Profiling the dynamics in the development of syntactic and phonological accuracy among learners of L2 Spanish in immersion condition: A multiple case study

Profiling the dynamics in the development

of syntactic and phonological accuracy

among learners of L2 Spanish in immersion

condition: A multiple case study

S2820757

MA in Applied Linguistics

Faculty of Liberal Arts

University of Groningen

Supervisors:

Professor Dr. Wander Lowie

Professor Dr. Kees de Bot

Table of Contents

Abstract... 3

Introduction ... 4

Background... 7

Emergentism, connectionism and Dynamic Systems Theory ... 7

Variability and modeling in language development ... 12

Relationships between growers ... 14

Literature on variability and modeling in language development... 14

Syntactic and phonological development in first and second language acquisition ... 19

Indicators of linguistic development ... 21

The role of attention and memory in the development of L2 ... 23

The present study: research questions and predictions ... 27

Methodology... 30

Participants ... 30

The task ... 32

Data Analysis and Materials ... 32

Measures... 34

General Measures ... 35

Error-Free Clause per Clause Ratio (EFC/C). ... 35

Detailed Measures ... 37

Correct Gender Agreement per Clause (CorrGA/C) ratio. ... 37

Correct Verb Noun Agreement per Clause (CorrVNA/C) ratio. ... 37

Correct Vowel per clause (CorrVow/C) ratio. ... 37

Correct /r/ phoneme per clause (CorrRphon/C) ratio. ... 37

Results of Variability Analyses ... 38

General Measures of Phonological and Syntactic Accuracy ... 38

Detailed Measures of Phonological and Syntactic Accuracy... 43

Modeling the dynamic relationship between phonological and syntactic accuracy ... 48

Discussion... 54

Conclusion ... 61

Abstract

Language as complex dynamic system is made up by a number of subsystems that interact and mutually affect each other over the course of development. This multiple case study explores the variability patterns and dynamic relationship in the development of phonological and syntactic accuracy among L2 learners. General and detailed measures of accuracy were used to calculate ratios of accuracy. Variability analyses, such as trajectory plots of observed and detrended data, min-max graphs, Monte Carlo simulations and static and detrended correlations were administered to explore the variability and interaction trends in the development of the two growers. In addition, three models of syntactic and phonological accuracy development were configured based on assumptions on the role of cognitive resources allocation and informed by the outcomes of the variability analyses which aimed at replicating the observed data, thus testing the validity of relevant predictions. Findings suggest that variability in linguistic performance along with interaction patterns are complementary sources of information on a learner’s state of development. They also suggest that phonological and syntactic accuracy display a competitive interaction at early stages of L2 acquisition, but that this relationship gradually becomes supportive in later moments of the L2 development. Mathematical models succeeded to reach a satisfactory fit of the observed patterns of interaction, which suggests that efficiency in cognitive resource allocation resulting from more exposure to and use of the L2 is a plausible account of the nature of such relationship.

Keywords: Second language development, accuracy, dynamic systems theory,

Introduction

In the field of applied linguistics, research has traditionally been carried out by resorting to Gaussian statistics. Clearly, by averaging results, comparing values, such as mean and

standard deviation between groups, and establishing relationships of causality among variables it is possible to study representative samples and subsequently generalize findings that concern larger populations. Although the contribution of conventional research methods to the field is unassailable, it is also true that there is a need in current times for new approaches that allow further advance in the endeavor of unveiling the complexity of language development (Verspoor, De Bot, & Lowie, 2011). In reaction to this need, and developed by advocates of linguistic theories such as emergentism and Dynamic Systems Theory (DST), variability analyses and mathematical modeling of language development offer an extraordinary possibility to understand the subtleties of language development over time. Moreover, instead of focusing on products of acquisition, as it has been the concern of most group studies in the past, a DST studies emphasize the process of development and are interested in understanding how language emerges over time.

subsystem, mathematical modeling can be used to flesh out and test the dynamics between interacting growers. Deviation from mean and variable behavior in data, then, are not considered error or noise that needs to be corrected for. Instead, variability is viewed as evidence of a self-organizing system, which constitutes potential evidence of imminent developmental jumps. Thus, advocates of a DST approach to second language development see language learning as an individual process, and claim that developmental studies can better serve the purpose of

understanding how language as a complex system evolves over time (Lowie & Verspoor, 2015). While both variability and modeling techniques are relative innovations in linguistic research, the enthusiasm that the DST approach to language development arises in a part of the academic community contrasts with the current need for more studies that adopt such perspective.

For this reason, the current study stands as a contribution towards a DST exploration and explanation of language development. For such purpose, two subsystems, namely syntax and phonology, were explored at the light of available theories and empirical research, by means of variability analyses and modeling techniques. Both of these subsystems are worth examining because what is known about them is mostly restricted to their separate developmental sequences (Gerken, 1996; Tomasello, 2000; Sebastián-Gallés & Bosch, 2005)andnot much is known about their dynamic interaction. Moreover, even if these subsystems might not be recognized as

separate entities by the learner, from an empirical point of view it might be important to test whether both subsystems operate differently or behave similarly over the course of development. While complexity, accuracy and fluency are indicators of linguistic performance usually

so far in written production and lexical growth, this study takes some distance from those previous works by shifting attention towards L2 oral production.

A task was set to collect longitudinal data from three English native speakers learning Spanish by immersion condition who differed from one another in the amount of prior exposure to and use of the target language. Theoretical and empirical research in the field of cognitive linguistics has repeatedly pointed at attention and memory as two important resources playing a fundamental role in language development (Doughty, 2001; Schmidt, 1990, 1993, 2010;

Robinson, 1995, 2003). It is also clear from research in this field that task complexity (Robinson, 2003), co-occurring tasks (Robinson & Lim, 1993), and planning time and previous knowledge (Chang, 1999; Yuan & Ellis, 2003) pose some restrictions to an individual’s efficient use of attention and memory. All of the previous seem to be potential factors that might influence the development of phonological and syntactic accuracy.

pronunciation, and the nature of the dynamic interaction between both growers in a developing interlanguage.

Background

Emergentism, connectionism and Dynamic Systems Theory

Since the end of the last century, emergentist theories have become influential in applied linguistics research, as they have re-shaped our conception of language acquisition. Emergentism was conceived as a response to the assumptions of the nativist paradigm and the Universal Grammar (UG) theory, which saw language as an independent faculty of mind, believed grammar knowledge to be innate, claimed the existence of a Language Acquisition Device (LAD), and held that the object of linguistic study should be grammar, considering other linguistic subsystems (semantics, morphology, phonology) pertaining to the “periphery of language” (Chomsky 1965, 1981, 1986). Although highly popular among academics, the advance of linguistic research and the technological innovations in the subsequent decades motivated many scholars to question the principles of UG, some of which showed to be difficult to prove or were refuted by empirical findings in other specialized fields. Emergentist accounts, in turn, could overcome some of those limitations to account for a number of L1 and L2

acquisition phenomena.

