as teachers and mark the option they thought was more appropriate. Emoticons with evaluative words were used and later translated into a numerical scale for data processing (“perfecto/perfect = 4, casi bien/almost ok = 3, regular/so so = 2, mal/bad = 1”). With the practice items, participants were guided to use the options 4–3 for correct items, and the 2–1 for incorrect ones, considering that “perfect” is not the same as “almost ok,” and that “so so” does not correspond exactly to “bad.” When data were codified, the four-option Likert scale was reduced to a 2-value scale to obtain the correction rates; that is, options 1 and 2 were collapsed for judgments assigned to incorrect experimental structures and options 3 and 4 were so for correct experimental structures. Correction rates were then translated into proportions for the analyses and comparisons carried out.

Out of the total number of experimental items (24 in year 2 and 32 in year 3), half were grammatical and half ungrammatical. Ungrammaticality implied a word order violation, as in Figure 2.

Figure 2

Ungrammatical experimental items: AN string and NN compound.

The task was administered, as part of the regular classroom activities, via Power Point presentations, providing each participant with their respective answer sheet.

Data from a group of English monolingual children from an elementary school in Edmonton (Canada) have been used in order to determine the validity of the task, especially in the case of some novel NN compounds. Native participants performed nearly at ceiling in all conditions with an NN accuracy rate of 86.11% (SD = 0.074) and an AN accuracy rate of 95.15% (SD = 0.075). These high accuracy rates ensure the robustness of the task. Furthermore, and in order to check the reliability of the task, the response patterns from the native control group were examined by a Chi-squared goodness of fit test. The results for participants’ tendency toward grammatical and ungrammatical responses showed that the differences between observed and expected frequencies were not significant for these English monolinguals (X ² = 0.193, df = 1, p = 0.659, V = 0.024). Thus, it can be concluded that the participants were not biased toward accepting the sentences as grammatical or ungrammatical in this study.

3.4 Research question and hypotheses

The present work is a 2-year longitudinal study (year 2 and year 3, as in Table 2) involving participants classified into two main comparable groups that only differ in the instructional input received on NN structures. Each group comprises two subgroups that differ in the length of exposure to L2 English that they have received at school (older and younger, as in Table 2).

Thus, the two variables for this investigation are the amount of exposure to L2 English at school, as seen in the different age groups (younger and older) and in the two years of testing (year 2 and year 3), and the NN exposure received by half of the participants as part of a specific instruction program (NN instructed and non-NN instructed groups).

By analyzing the participants’ answers in the experimental task regarding the two structures discussed – NN compounds and AN strings – and taking into account the two variables mentioned above, the research question that guides this study is the following: what shapes the L2 English production of L1 Spanish children: length of exposure or explicit instruction? That is, in order to minimize the negative effect of cross-linguistic influence (i.e., transfer) from Spanish into English, is length of exposure enough? Cross-linguistic influence would be reflected in the transferring of Spanish word order into English: noun-modifier instead of modifier-noun (as in Gómez Garzarán and Fernández Fuertes 2020, please see the comparative account in Section 1). More specifically, transfer is expected to occur in both AN strings and NN compounds, as they share the same underlying structure. Differences between the two, with a higher degree of transfer in NN compounds, could be attributed to NN compounds being unproductive structures in Spanish and not typically part of the English curriculum as such.

If amount of L2 exposure could be measured in two ways (i.e., in terms of overall exposure to L2 English in the school context and in terms of NN exposure as part of a specific instruction program), then cross-linguistic influence could be affected by either or both. In particular, three hypotheses can be put forward.

Hypothesis 1 states that L2 exposure will account for the improvement of the participants when judging NN compounds and AN strings. This way, older students will outperform younger students; and data from year 3 will show higher accuracy rates when compared to those from year 2. In both cases, whether length of exposure is seen in terms of age (i.e., younger/older, between group comparison) or in terms of year of testing (i.e., year 2/year 3, within group comparison), this difference will equally be shown regardless of whether data come from the non-NN instructed group or the NN instructed group. Improvement will affect the two structure types (i.e., NN compounds and AN strings) in the same proportion.

Hypothesis 2 states that intervention through the NN instruction program will explain the L2 improvement and so the NN instructed groups will outperform the non-NN instructed groups. Difference across instruction groups will be seen in NN compounds as it was the target structure of the NN instruction program; and it will equally be shown regardless of age (i.e., younger/older, between group comparison).

