2.1 Adult studies adopting miniature-language paradigms

The literature on L2 learning of morphosyntax has consistently shown that adults can acquire nonnative word orders (including form-meaning mappings) through relative brief exposure to an AL, although mixed results have been found for inflectional morphology (Boyd, Gottschalk & Goldberg, 2009; DeGraaf, 1997; DeKeyser, 1995; Friederici, Steinhauer & Pfeifer, 2002; Grey, Williams & Rebuschat, 2014; Morgan-Short, 2007; Morgan-Short, Sanz, Steinhauer & Ullman, 2010; Rebuschat & Williams, 2012; Rogers, Révész & Rebuschat, 2015; Williams & Kuribara, 2008).

Williams and Kuribara (2008) used Japlish, a semi-artificial language with English lexis and Japanese morphosyntax (word order and case) in a single-session design. They tested 41 L1 English university students on the acquisition of syntactic scrambling, after 25 of them were trained in the language in incidental conditions by means of a plausibility judgment task and using 194 bimodal (visual and aural) sentence stimuli. In a receptive GJT administered using the same modality they found that the exposure group showed significantly higher accuracy in judging the grammaticality of scrambling compared to the control group both on trained and novel items.

Grey, Williams and Rebuschat (2014) used Japlish to further investigate adult L2 learning of word order and case morphology. In the first of two sessions, they trained two groups of L1 English learners of Spanish, one with high and the other with low L2 proficiency, using an aural plausibility judgment task (20 minutes, 128 items). Immediately after training, and again after two weeks, they tested receptive performance in word order with an aural acceptability judgment task, and the acquisition of the relationship between thematic structure and case marking by means of a picture-matching task (aural sentence stimuli). At immediate posttest the tests revealed that, overall, learners performed significantly above chance on word order but not on case marking. However, accuracy on case marking improved at posttest, becoming statistically significant. The posttest also revealed significant positive relationships between proficiency in the AL and the participants’ ability to verbalize the language rules, as well as between proficiency in the AL and proficiency in Spanish.

Morgan-Short (2007) trained 42 L1 English adults in the language Brocanto2, in the context of a computer game similar to chess over three sessions, a maximum of five days apart (for a description of vocabulary training and the gaming environment see the Methods section). The language, based on Brocanto (Frederici, Steinhauer & Pfeifer, 2002), mirrors the gender morphology of modern Romance languages and displays SOV word order. In the training phase of each session (13 minutes) participants first listened to a set of sentence exemplars while watching the corresponding game constellations on screen (33 exemplars and metalinguistic information in the explicit condition, and 127 exemplars in the implicit condition). After exposure, participants practiced the game in a total of 44 alternating comprehension and production blocks (20 items per block) distributed across the sessions, whereby they listened to a sentence in Brocanto and had to perform the corresponding move (comprehension, see Methods section), or they had to utter a Brocanto sentence to describe a game move they had just watched on screen (production). The behavioral measures of language learning included two aural GJTs (one administered when participants had reached low proficiency and the other at the end of practice), plus one speeded aural GJT, a written GJT, and a free production task administered at the end of practice. Accuracy in the GJT at low levels of proficiency showed that participants in the explicit condition outperformed participants in the implicit condition, but only in the learning of gender agreement structures (noun/article and noun/adjective agreement). At the end of training no significant differences were found between conditions in any of the measures.

In a follow-up study Morgan-Short, Sanz, Steinhauer and Ullman (2010) investigated specifically the acquisition of gender agreement and trained 30 L1 English adults in Brocanto2 using a similar methodology. In that case the study found that although participants in both the implicit and the explicit condition showed significant gains between a GJT test administered at low levels of proficiency and a GJT administered at the end of training, only the implicit group improved significantly on noun-adjective agreement (p. 171).

2.2 Child studies adopting miniature-language paradigms

The main body of literature that has investigated how children learn nonnative sentence patterns and construction-meaning relationships using exposure to semi-artificial languages in laboratory conditions comes from developmental psychology (Boyd & Goldberg, 2012; Braine et al., 1990; Brooks, Braine, Catalano, Brody & Sudhalter, 1993; Casenhiser & Goldberg, 2005; Hudson, Kam & Newton, 2005, 2009; MacWhinney, 1983; Wonnacott, Boyd, Thompson & Goldberg, 2012). A common feature of AL studies looking at morphosyntactic learning in children has been an interest in the role of input, both in terms of frequency effects as well as the way input is structured.