Having its roots in the early work of John Stuart Mill (1930) who asserted that the properties of a system can be more than the sum of its constituent parts, emergentism holds that most linguistic phenomena can be explained by the interaction of non-grammatical

assumption then is that language emerges from the exposure to the complex and rich

environment the learner is immersed in, due to the interplay of a variety of processes associated with human psychological development (Ellis N, 1998). While the human capacity for pattern recognition and categorization in not questioned from an emergentist perspective, the existence of an innate, language-specific acquisition devise is. Emergentism also differs from nativist theories in that it acknowledges that the study of language should encompass its biological, cognitive and social dimensions, interchangeably affecting each other over time. Furthermore, it addresses not only grammatical properties, but also the lexical, morphological, phonological and semantic components of language, and claims that, rather than a fixed order of acquisition, it is the distributional and probabilistic characteristics of the input which account for the sequence in which certain aspects of language are learnt, given their degree of frequency and saliency (Ellis N, 1998; O'Grady, 2008b). While emergentism does not necessarily reject the study of linguistic and syntactic rules, as this is actually an important source of inspiration for emergentist empirical inquiry, their claim is that, rather than inherently rule-governed, emerging properties can

generate rule-like behavior in language (Ellis, 1998). In contrast, other representatives of the emergentist research agenda take some distance from referring to traditional linguistic sections and are more careful to distinguish between them. Elman (2004, 2009), for instance, claims against a strict division between lexicon and syntactic rules. He further adds that rather than a mental lexicon where words are associated to specific meanings and used according to syntactic rules, words can be understood as having syntactic, phonological and semantic properties that emerge as language develops, and as learners use them in specific events and contexts.

approach to the study of mind and its cognitive properties, connectionism provides the research techniques that make it possible to investigate how simple learning mechanisms can develop complicated representations when exposed to complex language evidence (Gasser, 1990). By using the appropriate software and programming tools, researchers have been able to simulate a variety of psychological developmental processes, language acquisition among them. This has allowed scientists to prove, among other points, that by modeling simple mechanisms and providing them with limited training on how to perform a linguistic task, they can extract the regularities in input and predict information that was not present in the training of the network (Elman, 1996, 2001). Such findings seem to hold true for fields such as syntax, morphology and phonology, showing how rule-like behavior can emerge when no explicit rules have been

provided (McWhinney, 2006). These results have been interpreted as supporting evidence for the claims of the emergentist research agenda and its advocates.

Finally, with roots in the fields of biology and mathematics, the Dynamic Systems Theory (DST) was initially thought of as a mathematical explanation of systems whose chaotic behavior could hardly be predicted. Very much compatible with emergentism and

notion of complex implies that such subsystems interact with one another in the course of time, that their development is non-linear, and that they might develop at a different and varying pace in different timescales (Larsen-Freeman & Long, 1991). Thirdly, the term dynamic refers to the fact that language as a system is not static, but that change over time is in its very nature, due to both internal re-organization and the effect of external conditions (De Bot et al, 2005). As N. Ellis sums up, commenting on De Bot, Lowie and Verspoor (2007) work, language is viewed as

(…) a complex dynamic system where cognitive, social and environmental factors continuously interact, where creative communicative behaviors emerge from socially co-regulated interactions, where there is little by way of linguistic universals as a starting point in the mind of ab initio language learners or discernable end state, where flux and individual variation abound, where cause-effect relationships are non-linear, multivariate and interactive, and where language is not a collection of rules and target forms to be acquired, but rather a by-product of communicative processes. (2007, p.23)

As other important tenets, DST postulates that language displays a variable behavior that is nevertheless interrupted by periods of relative stability, is an iterative process (each stage of development critically depends on the previous one), and that its properties are emergent (De Bot et al, 2005;Van Geert, 2008; Verspoor et al, 2011).

process from one another because their initial conditions (L1, previous L2 learning experience, language aptitude, motivation, time constraints, etc.) are unique. The interconnectedness of language subsystems suggests that the lightest change in one of them might have an impact in all the others, just as a change in learners’ external conditions (say, for instance, a learner suddenly finding himself in immersion setting) might have the same impact as well. It is important to mention here, though, that linguistic subsystems are not necessarily inherent internal

representations in a learner’s mind. This is to say, traditional distinctions among subsystems might not be relevant for learners themselves, because they might not tell between them when operating in an L2, but they might hold great value for scholarly work and language researchers, as they help us understand the complexity behind the phenomena underlying first and second language acquisition. Constant change should remind us of the fact that developmental patterns are non-linear and cannot be fully predicted. Nevertheless, although highly variable,

development in a dynamic system also features moments of relative stability or slowdown. Lastly, the iterative nature of language development implies that learning conditions might be both catalytic and limiting agents along the learning process, both in the short and long run.

alternative approaches have been devised in the last decades, and more statistical tools and procedures are being designed that allow bringing light into language complexity.

Variability and modeling in language development

If one is interested in exploring the subtleties of language acquisition, this can be better done by tracking its development over time (Spoelman & Verspoor, 2010), enterprise for which two different yet complementary methodologies exist, namely variability analyses and modeling techniques. These methodologies are implemented in longitudinal case-studies that make use of large amounts of data to explore and describe patterns of development, as well as to flesh out the dynamic relationships between specific subsystems, at a desired and/or appropriate timescale (Verspoor et al, 2008).

on a determined lapse of time, or over a certain number of samples, and Monte Carlo

simulations, which offer the possibility to test against chance observations in developmental data, among other procedures (Van Geert & Van Dijk, 2002; Verspoor et al., 2008).

Outcomes from variability analyses along with relevant literature concerning the

phenomena under study can be subsequently used to configure mathematical models which are later compared to smoothed data. Data is smoothed in order to reveal the overall fluctuations in development and for the sake of a clearer visual comparison against the devised model. Although many smoothing techniques are based on data averaging, DST studies usually resort to spline functions, which smooth the actual data while retaining important features on local variability, and which differ from techniques used to inspect general trends, such as linear and polynomial equations (Caspi, 2010). Similarly, different types of models have been used in applied linguistic research, such as static, dynamic, linear, non-linear, deterministic and stochastic models.

promising method for investigating language acquisition and development (Van Geert, 2008; Larsen-Freeman & Cameron, 2008).

