Enhancing Digital Learning through Contextual Reading Comprehension Analysis

Digital learning platforms increasingly assess reading comprehension through automated question-answering evaluation, but most deployed systems conflate literal recall with deeper inferential and critical-evaluation competencies-measuring a narrower skill than they report. This paper applies DeBERTa-v3 in a contextual multi-task configuration across four comprehension dimensions (literal comprehension, inferential reasoning, critical evaluation, vocabulary-in-context) to 28,400 reader-text pairs from 2,840 learners (10 texts per learner), achieving Pearson$\mathbf{r}=\mathbf{0. 8 6 4}$against expert scores. Subdimension accuracy ranges from r = 0.904 (literal comprehension) to$\mathbf{r}=\mathbf{0. 7 6 4}$(critical evaluation), consistent with critical evaluation being the most subjective and contextually variable of the four dimensions. Adaptive text-selection based on the model's comprehension profile produces 28.1% higher comprehension scores [(74.8-58.4)/58.4] relative to static-text control groups, and increases voluntary reading-session frequency by 41.3% [(68.4-48.4)/48.4]. Three limitations apply: figurative language achieves F1 of only 0.584 versus 0.904 for literal language, long-form texts (over 800 words) reduce accuracy by 9.3%, and culturally embedded references not represented in the training corpus degrade performance by 16.2%.

Hali tarjima qilinmagan