Casenhiser and Goldberg (2005, Experiment 1) investigated how 51 six-year-olds learnt a novel verb-final word order associated with five novel verb forms. The verbs also presented a novel meaning, whereby an entity (NP1) would appear in a location identified by NP2 in ways specified by the verb, e.g., sailing, dropping down, rising, rolling, etc. An exposure phase of the duration of three minutes included 16 video clips paired with audio descriptions (a control group watched the videos with no sound). The study compared two experimental groups, one exposed to an item set where one verb was four times more frequent than the others and one where the frequency of verb tokens was balanced, and a control group. The results of a forced-choice comprehension task, where participants had to match an aural sentence stimulus to the correct video (out of two), showed that both experimental groups performed significantly better than controls in learning the construction and that the skewed-input condition significantly outperformed the balanced-input condition.

More recently, Boyd and Goldberg (2012) adopted a similar exposure and testing methodology in relation to the comprehension of a novel NP1-NP2-V structure, where NP1 was an entity moving towards a goal (NP2) in ways specified by the verb. The study compared 18 five-year-olds, 18 seven-year-olds and 18 undergraduates and tested not only the overall learning of the novel construction (non-linking trials), but also the ability to learn the linking rules relating an NP’s thematic function to a specific syntactic position in the sentence (linking trials). For non-linking trials the forced-choice task found significant group differences, with adults outperforming children, and seven-year-olds outperforming five-year-olds. In the linking trials no significant differences were found between adults and older children, whilst both groups significantly outperformed smaller children, who were at chance.

Wonnacott, Boyd, Thompson and Goldberg (2012, Experiment 1) trained 42 L1 English five-year-olds, in a miniature language with a novel word order (V–NP1–NP2) associated to a novel approach semantics (NP2 approaches NP1 in a way specified by the verb), over three sessions. They manipulated input in the exposure (one-verb vs. four-verbs sets) and item familiarity in the testing tasks, repeated at the end of sessions 1 and 3. In the exposure (one block on session 1 and 3, and two blocks on session 2) the experimenter read aloud 16 sentences each followed by a demonstration of the relevant event using toy animals. The three testing tasks had four trials each – general comprehension trials and trials designed to test argument linking – and included acting out of the event semantics using toy animals after an aural stimulus, production after a visual stimulus (event enacted by the researcher) and a forced-choice task (selecting a video matching with an aural sentence stimulus out of two – linking trials only). Except for the forced-choice task, where children were at chance, the results of the other two tasks showed that there was an improvement of the understanding of the linking rules over time as well as a significant learning effect for the construction as a whole.

One point that has been scarcely addressed in child studies, and deserves more investigation, is the role of morphology in conveying semantic information in sentences displaying a higher level of morphosyntactic complexity. In natural languages case morphology has a function similar to word order in the encoding of thematic relationships, although the individual contributions of these two strategies to the thematic interpretation of noun phrases by learners of a novel L2 are not yet well understood. Although a number of studies have shown that children can learn the relationship between affixes and their meanings after relatively brief exposure to lexical exemplars (Braine et al., 1990; Brooks et al., 1993; MacWhinney, 1983), we have no evidence as yet that the ability to relate morphemes to an associated meaning extends to arguments in full sentences. Also, acquisition of the linking rules between syntactic position and thematic function has been mainly investigated in sentences with bare nouns and no determiners or verb modifiers, but the question whether exposure to an AL in implicit conditions would still result in learning of these relationships in sentences displaying a richer phrase structure remains open.

A further point to consider is the comparison of child and adult rate of learning in instructed contexts, a question that has been central to instructed SLA in recent years (García Mayo & Lecumberri, 2003; Muñoz, 2006). In general, these studies found that after extended periods of classroom instruction (one to three years), adults display higher rates of learning compared to children, although this difference is mitigated in the case of auditory receptive skills (García, Mayo & Lecumberri, 2003, p. 171). To the best of my knowledge the issue of whether rate of learning is age-dependent has not been addressed in laboratory AL studies employing a fully productive miniature language (see Ferman & Karni, 2010, for a study investigating differences in rate of learning of an isolated morphological rule). The distinctive advantage AL laboratory studies offer in this area of investigation is to provide an environment where instruction can be controlled, a situation difficult to achieve in classroom contexts, where teaching methods and materials can vary substantially depending on the age group. In view of the literature review, the research questions of the present study were formulated as follows:

RQ1: To what extent do children and adults differ in the rate of overall comprehension of a new AL in the early stages of learning?

RQ2: To what extent do children and adults differ in the rate of comprehension of the linking rules relating syntactic position and thematic function for argument noun phrases?

RQ3: Is there any evidence that children and adults learn the relationship between a case marker and the thematic function of the associated noun phrase?

3.1 Participants

Six children (M = 9.5; SD = 0.7) and eight adults (M = 29; SD = 8.9), all English monolinguals, were selected for this pilot study. All phases of the experiment complied with ethical procedures approved by the host institution and informed written consent was obtained from adults and from the children’s parents. The children were recruited by advertising the study to parents through schools and libraries in the Greater Manchester area, whilst the adults were recruited among students on campus and individuals who had previously expressed an interest in participating in a language learning study. All participants received £15 for their participation.