Relationships between growers

The variables under study, which are traditionally referred to in DST studies as growers, can relate to each other in three different ways; they can display competitive, supportive or

percussive relationships. In a competitive relationship, one grower develops at the expense of the

other one. This means that if one grower displays an ascending trend, the other one should display a descending one. A supportive interaction, in turn, is one in which both growers develop coordinately in an approximately similar scale. Last but not least, a percussive interaction is the one in which the development of one grower precedes and is a necessary condition for the development of the others (Verspoor et al, 2011). Being inherently dynamic, such relationships can vary over time; two variables could be competitors at initial stages of development, to eventually end up as supporters in later stages. Furthermore, variability is a process present at every time scale, so one can think of it not only in a small scale (miliseconds, seconds, minutes or days) but also in a bigger scale (weeks, months, or years) (Caspi, 2010). Because they offer a perfect opportunity to make sense of the unique developmental patterns that are the result of learner’s interaction with inner and outern conditions, variability analyses and modeling techniques seem to be a wise choice for any study which, similarly to the current one, attempts tot study the interaction of two or more linguistic subsystems.

Literature on variability and modeling in language development

some of these studies succeeded in identifying some variability patterns among language

learners, one vexing, recurrent finding was that there would always be a certain degree of “free” variability, which apparently could not be explained by any situational factor. This is how, in the early 90’s, researchers aligned with a DST approach to language development came to be

interested in what this free variability could mean in terms of cognitive and linguistic development. What was frequently interpreted as noise or error by group studies in data

collection stages became the main source of information for exploring language complexity and development in DST studies. Variability, from a DST approach, represents the intrinsic patterns of self-organization in language, which as has been already discussed, is a well-known property of dynamic systems. This became visible in the work of Thelen and Smith (1994). Investigating the development of human cognition, they concluded that the development of motor skills is characterized by a considerable amount of variability as many subsystems (e.g. brain, spinal cord, peripheral, sensory and neuromotor pathways) develop, each at its own pace, while affecting all the others. This reflected the unsteady transition between stages as the different subsystems converged and overlapped. It was concluded that variability was a necessary condition for development.

These ideas soon gained popularity in the field of first language acquisition. Van Geert (1991, 1994) was among the first to apply the DST principles into the field of language

Van Geert and Van Dijk (2002) showed how the mean length of utterance displays a gradual development while the use of preposition features a considerable amount of variability over time. Successively, Mariscal (2009) also used variability techniques to track the development of Spanish gender agreement among children younger than 3 years old. She found that in terms of language development, a mixture of target and non-target forms seems to be normal and, in fact, an indicator of developmental processes taking place.

In the field of second language development, R. Ellis (1994) pointed at free variability as an inherent condition of second language acquisition from an emergentist framework. Referring to Cancino, Rosansky and Schumann’s (1978) study on the acquisition of English negative forms by a group of six native Spanish speakers, he calls the attention to how the authors were puzzled by the amount of variability in data which could not be attributed to any specific reason. More recently, van Dijk, Verspoor and Lowie (2011) reanalyzed Cancino’s et al data using a set of variability techniques. They revealed that, for one of the learners that participated in the original study, the use of different negative constructions showed a classical pattern of variability with peaks and dips over the process. This was interpreted as evidence of a developmental hierarchy from less to more complex constructions. Similarly, Verspoor et al (2008) investigated the development of vocabulary use and sentence complexity in an advanced learner of L2 Dutch. For that, measures for vocabulary use, such as word length, type-token ration, and frequency of appearance of words from the AWL (Academic Word List); and for sentence complexity, such as length of noun phrase, and the number of words per finite verb in a sentence were used. Results showed the expected variable behaviour in most growers, and how some measures (such as the AWL) are quite unstable over time. Additionally, an unexpected trend towards a

Researchers interpreted this finding as evidence of the fact that the learner made sentences complex by using either one or the other strategy intercheangebly. In another longitudinal case study, Spoelman and Verspoor (2010) used a set of accuracy rates and complexity measures in order to explore case endings in the written production of a Dutch learner of Finish. Variability analyses revealed that errors in the use of case ending generally decreased over time, with the exception of the most complex ones. In addition, they found that written samples displayed considerable variability over the 54 collected samples and noticed how the interaction of different complexity measures changed in the course of development. Results also showed a supportive trend between sentence and word complexity as well as for word and NP complexity, while, sentence and noun phrase complexity revealed a competitive interaction.

utterances among participants, some specific observed peaks in the data, and the temporary regression in the proportions of use of the different types of utterances. These findings were interpreted to be evidence of transitions in the process of grammatical development. Caspi (2010) studied the development of complexity and accuracy for both syntax and lexicon in written production. Written samples were longitudinally collected from four participants for a period of 36 weeks. The study included indexes for lexical complexity (complex word ration and general word variation), lexical accuracy (correct lexical use ratio), syntactic complexity

(clause/sentence ratio and ratio of subordinating conjunctions to total clause number), and syntactic accuracy (correct word/error per clause ratio). Variability analyses informed a precursor hierarchical model in which lexicon was a precursor of syntactic complexity, which subsequently enabled accuracy development. Such a model could replicate actual data from three of the participants, showing that the aforementioned growers indeed displayed a dynamic

relationship of precursors and dependents in which the development of complexity is a necessary condition for accuracy in both lexicon and syntax. By configuring a similar precursor model and using spline smoothing techniques, Caspi and Lowie (2013) tested the dynamic interaction between four categories of vocabulary knowledge; namely, recognition, recall, controlled production and free production in a native speaker of Portuguese learning English as an L2. Results from variability analyses informed the design of a mathematical model that successfully reproduced empirically collected data from the participant. This confirmed the hypothesis that the four categories would compete for learner’s resources and conditionally supporting each other’s rate of growth, which confirmed the existence of a receptive-productive gap in

lexical difficulty. To test the assumption that the more exposure to English a learner has, the higher his/her increase in lexical difficulty over time, a logistic growth model was configured based on previous literature and informed in patterns of development that emerged from variability analyses. Such model successfully replicated the data obtained from three of the participants, supporting the assumptions made on early stages about the dynamic relationship between input and vocabulary development.

Syntactic and phonological development in first and second language acquisition

As compiled above, variability analyses and modeling techniques have been used to address a number of issues in first and second language development research. The current study, however, is concerned with the development of syntactic and phonological subsystems among L2 learners. Although language phenomena are rich and varied, syntactic and phonological development seem to have a predominant role in linguistic research.

