Real-Time Academic Essay Analysis Through Intelligent Language Feature Modeling

Academic essay analysis in intelligent writing-support systems depends on decomposing overall quality into teachable sub-dimensions-syntactic complexity, lexical diversity, cohesion, and argument structure-rather than returning a single holistic score that conceals which aspect a student most needs to develop. This paper fine-tunes RoBERTa in a multi-task configuration across four linguistic dimensions on 48,000 academic essays from 3,200 students (36,000 training, 12,000 test; 15 essays per student on average), achieving Pearson$\mathbf{r}=\mathbf{0. 8 6 4}$against expert-rater scores, outperforming a BiLSTM feature model ($\mathrm{r}=0.764$; +13.1%), a TF-IDF SVM (0.684; +26.3%), and a rule-based grammar checker (0.524; +64.9%) while operating at a mean latency of 284ms-within the 500ms real-time threshold for inline feedback. Sub-dimension accuracy ranges from$\mathbf{r}=\mathbf{0. 8 7 4}$for syntactic complexity to$\mathbf{r}=\mathbf{0. 7 9 6}$for argument structure. With deployment over one semester, the acceptance of revision has increased from 68.4% to 84.8% after 3 revision cycles and the mean grade increased by 11.7% [(72.4-64.8)/64.8]. Three constraints emerged: a decrease in performance of 16.2% for low-resource academic language varieties; a decrease of 6.9% for mismatches in discipline-specific vocabulary; and a 5.2% error rate in some instances of feedback (suggestions) that needed human proofreading prior to deployment in high-stakes summative contexts.

Hali tarjima qilinmagan