Hypothesis 3 states that a combination of overall exposure and explicit NN instruction will create a hierarchy among the four participant groups, with NN instructed groups taking the lead, so that the older NN instructed group will perform the best and the younger NN instructed group will outperform both non-NN instructed groups. In this last scenario, the cooperation of length of exposure and NN instruction will bootstrap the participants’ L2 final attainment with a more native-like performance.

If hypothesis 1 is confirmed, a difference between the answers of year 2 and year 3 groups will appear, as in García Mayo (2003), Muñoz (2011), and Shojamanesh et al. (2018). This difference will be so regardless of whether participants belong to the NN instructed or the non-NN instructed groups. A hierarchy or progression from lower to higher accuracy rates will be seen as in (3). In this hierarchy, groups are ordered in terms of the amount of exposure (from the lowest to the highest):

If hypothesis 2 is confirmed and NN instruction is the key factor, the difference will be spotted between the NN instructed and the non-NN instructed groups, as in Fernández Fuertes et al. (2020) and Gómez Garzarán and Fernández Fuertes (2020). The hierarchy in (4) shows the accuracy rates expected for the NN compounds (from the lowest to the highest):

If hypothesis 3 is confirmed and a cooperation between length of exposure and NN instruction is behind the results, a hierarchy or progression from lower to higher accuracy rates will be seen as in (5). In this hierarchy, groups are ordered in terms of the relative amount of cross-linguistic influence (from the highest to the lowest):

By addressing what is behind cross-linguistic influence effects and the relation between length of exposure and explicit direct NN instruction, we want to contribute to the ongoing debate on the factors impacting L2 acquisition in the school context through a longitudinal experimental study.

4.1 Length of exposure: year 2 versus year 3

The results corresponding to testing times (year 2 versus year 3) appear in Figures 3–5. Overall, for year 2 the mean correction rate is lower than the mean correction rate for year 3 (year 2: M = 62.66, SD = 16.33; year 3: M = 73.65, SD = 17.79). As data are not normally distributed, robust tests have been used to compare these paired samples. Results from the Wilcoxon Signed-Rank Test show a significant difference between years 2 and 3 with a large effect size (V = 571.5, p < 0.001, r = 0.63, n = 95). That is, the longer the exposure, the better the performance, as in Burstall (1975).

Figure 3

Correct judgments at the two testing times.

Figure 4

Correct judgments at the two testing times: grammatical versus ungrammatical items.

Figure 5

Correct judgments at the two testing times: NN compounds versus AN strings.

Where overall accuracy rates are shown (Figure 3), the higher the length of exposure, the better the performance in both age groups. That is, participants tested in year 3 have higher accuracy rates (younger: M = 68.72, SD = 17.19; older: M = 78.68, SD = 17.14) than participants in year 2 (younger: M = 56.77, SD = 13.14; older: M = 68.88, SD = 17.19). This increase in accuracy suggests that overall exposure does have an impact on these speakers’ performance (as suggested in hypothesis 1).

A robust mixed analysis of variance (ANOVA) with trimmed means at 20% was conducted to explore the interaction and main effects of age, as a between-subjects factor, and testing times, as a within-subjects factor. The ANOVA results show a significant main effect for testing times with a large effect size (F _(156.81) = 43.04, p < 0.001, η _p ² = 0.36) and for age with a moderate effect size (F _(149.74) = 13.18, p < 0.001, η _p ² = 0.13). However, there is no interaction effect between age and testing times (F _(156.81) = 0.03, p = 0.86, η _p ² = 0.004). This points again to a clear effect of length of exposure, that is, year 3 participants outperform year 2 participants in both age groups. This is in line with previous works on the role of length of exposure (García Mayo 2003, Muñoz 2011, Shojamanesh et al. 2018).

When considering the correct judgments given to grammatical and ungrammatical experimental items, participants seem to have no problem with grammatical items with their performance being almost at ceiling. The mean correction rate for year 2 (M = 88.22, SD = 11.43) was compared with that for year 3 (M = 91.57, SD = 10.14), using the Wilcoxon Signed-Rank Test. The results showed a significant difference between years 2 and 3 but with a small effect size (V = 3522, p = 0.007, r = 0.26). Ungrammaticality, however, is more difficult for them to detect in both testing times (year 2: M = 36.98, SD = 32.38; year 3: M = 55.68, SD = 31.68). Results from the Wilcoxon Signed-Rank Test (V = 3025, p < 0.001, r = 0.57, n = 95) show a significant difference between years 2 and 3 with a large effect size. This reconfirms hypothesis 1 in the sense that year 3 participants outperform year 2 participants in the judgment of both grammatical experimental items and, especially so, ungrammatical experimental items.