3.2 The miniature language

The artificial language, called BrocantoJ, is a modification of Brocanto2 (Morgan-Short, 2007, and subsequent studies), mirrors Japanese morphosyntax but has English phonotactics. BrocantoJ has 12 vocabulary items: three obligatorily transitive verbs (the moves nim, ‘capture’, praz, ‘swap with’, and yab, ‘release’), one intransitive verb (the move klin, ‘move’), four nouns indicating the token symbols (pleck, blom, neep, and vode), two adjectives for the tokens’ shapes (neimo, ‘square’, and troise, ‘round’), and two adverbs to indicate the possible directions of movement (zayma, ‘horizontally’, and noyka, ‘vertically’). Similar to Japanese, BrocantoJ is uniformly head-final and has two postpositional case markers (li for nominative and lu for accusative). The basic word order of the language is Subject-Object-Adverb-Verb, with no determiners and with adjectives preceding nouns inside the nominal phrase. As the learners were L1 English monolinguals, aspects of the language that were novel to them included sentence word order (e.g., verb and adverb position), case morphology, and the possibility of omitting subjects if the information can be inferred from the context. The following are examples of SV intransitive sentences, SOV transitive sentences with expressed subjects, and OV transitive sentences with null subjects. Each sentence relates to the corresponding move in Figure 1.

3.3 Procedure

3.4 Outcome measures of language comprehension

For the purposes of the present study, comprehension was operationalized as general sentence comprehension and comprehension of the linking rules (the relationship between arguments or case markers to the corresponding thematic functions). General sentence comprehension was measured by the number of accurately performed game trials overall, while accuracy in the liking rules was measured by the number of accurately performed trials in the relevant trial subset and by the accurate matching between case marker and token in the forced-choice task.

3.4.1 Linking trial subset

Two of the three moves expressed through the transitive verbs (‘capture’ corresponding to nim, and ‘swap with’ corresponding to praz, Figure 2a and 2b) were symmetrical, whilst the third (‘release’, corresponding to yab, Figure 2c) was asymmetrical. In symmetrical moves (unlike in the case of yab) players had to rely solely on the morphosyntactic information provided in the auditory stimuli (position of the NP in the sentence and case marking on the noun) to perform the correct move out of two alternatives. In other words, they had no contextual hint as to which token would capture the other (nim) or which token would have to move first to initiate the swap (praz). Hence nim and praz sentences, corresponding to 68 out of a total of 120 game stimuli, were identified as a subset suitable for the analysis of the development of the understanding of subject/object linking rules across blocks.

Figure 2 

Symmetric moves corresponding to the verbs nim(a) and praz(b), and the asymmetric move corresponding to yab(c).

3.5 Statistical analysis

As the participants’ sample was not sufficiently ample to allow the calculation of reliable inferences, the statistical analysis presented here has mainly descriptive value. Three main analyses were deployed. In order to ascertain that learning on each of the tasks had occurred, participants’ attainment was compared to the level of chance performance, i.e., the level of performance expected if they were guessing. To evaluate between-group performance, Chi-Squared tests were performed to compare the proportions of accurate to inaccurate responses in the two groups. For the analysis of within-group performance at different points in time, the McNemar test, a variation of the Chi-squared test suitable for repeated-measure designs, was used. All analyses were performed using the SPSS statistical package (version 20).

4.1 RQ1

The frequency data relative to the accurate moves per block were pooled across participants and analyzed per sentence type for the child and the adult group separately (Table 1). In order to ascertain when accurate performance represented evidence of learning, a level of chance performance of 14% correct (corresponding to 2.8 items correct per block) was adopted following Morgan-Short (2007, p. 143). Based on that, the threshold of significance above chance was found to lie between 26% and 27% correct responses (rounded up to 27%, 5.4 items correct per block), χ² (1) = 5.77, p = 0.024, Φc = 0.155, α = .05 (for this and all analyses henceforth α = 0.05 and p values are exact and two-sided).