Research in the field of first language acquisition, for instance, convincingly shows how the early recognition and subsequent production of individual phonemes makes possible the segmentation of syllables among children, which allows the learner to delimit chunks, words and sentences drawing on the distributional patterns of vocalic and consonant clusters (Eimas,

overlapped, phonological development seems to be a precursor of syntactic development. Finally, some other important characteristics of phonological and syntactic development in L1 are the fact that perceptive skills develop before productive ones, for both syntactic and phonological subsystems (Fenson, Reznick, Bates, Thal, Pethick, & Stiles, 1994), and that phonological development has shown to feature higher intra-learner variability than syntactic development (Hayes, 2004; Young-Scholten & Hannahs, 1997).

For most L2 adult learners, however, second language acquisition takes place at a moment in which their L1 is a mostly settled and balanced system. Because adult L2 learners normally have immediate communicative needs that compel them to start using the target language as soon as possible, time for familiarization with L2 phonology prior to syntactic development is not always granted. It follows that while in L1 acquisition phonological development precedes syntactic development, late L2 learners have to develop these two subsystems (and arguably others) at the same time, a process that highly depends on and is affected by a number of internal and external factors (De Bot et al, 2005), one of which is the allocation of attention and memory to simultaneous linguistic subtasks (Robinson; 1995, 2003). Empirical research has shown, for instance, that speech rate of most L2 speakers is usually lower as compared to that one of native speakers (Derwing & Munro, 1996), and that L2 learners’ pronunciation is more prone to error due to the influence of L1 phonology at both phonemic and prosodic levels (Archibald, 1995; Beckman, 1986; Eckman, 1991). This has been labeled as

cross-linguistic influence (Kellerman & Sharwood Smith, 1986). An implication is that while

command of L2 syntax, pronunciation of an L2 is highly unlikely (yet not impossible) to reach at a comparable level to that one of a native speaker (Edwards & Zampini, 2008). However, claims exist that an age effect also exists for the learning of an L2 syntax, and that late L2 learners will encounter a number of difficulties to fully master the grammar of their target L2 language (DeKeyser; 2000; Long, 2006). Although also present in first language acquisition, fossilization of innacurate instances of grammatical structures in the second language seems to be another widely accepted feature of L2 acquisition, albeit the same concern holds for L2 phonology (Selinker, 1973; Selinker & Lakshmanan, 1992). Finally, most linguists agree that second

language development in general is a process characterized by higher degrees of variation among and within learners than first language acquisition.

Wrapping up, we know that the conditions under which L1 and L2 acquisition take place are usually quite different, and that generalizations about each do not necessarily transfer to other. L2 development is allegedly far more unpredictable and variable than L1 development. Consequently, it is of particular interest for the present study to investigate not only how syntax and phonology relate to each other and how they interact over time, but also what variability in language production has to say about linguistic development.

Indicators of linguistic development

and accuracy (Skehan 1998; Ellis R, 2003, 2008). Yuan and R. Ellis (2003) further elaborate on Skehan’s ideas on these three indicators of linguistic development:

Fluency “concerns the learner’s capacity to produce language in real time without undue pausing or hesitation” (Skehan 1996:22). (…) Complexity “concerns the elaboration or ambition of the language that is produced” (Skehan 1996:22) and reflects, Skehan suggests, learners’ preparedness to take risks and to restructure their interlanguages. (…) Accuracy concerns the extent to which the language produced conforms to target

language norms. (p.2)

These constructs have been widely used in research, as both dependent and independent variables, in the study of both written and oral language production. One commonality among such studies is that most of them acknowledge, either implicitly or explicitly, the relationship between these three constructs with the mechanisms involved in the acquisition, representation and processing of a second language (Housen & Kuiken, 2009). Again, and similarly to the distinction between phonological and syntactical accuracy, the indicators of linguistic

performance do not necessarily represent segregated mental states in the learners mind. Instead, they might serve descriptive purposes and might be more relevant to language researchers than to language users. Nevertheless, they have extensively proven to be useful constructs to study the development of different first and second languages (Housen & Kuiken, 2009).

Although the three indicators of linguistic development are believed to interact and be mutually dependent (Skehan, 1996), a view that is very much compatible with a DST

which is believed to exist among the three indicators might also take place within each indicator, for different dimensions of linguistic development. The current study is interested in exploring such possibility. Second, accuracy seems to be the most transparent of the three indicators, which makes it possible to design a task in which accuracy can be studied and where learners’ concerns on complexity and fluency in their language production would be diminished. Third, a number of studies have been carried out that investigate the trade-off effect among the indicators,

(Larsen-Freeman; 2006; Robinson, 2001; Spada & Tomita, 2007; Yuan & Ellis, 2003), and those few that investigate how they develop over time have ordinarily given more relevance to

linguistic complexity (Spoelman & Verspoor, 2010; Verspoor et al, 2008). Thus, by focusing on the development of accuracy, this study seeks to offer insights that can complement our

understanding of linguistic development. Finally, since the aforementioned studies have mostly explored written production, and because written and oral production are two essentially different tasks that involve different planning approaches and a different cognitive resources allocation strategies (Chan, 2015), this study aims at generating some balance in literature by focusing on phonological and syntactic accuracy in oral production exclusively.

The role of attention and memory in the development of L2

learner to communicate comprehensible output depends, to a good extent, on the learners’ ability to produce utterances that respect the norms of the target language grammar (Chierchia &

McConnell-Ginet, 2000), but that accurately accord, at the same time, with its phonological system (Morley, 1991). Serious pitfalls in any of these two components could prevent the speaker from conveying comprehensible ideas and, hence, engage in effective communication.

Among all the factors identified as affecting L2 accuracy (also complexity and fluency), allocation of cognitive resources, mainly attention and memory, seems to be of utmost

importance (Robinson, 1995, 2003). This process of cognitive resources allocation is,

nonetheless, conditioned by the nature of language production tasks learners have to engage in. There are two dominant ways of understanding the role of attention and memory in language production. On the one hand, there are those who claim that attention and memory are inherently limited resources, and that while engaged in linguistic production, L2 learners wade through the process of allocating those limited resources to the three indicators of linguistic development, as Yuan and Ellis (2003) explain:

Second language (L2) learners, especially those with limited proficiency, find it difficult to attend to meaning and form at the same time and thus have to make decision about how to allocate their intentional resources by prioritizing one aspect of language over others (Anderson, 1995; Skehan, 1996; VanPatten, 1999). (p.1)

instances are not due to attentional capacity constraints, but are due to interference in the allocation of such resources, which results from the amount and the type of responses that are expected from the learner, as Robinson (2003) elaborates:

Interference models argue that increasing the number of stimuli and response alternatives or the similarity between them will sometimes lead to confusion, reducing performance efficiency. This can be caused by competition for the same types of codes during information flow, or by “cross-talk” between similar codes. (p.645)

Those who agree with an interference account of the allocation of attentional resources in L2 production usually resort to connectionist models of language processing to account for the constraints L2 learners experiment when attempting to efficiently produce L2 output.