Accuracy rates considering age show the same pattern (Table 3 and Figure 4). Grammatical items are correctly judged, with the lowest mean proportion corresponding to the older group in year 2 (87.55%). Ungrammatical items show lower accuracy rates, the highest mean proportion corresponding to the older group in year 3 (65.17%).

A robust mixed ANOVA with trimmed means at 20% was conducted in order to identify interaction and main effects of age (between-subjects factor), grammaticality condition (within-subjects factor) and testing times (within-subjects factor). The results show a significant main effect for grammaticality with a large effect size (F _(1,372) = 262.42, p < 0.001, η _p ² = 0.49), for testing times with a moderate effect size (F _(1,372) = 21.62, p = 0.001, η _p ² = 0.09), and for age with a moderate effect size (F _(1,372) = 21.91, p = 0.001, η _p ² = 0.09). An interaction effect between grammaticality condition and testing times was found (F _(1,372) = 10.93, p < 0.05, η _p ² = 0.06). Pairwise comparisons using Tukey HSD tests attested a significant difference between years 2 and 3 when it comes to judging ungrammatical experimental items (t = −3.034, p < 0.05). No such difference appears with grammatical items (t = 0.049, p > 0.05). That is, when age is taken into consideration, significant differences are only found with ungrammatical items. Therefore, length of exposure (as seen in the two testing times) seems to have an impact on how participants detect ungrammaticality, while grammaticality is more stable across the two years. Furthermore, the large effect size found in the grammaticality condition is also reflected when comparing within each testing time (year 2: t = 4.250, p < 0.001; year 3: t = 3.097, p = 0.011).

In this case, the pattern in Figure 4 matches the one in Figure 3: the higher the length of exposure, the higher the accuracy rate, both in the younger group and in the older group, and both considering grammatical and ungrammatical items. This suggests that the pattern observed in Figure 3 for all the experimental items together is to be mainly attributed to ungrammatical items. That is, it is the judgment of ungrammatical items what significantly improves from the first to the second testing time (i.e., from year 2 to year 3). Therefore, ungrammaticality seems to be a landmark for how the processing of noun premodification proceeds in the case of these L2 bilingual children and for these two target structures.

If correct judgment rates are separated for each of the two experimental structures, a similar rate of accuracy seems to be given to NN compounds and to AN strings. In the case of NN compounds, the mean correction rate for year 3 is higher than that for year 2 (year 2: M = 60.12. SD = 16.48; year 3: M = 72.47, SD = 17.49). Results from the Wilcoxon Signed-Rank Test show a significant difference between years 2 and 3 with a large effect size (V = 526.5, p < 0.001, r = 0.63, n = 95). In the case of AN strings, the correction rate increases from year 2 to year 3 (year 2: M = 65.19, SD = 19.42; year 3: M = 74.83, SD = 20.22). Results from the Wilcoxon Signed-Rank Test show a significant difference between years 2 and 3 with a moderate effect size (V = 697.5, p < 0.001, r = 0.42, n = 95).

Figure 5 shows these results in the two age groups.

A robust mixed ANOVA with trimmed means at 20% was applied on accuracy rates considering age (between-subjects factor), testing times (within-subjects factor), and structure types (within-subjects factor). Results show a significant main effect for age with a moderate effect size (F _(1,372) = 30.82, p < 0.001, η _p ² = 0.09), for testing times with a moderate effect size (F _(1,1387.86) = 30.31, p = 0.001, η _p ² = 0.08), and no significant main effect of structure type (F _(1,372) = 2.05, p = 0.16, η _p ² = 0.008). No significant interactions were found.

The lack of a difference between the two target structure types for the two age groups at the two testing times seems to suggest that both structures do share a common underlying representation (Cinque 2010, Gómez Garzarán and Fernández Fuertes 2020, Harris 1991a,b, Kayne 1995, Piera 1995, Snyder 2001).

4.2 Explicit instruction: NN instruction versus non-NN instruction

When data are classified in terms of the two NN instruction type groups, the results are as shown in Figures 6–8. Overall, for the NN instructed group the mean correction rate is higher than that of the non-NN instructed group (NN instructed: M = 74.58, SD = 18.35; non-NN instructed: M = 61.86, SD = 15.07). As data are not normally distributed, a non-parametric test has been used to compare these paired samples. Results from the Wilcoxon Signed-Rank Test show a significant difference between the NN instructed and the non-NN instructed groups with a large effect size (V = 11.95, p < 0.001, r = 0.63, n = 95). That is, the group that has received a specific pedagogical intervention program in NN compounds shows better results overall than the group that has received the traditional L2 instruction program. This lends support to hypothesis 2.

