Source-text transfer errors (STTEs) are pervasive in second language (L2) argumentative writing, particularly when learners integrate translated content from their first language (L1) into L2 drafts. These errors undermine argument coherence, linguistic accuracy, and academic rigor, yet traditional assessment methods lack the granularity and timeliness to identify and address them proactively. To bridge this gap, this study proposes a curriculum-embedded deep analytics framework that combines multitask BERT (MT-BERT) with just-intime (JIT) micro-interventions to predict STTEs in L2 argumentative writing. First, we construct a large-scale annotated corpus of L2 English argumentative drafts ( n = 8, 236 ) paired with their L1 (Chinese) source texts, labeling STTEs across five categories: lexical transfer, syntactic transfer, semantic distortion, pragmatic misalignment, and logical inconsistency. Second, we develop an MT-BERT model that jointly learns STTE prediction, L1-L2 semantic alignment, and argumentative structure classification, outperforming single-task BERT, RoBERTa, and traditional machine learning baselines. Experimental results show that the proposed MT-BERT achieves 0.892 F1-score for STTE detection and 0.876 for error category classification, with significant improvements in handling low-resource error types. Third, we design JIT micro-interventions tailored to predicted STTEs, delivering targeted feedback during the writing process via a curriculum-integrated platform. A quasi-experimental study demonstrates that the framework reduces STTE occurrence by 41.3% and improves writing quality (measured by holistic scores) by 15.7% compared to conventional feedback methods. This research contributes to L2 writing analytics by (1) developing a fine-grained STTE annotation schema and corpus, (2) proposing a multitask deep learning architecture optimized for curriculum-embedded scenarios, and (3) validating the efficacy of JIT micro-interventions in real-world L2 writing instruction. The framework offers scalable, adaptive support for L2 learners and actionable insights for instructors, advancing the integration of natural language processing (NLP) into formative assessment practices.
Can Cheng (Sun,) studied this question.