In line with the theoretical considerations the present study holds, and because Levelt’s model has been questioned for not being dynamic and exclusively addressing L1 production, whereas connectionist models better account for bilingual processing (De Bot et al, 2005), it is believed that an interference account of cognitive resource allocation can best explain how L2 learners cope with the process of oral production. In this sense, it is claimed that L2 learners experience attentional shifts while engaged in L2 production, which might result in the trade-off effect between the indicators, especially when learners’ proficiency is in the course of

As for language production tasks, generalizations have been made as for the impact of task planning, availability of related knowledge, and cognitive complexity in accuracy, linguistic complexity and fluency. Some of those generalizations which are specifically concerned with accuracy are that tasks based on familiar information, with a clear structure and requiring

interaction with another speaker might be less cognitively demanding and might be favorable for accurate language production. In addition, L2 learners also show to be more accurate when planning time is provided prior to production tasks, and when post tasks allow some kind of revision of learners’ production, this also tends to favor more accurate output in subsequent tasks. On a divergent note, accuracy might be affected when tasks require manipulation of complex information (Though some researchers might argue otherwise). Similarly, monologic linguistic development requiring the use of less frequent lexical items, and in which pre-planning is not allowed. Finally, performing two sub-tasks at the same time might be unfavorable for accurate L2 output as well, both in oral and written production (see Skehan, 2001; Housen & Kuiken, 2009; Skehan, 2009; and Robinson, 2003, for a comprehensive review). Finally, as these generalizations have been documented and experimentaly explored by research in cognitive and psycholinguistic research during the last 20 years, a commonplace conclusion is that allocation of cognitive resources is hihgly learner dependent (Robinson, 2002).

Unfortunately, there are not many studies that have investigated accuracy only, or that have focused on the relationship between phonological and syntactic accuracy over time. As previously discussed, most studies that have dealt with accuracy have done so by also

that the highest amount of variability takes place at the earliest stages of development. Interestingly, no significant relationship was found between accuracy and complexity. Caspi (2010) studied the development of complexity and accuracy for lexicon and syntax, again in written production. Accuracy seemed to compete with complexity in the lexical dimenion, whereas both growers supported each other at the syntactic dimension. This, however, does not tell us much about accuracy development in oral production. Kuiken and Vedder (2012) studied the effects of task complexity on syntactic complexity and accuracy in both written and oral mode. A finding relevant for this study is that a significant influence was found for proficiency level on accuracy rates on the oral mode. Also, results showed more errors in written than oral production. Nonetheless, errors for oral production were concerned only with grammar, and not with phonetic aspects of the L2. Lastly, in a longtudinal multiple case study, Ferrari (2012) investigated how different tasks influenced measures of complexity, accuracy and fluency in oral production among four L2 learners of italian and 2 native speakers of the same language. Results showed that, all in all, accuracy increases over time, but decreases at some focalized moments as a result of a trade-off effect with complexity, and that the type of oral task (monological vs interactive) does make a difference in accuracy rates, as well as for complexity and accuracy, for both L2 and native speakers.

The present study: research questions and predictions

“(…) but would include any activity, inside or outside the classroom that involves mental

activity around language: understanding, speaking, recall or language, meaningful practicing and so on.” (p.204). In line with this view and for the purpose of the present study, language use was defined as the non-continuous amount of time in contact with the L2, either by formal instruction or incidental learning, as reported by learners themselves. Two general research questions

corresponding with the two methodological procedures reviewed in this section (namely, variability analyses and modeling techniques) were formulated. The first is as follows:

1) What do variability patterns reflect about the dynamic relationship (if any) between phonological and syntactic accuracy within the oral production dimension, for learners of L2 Spanish in immersion condition with differing degrees of L2 use?

As discussed earlier, emergentist accounts of second language acquisition allow to hypothesize that, overall, the more exposure and more use learners make of the target language, the better they should perform in an oral production task, as more time was granted for previous stabilization of the subsystems and the regularization of specific aspects of the L2 phonology and syntax. Additionally, literature on the role of memory and attention in linguistic performance suggests that becoming more proficient in an L2 implies making a more efficient allocation of attention and memory during L2 production. It follows that at early stages of development both subsystems should compete for the allocation of attention and memory, while at more advanced ones both growers should behave with considerably less competition and much more

a) Data from the learner with the lowest degree of exposure to and use of the L2 should display a clear competitive interaction between syntactic and phonological accuracy, although a trend towards support is also expected.

b) Data from the learner with intermediate degree of exposure to and use of the L2 should display a milder competitive interaction between syntactic and phonological accuracy, with an increasing degree of support between growers.

c) Data from the learner with the higher degree of exposure to and use of the L2 should reveal a clearer supportive interaction between syntactic and phonological accuracy, although to some extent, a minor degree of competition is still expected.

This leads to the second research question for the present study, which states as follows:

2) Can a mathematical model successfully replicate the interaction patterns in the data of at least one of the learners, on the basis of a hierarchical order in which syntactic and phonological accuracy are competitors in early stages of familiarization with the L2 but display a tendency towards coordination for the more advanced stages of development?

Methodology

A multiple case study of intra-learner variability was set up in which variability analyses and modeling procedures were used to test the dynamic interaction of phonological and syntactic accuracy among learners of L2 Spanish in immersion condition. Global and detailed measures were administered. Three native speakers of English learning Spanish as an L2 in immersion condition were recruited for the study. They were asked to perform an oral production task, twice a week, for a period of 9 weeks. In total, 18 samples were collected per participant, which were subsequently transcribed. A variety of non-linear time analysis techniques were administered to graphically plot the development and interaction of phonological and syntactic accuracy over time. Data was smoothed by means of a cubic spline function, and a mathematical model informed in the available literature was configured that aimed at replicating the growth rate and interactions as displayed in the smoothed data.

Participants

Three participants, whose L1 was English and who were living in Colombia (a Spanish speaking country) at the moment of the study volunteered to take part in this project. Participants differed from one another in the amount of time that they had been exposed to and had made use of Spanish as L2 prior to the beginning of the study, either by formal instruction or incidental learning.

months, where she learnt Spanish mainly incidentally but also by taking about 5 weeks of Spanish lessons. At the moment of the study, she was exposed to Spanish for about 50% of her time, mainly in general day-to-day activities, like doing the groceries, talking to neighbors and at partially at work. In total, she had been previously exposed to Spanish for about one year. In this study, H will be referred to as the beginner learner.

