A deep learning-based model for automatic syntactic complexity assessment in L2 English writing: development and pedagogical application

Syntactic complexity serves as a critical indicator of second language writing proficiency, yet traditional assessment methods face challenges in scalability and consistency. Grounded in processability theory and usage-based approaches to second language acquisition, this study proposes a deep learning architecture for automatic syntactic complexity assessment in L2 English writing. The model integrates pre-trained BERT representations with graph attention networks to capture hierarchical syntactic structures, employing a multi-task learning framework that simultaneously predicts multiple complexity dimensions operationalized through both coarse-grained and fine-grained indices. Experimental results on learner corpora, validated through five-fold cross-validation, demonstrate that the proposed model achieves a Pearson correlation coefficient of 0.923 with expert human ratings, outperforming traditional rule-based tools and baseline neural approaches. Furthermore, a semester-long quasi-experimental study involving 186 Chinese university students indicated that the integrated instructional package incorporating automated syntactic feedback was associated with greater writing development, with the experimental group showing effect sizes ranging from 0.71 to 0.89 across complexity measures. These findings provide preliminary evidence that deep learning-based assessment shows potential for supporting L2 syntactic development in educational contexts, though further research is needed to disentangle the specific contributions of the automated model from other instructional factors.