Attention-enhanced multi-task neural network for English reading comprehension difficulty classification and cognitive load prediction

Accurate assessment of reading comprehension difficulty represents a fundamental challenge in adaptive learning systems and educational content development. Traditional readability formulas rely on surface-level linguistic features, failing to capture the multidimensional cognitive processes underlying comprehension. This research proposes a unified attention-enhanced neural network framework that simultaneously addresses difficulty classification and cognitive load prediction through multi-task learning. The model employs hierarchical attention mechanisms operating at word, sentence, and discourse levels to identify salient linguistic features contributing to comprehension complexity. Experiments conducted on three benchmark datasets—including the publicly released OneStopEnglish corpus, the Cambridge English Readability Dataset, and a cognitive-load-augmented version of RACE (RACE-CL)—demonstrate substantial performance improvements, achieving 89.47% (± 0.83) accuracy in difficulty classification and 0.891 Pearson correlation in cognitive load prediction, representing 7.23% and 18.3% improvements over single-task baselines respectively. Crowdsourced cognitive load annotations attained Krippendorff’s α = 0.82 with NASA-TLX convergent r = 0.78, supporting their use as supervised targets. Comparisons against modern transformer baselines (RoBERTa, DeBERTa-v3, DistilBERT) confirm that the proposed hierarchical-attention architecture remains competitive while requiring substantially fewer parameters. A pre-registered randomized controlled deployment with 2,347 learners produced a 23.6% gain in reading comprehension outcomes (Cohen’s d = 0.93, p < .001) over a static-content control. Ablation studies confirm that both hierarchical attention and multi-task learning contribute significantly to the observed improvements. The proposed framework advances automated reading assessment by integrating computational linguistics with cognitive load theory, enabling precise difficulty estimation and actionable insights for instructional design optimization.