Students' ability to self-regulate their individual and collaborative learning activities while performing challenging academic writing tasks is instrumental for their academic success. Presently, the majority of such learning activities often occur in computer-supported collaborative learning (CSCL) settings, in which students generate digital learner data. Examining this data may provide valuable insights into their self-regulated learning (SRL) behaviours. Such an understanding is important for educators to provide adequate support. Recent advances in the fields of learning analytics (LA) and SRL offer new ways to analyse such data and understand students' dynamic SRL processes. This study uses a novel psychological network method, i.e., Gaussian Graphical Models, to model the interactions between the students' SRL tactics and how they influence language learning in a CSCL setting for academic writing. The data for this study was generated by first-year foreign language students (n=119) who used a Facebook group as a collaborative space for peer review in an academic writing course. The theoretical lens of strategic self-regulated language learning was applied. The findings show a strong connection between the following tactics: writing text, social bonding and acknowledging. Strong connections between students' reflective activities and their application of feedback, as well as between acculturating, organising and using resources were also identified. Centrality measures showed that acculturating is most strongly connected to all other tactics, followed by acknowledging and social bonding. Expected influence centrality measures showed acculturating and social interactions to be strong influencers. Students' academic performance and their use of tactics showed little correlation.
Academic writing is a vital aspect in a student’s learning path as it is one of the major ways
to achieve several academic goals, enhance critical thinking skills, stimulate creativity, and
create awareness, knowledge and skills about the use of academic and professional discourse
[
Lately, significant attention has been paid to how online collaborative learning spaces can
enable and foster language learners’ autonomy development, their self-regulation in learning
and such skills as writing and speaking in second and foreign language learning [
In recent years, the context of computer-supported collaborative learning (CSCL) has
allowed students to generate a range of new types of digital (log) data, the careful analysis of
which can offer researchers and practitioners new valuable insights into foreign language
students’ SRL processes. This approach would in addition reduce the need for subjective
assessment methods (e.g., surveys and think-alouds) [
This exploratory study aims to further contribute to this emerging trend by focusing on the application of a (to the educational research field) novel psychological network method of Gaussian Graphical Models (GGMs; an undirected network of partial correlation coefficients) to study foreign language students’ SRL activities / tactics in the setting of an academic writing course. SRL tactics refer to the specific, applied ways in which a strategy (e.g., planning or time management) is being employed to meet a goal in a specific learning situation [13].
Networks are used across similar studies to represent different entities (referred to as nodes), and the relationships between them (referred to as edges). The nodes can represent various structures or elements (e.g., humans, cities, countries, or concepts), while edges can be friendships, roads or influence. Psychological networks are a special type of networks where the nodes are variables and the edges represent correlations or a level of dependence between nodes [14, 15]. Psychological networks have been increasingly adopted in psychology and behavioural research to model the interactions and relationships between constructs including behaviour, emotions and mental phenomena (e.g., [16, 17]). The approach stems from the conceptualisation of human behaviour as a complex system with multiple elements that interact and influence each other. This is also the case for education where, for example, learners' goal setting activities often relate to their task management and engagement in the learning process, and both can be linked to their self-reflection activities which, in turn, influence the process. Such interactions cannot be studied individually, but need to be combined in a robust holistic model that maps the various elements of the process. Therefore, in this study, we argue that the application of psychological networks can be a valuable approach that would broaden our understanding of these dynamic interactions in learners’ SRL development. In contrast to the commonly used network models (e.g., social network analysis), psychological networks offer rigorous probabilistic network models that account for spurious correlations and conditional independence. Psychological networks have a large battery of confirmatory statistics to verify the estimated networks and offer a hypothesis generation model that helps advance our understanding of the (learning) process [16, 18].
In summary, this study aims to test the application of GGM methods to uncover the dynamics of foreign language students’ SRL tactics when collaborating online with their peers on academic writing tasks. The following research questions have been formulated: 1. What information can be obtained from analysing foreign language students’ SRL tactics in their academic writing process using Gaussian Graphical Models? 2. How does academic achievement relate to students’ approaches to self-regulated language learning?