Figure 6

Correct judgments in the two NN instruction type groups: younger versus older.

Figure 7

Correct judgments in the two NN instructed groups: grammatical versus ungrammatical items.

Figure 8

Correct judgments in the two NN instructed groups: NN compounds versus AN strings.

In the case of overall judgments in the two age groups (Figure 6), the NN instructed groups perform better than the non-NN instructed groups (for the younger groups, NN instructed: M = 70.63, SD = 17.47 and non-NN instructed: M = 55.49, SD = 11.26; and for the older groups, NN instructed: M = 78.36, SD = 18.55 and non-NN instructed: M = 68.80, SD = 15.73).

A robust mixed ANOVA with trimmed means at 20% was applied on accuracy rates considering age (between-subjects factor) and NN instruction type groups (within-subjects factor). Results show a significant main effect of NN instruction type groups with a moderate effect size (F _(1,186) = 20.84, p = 0.001, η _p ² = 0.12) and a significant main effect of age with a moderate effect size (F _(1,186) = 17.06, p = 0.001, η _p ² = 0.10). No significant interaction was found between age and instruction group (F _(1,186) = 0.92, p = 0.34, η _p ² = 0.005). These results confirm that the NN instructed group achieved higher accuracy rates than the non-NN instructed group (t = −5.17, p < 0.001). The main effect of age shows that the older group always outperforms the younger group without taking instruction into account (t = −4.70, p < 0.001).

A similar pattern is seen in the case of the overall correct judgments when grammatical and ungrammatical items are separated. In the case of grammatical experimental items, the comparison between the correction rate for the NN instructed group (M = 90.34, SD = 10.86) and for the non-NN instructed group (M = 89.47, SD = 10.99), calculated with the Wilcoxon Signed-Rank Test, is found to be non-significant (V = 4188.5, p = 0.383, r = 0.06). For ungrammatical stimuli, however, the Wilcoxon-Rank Test revealed significantly lower accuracy rates (V = 2575, p < 0.001, r = 0.37) in the non-NN instructed group (M = 34.18, SD = 28.84) than in the NN instructed group (M = 58.74, SD = 33.10). These results confirm that instruction plays a salient role when judging ungrammatical items.

Accuracy rates considering age show the same pattern (Table 4 and Figure 7). Grammatical items are correctly judged, with the lowest mean proportion corresponding to the older non-NN instructed group (88.90%). Regarding ungrammatical items, accuracy rates are lower, with the highest accuracy rate corresponding to the older NN instructed group (65.90%).

The results from the robust mixed ANOVA show significant main effects for the three factors at alpha-level <0.001 (age and NN instruction groups as between-subjects factors and grammaticality condition as a within-subjects factor). In addition, significant interaction effects were found between NN instruction groups and the grammaticality condition (F _(1,186) = 18.47, p < 0.001, η _p ² = 0.06), and between age and the grammaticality condition (F _(1,186) = 20.12, p < 0.001, η _p ² = 0.05).

Grammatical items are judged correctly with high accuracy rates, with no differences between age groups (t = 0.02, p = 0.95) or instruction groups (t = −0.25, p = 0.58). Ungrammatical items are more correctly judged in the NN instructed groups than in the non-NN instructed groups (t = −7.28, p < 0.001), which suggests that it is NN instruction, rather than age, what plays a role. These results suggest that the specific pedagogical intervention program targeting NN compounds has a positive effect making the NN instructed group perform better than the non-NN instructed group.

When considering only ungrammatical items, an interesting observation can be made. Even if no significant differences appear between the older non-NN instructed group and the younger NN instructed group (non-significant interaction between age and instruction: F _(1,186) = 1.31, p = 0.25, η _p ² = 0.004), this last group obtains higher accuracy rates for ungrammatical items (older: M = 48.65, SD = 29.13; younger: M = 51.26, SD = 32.43). This is so in spite of their having had 1 year less of English NN instruction when compared to the participants in the older non-NN instructed group. Therefore, NN instruction seems to accelerate sensitivity to ungrammaticality.

