Session 3A

Because of its complexity, writing ability develops at a varied pace and is characterized by a high variability in ultimate attainment, both in L1 and L2 (Bereiter & Scardamalia, 1987). Variability in L2 writing can also be attributed to individual differences, both cognitive and affective, that learners bring to the writing task (Kormos, 2012; Papi, Vasylets & Ahmaddian, 2022). Working memory has been posited as one of the central cognitive resources which are believed to support L2 writing, although the empirical evidence about its effects have been mixed (Ahmaddian & Vasylets, 2021). Another open question is the way implicit and explicit knowledge are deployed in L2 written performance (Manchón & Vasylets, 2019). To fill this research gap, we explored the predictive role of the explicit knowledge, implicit knowledge and working memory in L2 written performance. In the experiment, 55 Spanish learners of L2 English performed a working memory test (L1 Spanish/Catalan complex reading span), implicit and explicit knowledge tests (timed and untimed grammaticality judgment tasks) and a writing task (an expository essay on the predefined topic performed within the established time limit and without access to external sources). In the analysis, moderate negative correlations were obtained between explicit knowledge, implicit knowledge, working memory scores and the ratio of errors. Hierarchical multiple regression analysis showed that all three predictor variables independently and significantly contributed to the error ratio, jointly explaining 31% of the variance in the dependent variable. The predictors had similar magnitude of contribution and the standardized beta values had negative signs, showing that with higher explicit/implicit L2 knowledge and working memory the number of errors decreased. These results allow for the conclusion that learners can deploy explicit and implicit knowledge to the same extent in L2 written production. Another relevant finding was that in timed L2 written performance working memory could play a role which is equally important to that of L2 knowledge sources. In terms of its implications, the study contributes to a deeper understanding of the role of working memory in L2 writing performance as we could evaluate the magnitude of its impact in comparison to that of the long-term memory sources, to which working memory is intrinsically connected (Erçetin & Alptekin, 2013). Additionally, this is the first study to show that L2 written performance can rely on both implicit and explicit sources of knowledge, Implications and avenues for future research are discussed.


Ahmaddian, M., & Vasylets, O. (2021). The role of cognitive individual differences in writing performance and written corrective feedback processing and use. In R. M. Manchón & C. Poilo (Eds.), Handbook of Second Language Acquisition and Writing. Routledge.
Kormos, J. (2012). The role of individual differences in L2 writing. Journal of Second Language Writing, 21(4), 390–403. 
Manchón, R. M., & Vasylets, O. (2019). Language learning through writing: Theoretical perspectives and empirical evidence. In A. Benati, & J. W. Schwieter (Eds.), The Cambridge handbook of language learning (pp. 341–363). Cambridge University Press. 

Several studies have shown the positive effects of bilingual education on children's literacy achievements (e.g., Bialystok 2018). However, it is still unclear whether the same effects are visible also among children exposed to multilingual education. Although there have been investigations on how multilingual pedagogies enhance children's confidence and motivation (García & Wei 2014 and Cummins 2019), no study has shown in how far they boost children's literacy and cognitive skills. 
The present study was conducted with 89 fourth and fifth graders of a primary school in Italy. 62 of them spoke a home language different from Italian (e.g., Albanian, Arabic and Chinese). Children of 2 out of 4 classes (one 4th and one 5th grade) had been exposed to multilingual education since the 2nd and 3rd grade, respectively. Multilingual practices were implemented within the Italian school curriculum and consisted in the involvement of children's home languages across different literacy-related activities, from vocabulary building to the development of oral and written narrative skills. The two groups of children (exposed vs. not-exposed to multilingual education: MULT vs. non-MULT, respectively) were matched for age, SES, number of children speaking Italian (or another language) at home, Italian language proficiency and amount of exposure to Italian outside school. 
Children had to write a narrative in Italian (Schneider et al. 2005), which was coded based on story-grammar (i.e., story structure and use of mental-state words). Furthermore, they had to perform an advanced theory-of-mind task tapping into their mentalizing skills (Devine & Hughes 2013). Children could answer in whichever language they preferred. Narrative and ToM-abilities has been considered as crucial indicators of children's literacy and metacognitive development (Dore et al. 2018).
We perform a binary logistic regression with group (MULT vs. non-MULT) as outcome variable and story-grammar score and ToM-score as predictors. The results show that the higher the two scores, the greater the likelihood that the model classifies the children as belonging to the MULT-group compared to the non-MULT one. The interactions between story-grammar and children's home language (Italian vs. other) does not contribute significantly to the model. The same holds for the interaction between ToM-score and children's home language. 
The study shows that children benefit from being exposed to multilingual practices at school. The inclusion of children's home languages during classroom activities enhances their metalinguistic and metacognitive abilities. The benefits of multilingual pedagogies involve all MULT-children, independently of whether their home language is Italian or not.