Overall accuracy in the game was found to be significantly above chance from block 1 and throughout the game for both age groups (Figure 3). An analysis of accuracy scores conducted per sentence type shows that both child and adult responses were significantly above chance for SV, SOV and OV sentences (Figure 4). Counts relative to accurate and inaccurate responses for children and adults were analyzed to ascertain whether the difference between the proportions of accurate to inaccurate responses in the two groups was significant. Overall, the proportion of correct adult responses was significantly higher but with a small effect size, χ² (1) = 17.41, p = 0.000, Φc = 0.104), although a score analysis by block revealed that the difference was significant only at the end of session 2 (block 3), with a small to medium effect size, χ² (1) = 18.31, p = 0.000, Φc = 0.256. Response comparisons per sentence type show that in this sample adults were overall significantly better in performing moves associated to SOV sentence stimuli compared to children, although the effect size was again small overall (χ² (1) = 13.57, p = 0.000, Φc = 0.107), whilst between-group differences for SV and OV stimuli were not found to be statistically significant. Further analysis that narrowed down the comparison to specific blocks, sentence and verb types found significant effects with larger effect sizes. The between-group difference in block 3 was significant only for SOV sentences, χ² (1) = 20.29, p = 0.000, Φc = 0.286, and, among these, only for nim SOV sentences, with a medium to large effect size, χ² (1) = 17.03, p = 0.000, Φc = 0.450.

An analysis of the within-group gains between blocks was also conducted, this time using the McNemar test, a variation of the Chi-squared test that allows the comparison of proportions relative to related samples. The effect sizes of the within-group analysis were small overall and the analysis found a significant increase in the proportion of accurate child responses between block 3 and 4 (corresponding to the end of session 2 and the beginning of session 3), Z = –4.11, p = 0.000, OR = 1.082. For adult responses, an intrasessional significant gain in accuracy was found between block 2 and 3, Z = –2.06 p = 0.038, OR = 1.008.

Looking at the performance on difference sentence types within-groups (Figure 4), children in this sample were significantly more accurate on SV and OV sentences compared to SOV sentences (Z = –2.46, p = 0.014, OR = 1.009 and Z = –2.33, p = 0.019, OR = 1.007, respectively), with no significant differences between SV and OV. No within-group significant differences in the proportion of correct responses per sentence type were found for the adult group.

4.2 RQ2 and RQ3

In order to assess the learning of correct linking rules, a subset of the original data was selected. This included only moves for which accuracy directly depended on the learning of the relationship between the noun-phrase position and its thematic function in BrocantoJ (referred to as symmetrical contexts in the Methods section). Note that linking trials with OV sentences specifically probed the ability to understand linking rules in a context where the only disambiguating element was the case marker, and hence provided direct evidence for the learning of its form-meaning mapping function. Overall, accuracy for linking was again found to be significantly above chance from block 1 and across blocks for both age groups (Table 2 and Figure 5).

The analysis found that overall adults were significantly more accurate than children on stimuli in symmetrical contexts although the effect size of this difference was quite small, χ² (1) = 9.53, p = 0.002, Φc = 0.099 (Figure 6). However, a significant difference with a larger effect size was found for SOV sentences, χ² (1) = 16.22, p = 0.000, Φc = 0.143, whilst there was no significant difference for OV sentences. A comparison of linking accuracy across blocks revealed two statistically significant differences in block 1 and in block 3, with adults performing better compared to children (block 1, χ² (1) = 6.68, p = 0.011, Φc = 0.200; block 3, χ² (1) = 7.93, p = 0.008, Φc = 0.217). As for within-group differences, no statistically significant difference emerged relative to how well linking was learnt in SOV versus OV sentences.

Further within-group analysis of the gains in linking accuracy across blocks found that the number of child-accurate responses improved significantly between block 3 and block 4 but with a small effect (Z = –2.193, p = 0.028, OR = 1.041), and very close to significantly between block 1 and 2 (Z = –1.91, p = 0.055, OR = 0.964). No comparable pattern was found for the adult group, which improved performance on linking in a steady way but by smaller increments.

Participants’ responses to the forced-choice task were also recorded and pooled according to the age group (Table 3). As the response to each task item was binary, the level of chance performance was set at 50%. Based on that, the threshold of significance above chance was calculated to be between 67% and 68%, and rounded up to 68% (χ² (1) = 4.85, p = 0.042, Φc = 0.184).

The number of child-accurate responses was below the one expected for chance performance overall (45.8%), as well as when computed for the individual case markers li (47.2%) and lu (44.4%). The number of correct adult responses, on the other hand, was significantly above chance overall (79.1%), as well as when computed for the individual case markers li (79.1%) and lu (79.1%). In both cases the within-group differences in proportion between correct li and lu responses resulted nonsignificant.

In the debriefing interview, when asked whether they had noticed anything special about any of the words, both children and adults reported noticing that li and lu were not mentioned in the vocabulary training, but were included in the stimuli. Only two adult participants mentioned and described parts of the word order of BrocantoJ correctly. Some participants in both groups said they had tried to work out the meaning of li and lu during the game. They reported they thought the words could have the function to indicate place/direction, were conjunctions (meant ‘and’), had an ordinal meaning (first, second), indicated which element moved, or had a function similar to punctuation. Another adult participant mentioned that during the game she thought the item before li ‘would probably do something’, but wasn’t sure if that was a rule. None of the participants was able to describe the linking pattern or the function of both case markers correctly.