The distribution of correct judgments across the two experimental structures is rather balanced. In the case of NN compounds, the mean correction rate for the NN instructed group is higher than that of the non-NN instructed group (NN instructed: M = 73.29, SD = 18.05; non-NN instructed: M = 59.45, SD = 15.26). Results from the Wilcoxon Signed-Rank Test show a significant difference between the two instruction groups with a moderate effect size (V = 2577.5, p < 0.001, r = 0.37). In the case of AN strings, the correction rate increases from the non-NN instructed to the NN instructed groups (NN instructed: M = 75.86, SD = 21.18; non-NN instructed: M = 64.28, SD = 17.84). Results from the Wilcoxon Signed-Rank Test show a significant difference between the two instruction groups with a small effect size (V = 3078, p < 0.001, r = 0.27).

In Figure 8, data corresponding to NN compounds and AN strings are presented separately by age groups.

The robust mixed ANOVA results confirm a significant main effect for the two between-subjects factors, instruction with a moderate effect size (F _(1,372) = 37.11, p < 0.001, η _p ² = .10) and age with a moderate-large effect size (F _(1,372) = 28.67, p < 0.001, η _p ² = 0.08). Non-significant main effects were found for the within-subject factor structure types (F _(1,372) = 28.67, p < 0.001, η _p ² = 0.08). No significant interaction effects were detected at alpha-level 0.05.

In fact, no significant differences are found between the accuracy rates corresponding to the two structures in any of the two NN instructed groups regardless of age. What makes groups differ is the total accuracy (Figure 8). This keeps on pointing to the connection between NN compounds and AN strings (Gómez Garzarán and Fernández Fuertes 2020) in that explicit instruction in NN compounds (i.e., the complex predicate structure) boosts structures having the same underlying representation (i.e., noun premodification either by another noun or by an adjective). That is, it is the NN instruction program what explains the difference between the NN instructed and non-NN instructed groups, with the difference between NN compounds and AN strings being constant throughout.

The analyses presented above show that both length of exposure (i.e., comparing year 2 versus year 3; and older versus younger groups) and the specific NN instruction program (i.e., comparing the NN instructed group versus the non-NN instructed group) have an impact on these child English L2 learners. That is, the longer they have been exposed to L2 English, the better their performance in the acceptability judgment test; and, if they have been exposed to the pedagogical intervention program targeting NN compounds, their performance is better when compared to those who have not been exposed to such program. This is so for both age groups (younger and older) and for both structure types (NN compounds and AN strings).

We further explore the relation between instruction type groups and testing times. A robust mixed ANOVA on accuracy rates revealed a statistically significant main effect for instruction with a large effect size (F _(151.53) = 17.43, p < 0.001, η _p ² = 0.14) and for testing time with a moderate-large effect size (F _(153.71) = 54.80, p < 0.001, η _p ² = 0.12). However, no significant interaction effect was found (F _(153.71) = 1.82, p = 0.18, η _p ² = 0.006). This means that the differences found between the NN instructed and the non-NN instructed groups are kept constant throughout the two testing times, even if both groups significantly improve their performance from testing time 2 (i.e., year 2) to testing time 3 (i.e., year 3).

When grammaticality condition is examined in the relation between instruction type groups and testing times, the robust mixed ANOVA results showed a statistically significant main effect for the grammaticality condition with a large effect size (F _(1,372) = 251.60, p < 0.001, η _p ² = 0.50) and for instruction (F _(1,372) = 27.41, p < 0.001, η _p ² = 0.11) and for testing times (F _(1,372) = 25.15, p < 0.001, η _p ² = 0.09), both with a moderate effect size. The grammaticality condition, in turn, generates a significant interaction effect with instruction (F _(1,372) = 24.16, p = 0.001, η _p ² = 0.08) and with testing times (F _(1,372) = 13.51, p < 0.001, η _p ² = 0.05) with a moderate effect size. There was no three-way significant interaction effect, which entails that ungrammatical items are challenging for the NN instruction group and specially so for the non-NN instruction group, but no differences appear between year 2 and year 3 in this respect. This suggests that the NN instruction program is more determinant than the length of exposure.

When distinguishing between NN compounds and AN strings, the robust mixed ANOVA results confirmed a significant main effect for instruction (F _(1,372) = 50.40, p < 0.001, η _p ² = 0.11) and testing times (F _(1,372) = 41.84, p < 0.001, η _p ² = 0.09) both with a moderate effect size. No significant main effect was found for structure type (F _(1,372) = 1.64, p = 0.20, η _p ² = 0.009) and no interaction effects at alpha-level < 0.05 appeared. Participants seem to be treating both structure types in the same way regardless of their respective productivity in their L1 (i.e., NN compounds are not productive in Spanish while AN strings are). This further points to both structures having a common underlying representation, a fact that seems to be behind these speakers’ L2 judgments.