Applied Linguists have long been interested in the effects of anxiety on second language (L2) acquisition and performance (Teimouri, Goetze, & Plonsky, 2018). At the same time, little research has focused specifically on writing anxiety, with the notable exception of Cheng's (2004) Second Language Writing Anxiety Inventory (SLWAI). Although Cheng acknowledged that the published inventory was only in the "initial stages of development," the original SLWAI continues to be used to measure writing anxiety in current research on the role of individual differences in writing performance (Chen, 2016; Uludag et al., 2019). Seeing as Cheng's participant group was 80% female, respondent gender surfaces as an important factor to consider when judging the appropriateness of any inferences made on the basis of said inventory scores.
The present study realizes Cheng's call for validation research by examining differential item functioning and gender performance on the SLWAI in two novel contexts. 200 students from a university in Brazil and 222 participants from a university in southern Spain completed a comprehensive inventory, in Portuguese and Spanish respectively, that included Cheng's (2004) instrument, unique items from Lee's (2005) Writer's Block questionnaire, and relevant statements from the original Foreign Language Classroom Anxiety Scale, or FLCAS (Horwitz et al., 1986). Under the umbrella of Item Response Theory and utilizing the one-parameter logistic--or Rasch--model, the performance of each item was assessed by comparing the items' z-scores to the critical z-score and by examining Item Characteristic Curves.
Item analyses revealed that certain anxiety-related statements functioned differently for male and female university students. Six items showed uniform differential item functioning (DIF), with females reporting significantly higher values for those items, and two items showed non-uniform DIF. These results suggest that the SLWAI would be improved by the exclusion of several items and by the addition of eight items from Lee (2005) and Horwitz et al.'s (1986) questionnaires. The paper will also identify potential causes of the discovered bias with reference to literature in the field of psychology. Ultimately, this study contributes to the ongoing validation of the SLWAI, and it provides the field with a sound, less-biased instrument with which to measure second language writing anxiety.
Chen, P.-J. (2016). Learners' metalinguistic and affective performance in blogging to write. Computer Assisted Language Learning, 29(4), 811-835.
Cheng, Y. S. (2004). A measure of second language writing anxiety: Scale development and preliminary validation. Journal of Second Language Writing, 13, 313-335.
Horwitz, E. K., Horwitz, M. B., & Cope, J. (1986). Foreign language classroom anxiety. The Modern Language Journal, 70(2), 125-132.
Lee, S. (2005). Facilitating and inhibiting factors in English as a foreign language writing performance: A model testing with structural equation modeling. Language Learning, 55(2), 335-374.
Teimouri, Y., Goetze, J., & Plonsky, L. (2019). Second language anxiety and achievement: A meta-analysis. Studies in Second Language Acquisition, 41(2), 363-387.
Uludag, P., Lindberg, R., McDonough, K., & Payant, C. (2019). Exploring L2 writers' source-text use in an integrated writing assessment. Journal of Second Language Writing, 46.

One major area in learner corpus research is text quality, which concerns semantic–pragmatic aspects of language use to influence overall text quality (e.g., Crossley et al., 2019). Despite increasing interests in employing various NLP techniques (e.g., Dascalu et al., 2017), little attention has been paid to how similarly/differently each technique reveals L2 constructs such as learner proficiency. In addition, NLP-based L2 research is heavily biased towards L2-English, which does not ensure the generalisability of its implications. Against this background, we investigate the relationship between learner proficiency and text similarity of L2-Korean learners' written production (relative to native speakers' writing) measured through transformer-architecture neural-network models, which are cutting-edge techniques in machine learning. 
Method (Table). Thirty-three L1-Czech L2-Korean learners (age: mean = 24.0; SD = 2.69) were asked to write argumentative essays on two topics. Learner proficiency was measured separately using the Korean C-test (Lee-Ellis, 2009; ranging from 0 to 188; mean = 103.74, SD = 25.66). Essays from 25 native Korean speakers were collected as a reference text. After electronically converting all the essays with typos and spelling/spacing errors uncorrected, we computed cosine similarity scores between individual learner writing and the reference text by employing two transformer-architecture neural-network models-BERT (Devlin et al., 2018) and GPT-2 (Radford et al., 2019). The similarity scores (predictor) and proficiency scores (outcome) were then submitted to linear regression. In addition, we recruited 10 raters to holistically evaluate learner essays for content, organisation, and language use (as employed in the TOPIK writing evaluation process).
Results (Figure). Whereas the rating–proficiency tendency was uniformly positive, the similarity–proficiency relationships were idiosyncratic. This indicates that the similarity scores, obtained automatically from the neural-network models, fundamentally differ from the rating scores obtained holistically from human evaluation. In addition, no uniform tendency was found across the topics for both models; only a few regression analyses yielded significance for the similarity–proficiency relationships. 
This eccentric performance in predicting proficiency indicates two possibilities: (i) the operation of these models may have been greatly influenced by such factors as essay topics in an asymmetric manner; (ii) the transformer-architecture models may not have been adept at extracting a centralised tendency from learner writing in general. Both possibilities are attributable to the properties of the transformer architecture's internal algorithms: they utilise raw sentences (with no POS information) as a basic data-processing unit, assuming that sequences of portions of the sentences comprises a context allowing those sequences to share certain distributions/meanings. Because of this nature, the transformer architecture may not work well in managing learner writing (due to learner language characteristics; Meurers & Dickenson, 2017) compared to its state-of-the-art performance in many downstream NLP tasks.
Together, our findings suggest that the application of NLP techniques to learner corpora needs a researcher's sound understanding of how their algorithms operate in conjunction with various factors that possibly affect their operation.
EuroSLA31_1643138666abstract_ts.pdf