The second participant, to whom we will refer to as S, was a 25 year old Australian living and working in Colombia as the general assistant in a rural kindergarten. She spoke no other foreign language apart from Spanish. In 2012, she took Spanish lessons for three months, a one and a half hour class each week. She then travelled for one year in South America, where she spoke in Spanish the majority of the time. She then returned to Australia, where she regularly spoke with friends in Spanish. At the moment of the study, she was exposed to Spanish the majority of the time, as she lived with a Colombian family, spoke with the children and other staff in the kindergarten in Spanish and engaged in administrative tasks in Spanish such as updating the notice board and making and receiving phone calls. In total S had been previously exposed to Spanish for a period of about two years. S will be referred to as the intermediate learner.

which somehow conditioned her L2 interaction. Overall, she had been in previous contact with Spanish for about 5 years. M will be referred to as the advanced learner.

The task

An oral production task was designed to collect data for the study. Cartoon-like or single-scene pictures were sent to the participants twice a week for 9 weeks. Two samples were

collected on a weekly basis, for a total of 18 samples per participant. In the task, participants were asked to come up with a coherent story based on the sequence of events or the details on the picture. In line with Robinson (2001, 2003), who claims that monologic tasks where planning time is reduced increase the depletion of cognitive resources’ allocation during language

production, participants were asked to narrate the story without pre-planning time granted. Finally, participants were asked to record themselves and submit the samples to the main researcher right after performing each task.

Data Analysis and Materials

Each recording with an oral production sample was transcribed and coded in CHAT (Codes for the Human Analysis of Transcripts) format, using the CLAN (Computerized

development or error typology should minimize concerns on the researcher’s bias towards percieved improvement in the later samples (Caspi, 2013). Finally, clear parameters on what was and what was not to be considered an error in the current study were devised prior to data

coding, which are presented later in this report. Once data was properly coded, ratings were plotted in trajectory line graphs to inspect patterns of variability and trends of development. This allowed a visual inspection of possible interactions between growers. Then, Min-max graphs were created to explore variability patterns for each participant, which present the bandwith of observed scores. By using moving windows of 5 measurements, a frame that moves up one position at a time was created. Each window overlaps the preceding ones using all the same meassurement occassions minus de first and plus the next one. Thus, the wider the bandwith the greater the amoung of variation (Verspoor et al, 2008). Significance in the differences on the amount of variability among participants was calculated by means of Monte Carlo simulations, using the Microsoft Excel add-in Pooptools (Hood, 2004). The Monte Carlo similuation is a resampling techniques that randomly draws a large number of subsamples from the original sample. By simulating variability under the null hypothesis, the level of statistical variability can be calculated (Verspoor et al, 2008). The alpha level for statistical significance was set up at an alpha level of 0.05. Data was subsequently detrended by substracting the slope in the observed data for a clearer visual inspection of the interaction between growers. At this point, static and moving Pearson correlations between the general and detailed measures were calculated. All of the aforementioned procedures were carried out using Microsoft Excel, and were originally devised by van Geert and van Dijk (2002).

good deal of the original variability (Caspi 2010, 2013). The cubic spline function was preferred over polynomial or linear smoothing because the latter ones are techniques based on data

averaging. As discussed previously in this report, averaging data is not in line with dynamic accounts of second language development since it shadows local variability and gives an imprecise impression of the developmental trends. Later on, outcomes from the variability analyses combined with the insights from literature in the field of second language development were used to devise three mathematical non-linear models, which were set up on the basis of a hierarchical order in which syntactic and phonological accuracy are competitors in early stages of familiarization with the L2 but display a tendency towards coordination for the more advanced stages of development. For each participant, a model was created in which phonological and syntactic accuracy were configured to interact according to the learners’ status of beginner, intermediate or advanced as determined by their degree of language use and prior exposure. The resulting models were later optimized by modifying the corresponding control parameters in order to reach the best possible fit to the smoothed data. By comparing the initial model configuration and the necessary modifications during model optimization, and visually

comparing both resulting graphs, it was possible to determine to what extent the hypothesized assumptions held plausible. The models were configured using a Microsoft Excel BVA macro code, originally developed by Paul van Geert, which allows modeling interactions between growers and thus simulating a variety of developmental processes (Lowie, Caspi, Van Geert & Steenbeek, 2011).

Measures

al. 1998), the present study relied on two general measures of phonological and syntactic accuracy as the main source of information for variability and modeling analysis. However, detailed measures were also devised to further explore the subtleties of phonological and syntactic development and shed some light on very specific, fine-grained components of each linguistic subsystem, and to understand how they interacted with the general ratings. Measures reflected the ratio of correct usage of each specific grammatical or phonological rule in each sample. Since the range of potential syntactic and phonological features to be studied is vast, pervasive and particularly problematic aspects of Spanish language for English native speakers were selected.

General Measures

Error-Free Clause per Clause Ratio (EFC/C). In order to measure the overall accuracy

of participants’ samples for phonology and syntax, clauses were chosen as the unit of analysis and the EFC/C ratio was utilized. This measure was preferred over the more popular EFC per T units due to the learners’ distinct level of L2 use and exposure. The advanced learner’s samples were expected to be longer and more complex than those of the beginner learner, which

increased the likelihood of errors appearance. Because error-free clauses were to be measured (instead of number of errors per clause), it was preferable to have a shorter unit of analysis that could make justice to all learners regardless of their level of proficiency, language use, or language complexity

The EFC/C ratio has been used in previous studies (Tapia, 1993; Ishikawa, 1995),

measure the EFC/C ratio, the transcripts were split into clauses. The guiding criterion for identifying a clause was that it should contain a subject (either explicit or implicit) and a verb. Nonfinite clauses containing a full predicate were also counted as separate clauses. Other phrases that did not contain a noun or a verb were linked to the next clause. Then errors were coded in each sample, after which the total number of error-free clauses was divided by the total number of clauses. Finally, as Spanish is a language featuring a lot of synchronic variation, in both syntax and phonology, the guiding reference for normative use of the L2 was the Colombian variety, further narrowed down to the Paisa accent (as used in the region of Antioquia, in the north-west of Colombia), where all participants were residing at the time of their participation in this study.