Academic writing in the target language is a complex and challenging process in which learners of second and foreign languages need to grow accustomed to a wide range of linguistic rules, interactional goals and socio-cultural contexts [19]. In order to do so successfully, learners can benefit from applying self-regulated learning strategies in their writing practice. Self-regulation in writing refers to “self-initiated thoughts, feelings and actions that writers use to attain various literary goals, including their writing skills as well as enhancing the quality of the text they create” [20, p.76]. Research has shown that learners who received SRL strategybased writing instruction were able to use a wider range of SRL strategies and reported increased levels of performance and linguistic self-efficacy [21, 22].
At present, a considerable number of academic writing instruction, including related tasks
and learning activities, often occur in CSCL settings [
Networks offer a rich framework for the analysis of different phenomena and therefore, have
been used extensively over the years to understand learners’ behaviour in education [23]. Social
network analysis (SNA) has, for example, been used to study interactions between students
and teachers, interactions between students and learning materials, and co-enrolment in similar
subjects [24, 25]. Network representations in learners’ networks can be considered rather
straightforward: the nodes represent learners and the edges among them are the connections
they make by verbal interactions or connections on online platforms [26]. Epistemic network
analysis (ENA) is another network representation that has recently gained popularity to
represent the interactions between different phenomena and in particular, the utterances in
students' interactions (e.g., [
A psychological network approach is another method to represent the relationships among variables (coded interactions in our case), where the nodes represent the coded interactions and the edges represent the association between the variables. The associations are measured by partial correlation. A positive partial correlation between variables means that there is an association between the two variables after controlling for all other variables in the network, similar to regression models. Controlling for other variables gives a robust estimate of the association and limits the problem of spurious correlations due to confounders. For example, in a bivariate correlation, a researcher reports a positive association between coffee consumption and academic achievement, which may be due to unmeasured study time, while in psychological networks, adding coffee to the network model will show the association between study time and achievement, and not exclusively the influence of the coffee. Psychological networks are commonly reported as regularised networks by applying a shrinkage parameter that drops small and insignificant edges from the model and prevents overfitting. This is particularly important, since correlating many variables together may result in ‘noisy’ trivial correlations. Furthermore, psychological networks offer several confirmatory tests to verify the significance of edges or centrality measures.
Psychological networks models have witnessed rapid growth, improvements and
refinements of the methods, interpretation, verification and rigour [16]. Such advancement has
led to an increased adoption in many fields, including exploratory analysis, modelling
psychological phenomena, studying human behaviour, modelling the interplay between
different attitudes, and the understanding of human personality and emotions. In the field of
psychology, which has also been studying individuals’ self-regulation process for decades
(e.g., [
This study makes use of a dataset generated by a group of first-year foreign language majors of English (n=119) as they collaborated with their peers on a Facebook group, integrated in an academic writing course at the University of Antwerp (Belgium), and which served as an online collaborative space for peer review. Learners exchanged written work and brainstormed about the different rules and principles they had to adhere to, as well as the form and argumentation of their writing, and the goals, perspectives and organisation of the course. During the 12 faceto-face contact hours for the course, learners were introduced to the foundations of academic writing and had to hand in three 300-word essays over the course of the semester. While students were introduced to the basics of peer review, there was no explicit self-regulated learning strategy-based instruction or training. Students generated 2594 posts and comments on the Facebook group over a period of three months.
Coding: Based on the Strategic Self-Regulation (S2R) model [13], including the taxonomy
of cognitive, affective and sociocultural-interactive activities in learning, and using the
principles of digital conversation analysis [
Analysis: The frequency of each code contributed by each student was calculated. The R
package Huge was used to apply Gaussianisation (applying a smoothing invertible
transformation to transform the distribution as close to Gaussian as possible) for all variables
to conform with the assumption of normality. We constructed a GGM, i.e. an undirected
network in which continuous variables (elements of SRL in our case) are represented by nodes,
and the strength of partial correlations between the variables is represented by the edges. An
edge between two nodes indicates conditional dependence after controlling for all other nodes
in the network, while the absence of an edge indicates that the two nodes are conditionally
independent after controlling for all other nodes. In particular, a regularised partial correlation
network was estimated using the R! Bootnet package (https://www.r-project.org). LASSO
shrinkage [
The descriptive statistics of the coded interactions show that acknowledging, writing text and writing arguments are the most frequently employed tactics. Large standard deviations (sd) show the wide variety in the use of SRL tactics as well as their frequencies (see Table 1).