Coding of phonological errors was limited to the incorrect pronunciation of vowels (reduced vowels, since vowel reduction in unstressed syllables does not take place in Spanish), incorrectly pronounced /r/ (thrilled) phonemes, unpronounced (thrilled) /r/ or (flapped) /ɾ/ where mandatory, errors in the pronunciation of the /tr/ cluster (characterized by an aspirated /t/

followed by a retroflex /r/), errors in word stress, pronunciation of voiced sound /z (z, c, s are all pronounced as /s/ in this variety of Spanish), pronunciation of the phonemes /ʒ/ and /ʤ/ when /h/ was required (as in general), and pronunciation of the phoneme /h/ when non-required (as in hotel). Neither unnatural intonation nor rhythm was coded, nor was any other case of

mispronunciation even if evident.

Coding of syntactic errors was limited to errors in grammatical gender agreement (as evidenced in pronouns, articles, and adjectives), errors in verb inflection (where the subject was either explicitly or implicitly stated), incorrect use of prepositions, incorrect use of used

inaccuracies in the use of verbs ser and estar. No other syntactic errors were coded even if evident.

Detailed Measures

Correct Gender Agreement per Clause (CorrGA/C) ratio. To measure the CorrGA/C

ratio, errors in the use of grammatical gender agreement were coded for each clause. Then, the number of clauses absent error was divided by the total number of clauses. Errors were of the type La niño bonita esta jugando. When an error was identified for two words linked to the same noun, this was considered just one error.

Correct Verb Noun Agreement per Clause (CorrVNA/C) ratio. Calculated as the total

amount of clauses free from noun-verb agreement errors divided by the total number of clauses. Shift in verb tense was coded as error only when associated by means of conjunctions to a previous verb in a different tense. Errors were of the type Yo juegas or los niños estaban contentos y cantan.

Correct Vowel per clause (CorrVow/C) ratio. Errors in the pronunciation of vowels

due to evident vowel reduction in unstressed syllables were coded. Then the total numbers of clauses absent error were divided by total number of clauses. Because identifying vowel

reduction is a difficult task, whenever there was doubt on whether a vowel had been reduced or not, a conservative approach was taken and the possible error was not coded.

Correct /r/ phoneme per clause (CorrRphon/C) ratio. Error in the pronunciation of the

mandatory were also coded as errors. (e.g. partir, as a possible consequence of non-rhotic L1). Then, the number of clauses absent error was divided by the total number of clauses.

Results of Variability Analyses

General Measures of Phonological and Syntactic Accuracy

First of all, data sets from the EFC/C for phonological and syntactic accuracy were plotted in individual line graphs in order to inspect the trajectories and patterns of variability over time for each of the participants. Graph 1 shows the developmental trajectories of

phonological and syntactic development as measured with the EFC/C ratio. As for linear trends, it is clear how H displays a moderate increase in the EFC/C for both phonology and syntax, whereas both S and M show a low growth rate. A visual inspection of variability patterns

suggests that learners differ a great deal from one another in the degree of variability. Interaction trends suggest competition between the EFC/C for phonology and syntax in the case of H. S displays a more coordinate behavior among both rates, but features some extreme peaks and dips for both growers. Finally, M seems to be the most stable of all participants, showing a coordinate interaction and less variability as compared to H and S.

Graph 1. Rate of development and linear trends as measured by the EFC/C ratio for phonological and syntactic accuracy for H, S, and M.

The initial observation that H’s rate of development for syntax competes with that on of phonology was further supported. As for S, indeed, a highly variable behavior emerges from data for both syntactic and phonological accuracy, but both growers seem to display a mostly

supportive relationship. There are considerable peaks and dips, as can be seen in data point 4 and 13 for syntactic accuracy and data point 2, 10 and 17 and 18 for phonological accuracy.

Similarly, M also shows a seemingly supportive interaction between syntactic and phonological accuracy over most of the data set. M’s data, however seems to display the most stable behavior as compared to both H and S.

points oscillated. A visual inspection of graph 3 revealed that, indeed, S seems to be the most variable of the three participants in both growers. M, on the other hand, seems to be the more stable of the participants, especially for the phonological accuracy rate. H, in turn displays moderate variability, and it can be seen how ranges follow different levels in the graph, further supporting a possible competitive relation.

Graph 2. Patterns of variability with detrended (EFC/C) values.

result was found for the difference between S and M. Finally, static and dynamic moving correlations were calculated for each of the participants. A Spearman’s Rho test revealed a significant positive correlation between the development of syntactic and phonological accuracy for S (ρ=0,61; p<0,01, two-tailed). However, because static correlations usually fall short to reveal the oscillations in the correlation values over time, moving Pearson correlations among detrended EFC/C values for phonological and syntactic accuracy were also calculated, as graph 4 shows. A moving Pearson correlation shows temporal changes in the coefficient values in a moving window of several observed values. For the current data sets, such moving correlations were calculated on the basis of an overlapping window of 5 measurement points.

Graph 4. Dynamic (moving Pearson) correlations between phonological accuracy and syntactic accuracy (EFC/C) for H, S and M

Detailed Measures of Phonological and Syntactic Accuracy

Line graphs were plotted to inspect the development of the CorrVow/C and the

Graph 5. Variability and developmental trend in the four detailed measures for H, S and M

Nonetheless, a closer inspection to variability patterns in the detrended data reveals some interesting facts. For H, the measure for CorrGA/C seem to display a competitive behavior against the CorrNVA/C and the CorrRphon/C. S’s CorrRphon/C ratio seems to be highly

variable over the 18 samples and compete with the CorrGA/C but support the CorrNVA/C ratio. Despite some temporal competition between CorrNVA/C and CorrRphon/C (as observed

Graph 4. Variability patterns between detailed measures for H, S and M in observed and detrended data.

Because developmental data measured in ratios is hardly ever normally distributed, a Spearman’s Rho test was administered to identify possible correlations within the data set of each of the participants, between the general and the detailed measures. Table 7 summarizes such findings, where significant values are displayed in bold. Most significant correlations were found for the data of S. Moreover, most correlations were found for the general and the specific

measures. It is interesting to note how the CorrNVA/C showed moderate and strong correlations with the EFC/C for syntactic accuracy for all of the participants. Similarly, a moderate

Table 5 Spearman Rho correlations between global and detailed measures for H, S and M (** significant at an α level of 0,01; * significant at an α of 0,05)

Graph 8. Interaction between the EFC/C for syntactic accuracy and the CorrGA/C

Modeling the dynamic relationship between phonological and syntactic accuracy

Variability analysis revealed a possible competitive relationship between phonological and syntactic accuracy for H, and an apparent supportive relationship between the same growers for S and M, which is mostly in line with the expectations previously set related to the first research question on the basis of each learner’s degree of previous language use and exposure. To test whether the theoretical and empirical assumptions upon which the present study was based held plausible, three mathematical models were configured, one for each participant.