The psychological network of self-regulating behaviour shows that there is a strong correlation between writing text and social bonding as well as between acknowledging and social bonding, while there is only a weak correlation between acknowledging and writing text (Figure 1). By using the principles of the psychological network, this means there is a strong conditional dependence between writing text and social bonding, and between social bonding and acknowledging after taking into consideration all other variables in the network. We have also identified a strong conditional dependence between writing arguments and reflecting, between writing arguments and acculturating, between organising and acculturating, between organising and acculturating, and between reflecting and applying feedback. A moderate correlation was found between writing arguments and using resources, as well as between using resources and acculturating. Other notable results include observations such as the fact that applying feedback is only weakly connected to all other relevant tactics, except for reflecting, and that planning does not seem to have strong connections whatsoever.
In Figure 1, we can observe the formation of three groups of strongly dependent and, therefore, connected behaviours: 1. a social group, comprised of social bonding and acknowledging, supplemented, in part, with writing text 2. a reflective group, comprised of applying feedback and reflection, and 3. a managing group, comprised of acculturating, organising and using resources, as well as writing arguments. The pie rings around each node show how much the node’s connections explain its value. The value of the nodes acculturating, social bonding and acknowledging, for instance, can be largely explained by their connections. Most other tactics can be moderately explained by their connections while planning showed very low predictability in this regard.
Unlike traditional social networks, the absence of edges in psychological networks signify conditional independence. In the present data set, there are some tactics which act independently. Organising, for example, is independent from writing arguments, which is, more or less, to be expected. More surprising is the fact that planning is independent from applying feedback, social bonding, using resources and writing arguments. Similarly, writing text is independent from reflecting, using resources and writing arguments. The centrality measures (Table 2) offer another window into the interconnectivity of the different tactics in the self-regulated learning network. Acculturating and acknowledging were the most strongly connected tactics, followed by social bonding and reflecting. We see that acculturating, acknowledging and social bonding were the elements with the most expected influence on all other tactics.
node
Strength Expected Influence node
These results indicate, most notably, that students in our sample chose to discuss
formspecific elements of their writing process separately from the content and argumentation of
their essays. It can also be observed that making plans about which steps to take next does not
naturally find its way into peer discussions here, which might indicate the need for
SRLspecific training in similar CSCL settings in the future [13]. Social bonding seems to form a
bridge between several key tactics in the present sample as it relates to both product-oriented
tactics (such as writing text) and process-oriented tactics (such as organising) (cf. [
The addition of students’ grades as a variable in the network allowed us to estimate the association between different SRL tactics after controlling for students’ academic performance (Figure 2). This approach enabled us to investigate if the use of SRL tactics is dependent on academic performance or not.
The structure of the Grade_network was fairly similar to the SRL_network where academic performance was not a variable, with some striking similarities (Table 3). First, we see strong connections between writing text and social bonding, and social bonding and acknowledging again. We also see similar trends for acculturating, using resources, organising, reflecting and applying feedback. Grades were only weakly correlated with writing text, organising and very weakly with planning and applying feedback. In other words, we can observe that a student’s grade can hardly be explained by its connections to other variables. The centrality measures of strength and expected influence are also similar to the self-regulation network. These similarities indicate that students who score high marks in our sample may have similar approaches to language learning compared to students with lower marks, yet that they might apply tactics to a different degree. In other words, most students practice writing text, social bonding and acknowledging in their collaboration, but might do so more or less frequently, which might influence their academic success in the end.
Table 3 Variables and corresponding explained variance for Grade_network and SRL_network
Variable
In accordance with the first research question, GGM has allowed us to model the complex interactions and dependencies among different SRL tactics using a robust estimation method which only takes into account relationships after controlling for all other variables in the network. This means that every visible link in the network is both significant, independent of collinearity with other variables and represents a substantial dependence or association between the nodes (in our case SRL tactics). GGM has also allowed us to verify the network model using simulation and bootstrapping. Furthermore, the conservative estimation approach (i.e., regularisation) helps discard insignificant relationships and prevents overfitting, making the graphs easier to read and interpret. The absence of edges in a GGM helps one understand the lack of dependence among nodes, which in other methods, cannot be accounted for. Additionally, the paths between nodes that a GMM can generate, help us understand the structure of a network and help generate a hypothesis about the structure of the studied phenomena.
Answering the second research question, we were able to observe that students’ academic achievement does not immediately or inherently relate to students’ approaches to self-regulated language learning in our sample. Other factors such as frequency of use and time spent on learning might be more influential in this regard, but this should be the subject of further research.