Under the assumption that a beginner learner (as determined by a the lowest degree of language use, 1 year in H’s case) would feature a competitive interaction between both growers, a model was set up for H where phonological development competed moderately with syntactic development by change. Rate of growth was configured as intermediate and positive for both growers, keeping in mind the hypothesized limitations in cognitive resources allocation for a beginner learner who has to cope with two non-stabilized linguistic subsystems. Also, variability analyses suggested a moderate growth rate over the 18 samples. Initial values were informed both by variability outcomes and what is known about our participant’s language use as

intermediate for syntactic development and low for phonological development. No delay was set up, which is in line with the fact that our learner has already one year of exposure to and use of the language. Carrying capacity was configured as moderate for both growers keeping in mind, again, the hypothesized effect of cognitive resources allocation towards two different subsystems of the language, which was believed to limit the learner’s support for development, yet in

level. Subsequent optimization of the initial set up by modifying the control parameters, allowed reaching a closer fit to the observed data was obtained.

Order Parameters Control Parameters Number of Growers 2 R ate of Gr owth Initial Va lue De lay S uppor t Va ria ti on Type of relationship Strong competition By Change (-1,5/-1,5) Precursors No

Growers Phonological Development 0,03 0,4 1 0,9 0,014

Syntactic Development 0,03 0,9 1 0,8 0,014

Table 10. Order and control parameters after model optimization

Table 10 summarizes the order and control parameters after model optimization.

Competition was optimized as proportional at a level of -1,5 for each grower, whereas variation was slightly corrected. Initial values were accommodated, although the same criteria as in the initial set up remained. Delay’s values were left as initially generated by the model. While the smoothed data for the interaction between phonological and syntactic accuracy for participant H revealed a competitive relationship between both growers, the modeled data was able to reach a close fit of such interaction over time, graph 8 illustrates. Because the smoothed data

corresponds to only 18 data points while the growth model stands for 300 iterations, and for the sake of a more fair comparison, the growth model shown below corresponds to the interval within the model that more closely replicated the smoothed data.

outcomes for S, which revealed a seeming coordination between both growers, featuring high levels of variability. For both growers, growth rate was configured as low and positive, bearing in mind that for intermediate learners the amount to grammatical and phonological aspects of language that need to be controlled for is much higher than for low learners.

Table 11. Smoothed data and interval in the optimized growth model simulating the interaction between phonological and syntactic accuracy for H

linguistic knowledge so far. Finally, no delay or variation was set up for any of the growers. The resulting model displayed a supporting relationship between the growers that lacked the

oscillations visible in the smoothed data for S. Therefore, control parameters were optimized to reach a better fit.

Order Parameters Control Parameters Number of Growers 2 R ate of Gr owth Initial Va lue De lay C arr ying C apa cit y Va ria ti on Type of relationship Support By Change (1/1) Precursors No

Growers Phonological Development 0,025 0,5 1 1 0

Syntactic Development 0,02 0,9 1 1 0

Table 12. Control parameters after model optimization for S

Graph13. Smoothed data and interval in the optimized growth model simulating the interaction between phonological and syntactic accuracy for S

Finally, the initial set up for M’s model was similar to that of S, based on the assumption that and advanced learner (as determined by M’s 5 years of L2 use) should display a mostly supportive interaction between growers, a model was configure in which phonological

effect might exist which compromises a learner’s ability to fully master the phonetics of an L2. Because the smoothed data shows an apparent delay in syntactic development, such grower was configured as delayed. Carrying capacity was set as high keeping in mind that the advanced learner should have a higher level of automatization and control over certain aspects of the L2. Variation was configured at zero. The resulting model could replicate the supportive relationship and the overall oscillation patterns in the smoothed data. Optimization was carried out to reach a better fit as shown in Table 12.

Order Parameters Control Parameters Number of Growers 2 R ate of Gr owth Initial Va lue De lay C arr ying C apa cit y Va ria ti on Type of

relationship Strong support

By Change (1/1)

Precursors No

Growers Phonological Development 0,04 0,8 1 1 0 Syntactic Development 0,03 0,75 30 1 0

Table 14. Control parameters after model optimization for M.

Graph 15. Smoothed data and interval in the optimized growth model simulating the interaction between phonological and syntactic accuracy for

Discussion

As for the first research question, which queried what variability patterns could tell about the relationship between syntactic and phonological accuracy among learners with different degree of prior L2 use and exposure, analyses yielded interesting results that can help us better understand the subtleties of linguistic development. Given the hypothetical limitations in the process of cognitive resource allocation in an oral production task, especially those of attention and memory, when linguistic subsystems are not stabilized, it was expected that data from the beginner learner should reveal a clear competitive interaction between both growers. This was indeed the case as shown by the different variability analyses, which revealed a competitive interaction between rates of phonological and syntactic accuracy development over time. A second expectation related to the first research question was that a milder competitive interaction with a clear trend towards support would emerge from data collected from the intermediate learner, S, who should have used the L2 for a longer period of time and therefore could make a more efficient allocation of such resources. This hypothesis was partially met, as no competition was evidenced but a strong support between phonological and syntactic accuracy did emerge from her data. The last expectation was that given the higher amount of L2 use, data from the advanced learner, M, would feature a predominantly supportive relationship between both growers with some short, focalized moments of competition. Again, this was indeed the case as evidenced in the variability analyses. The delay in syntactic accuracy, which is more evident in the data smoothed by means by a cubic spline function than in the observed values, was

Another interesting yet unexpected outcome from the variability analyses, namely that the intermediate learner partially behaved as hypothesized (no competition, but indeed, strong

support), prompts us to take a look at the data from a different angle. If both S and M show a supportive relationship among both growers, does this mean that S and M are equally advanced learners? Do they share a more or less similar moment in their developmental process? If one were to pay attention only to the dynamic relationship between phonological and syntactic accuracy, one might feel tempted to say that, at least in terms of phonological and syntactic accuracy, they do. However, the amount of variability in general and specific measures shows a different picture. S is more variable than M, a difference that was statistically significant only for phonological accuracy. Moreover, even if not significant, plotted data also reveals similar

conditions for phonological accuracy. It could be said then that in terms of the dynamic relationship between phonological and syntactic accuracy, S and M have reached a point of support among both subsystems; but in terms of variability, M features a more stable behavior than S.

The above referred findings support several of the theoretical assumptions upon which the current study is based. On the first place, it is in line with De Bot et al, (2005); Van Geert and Van Dijk, (2002); Spoelman & Verspoor, (2010); and Van Dijk, Verspoor and Lowie’s, (2011) in that, from a DST approach to language development, variability is a vital source of