Automated Essay Structuring Support via Discourse Pattern Recognition

Essay structuring support systems that identify discourse moves-thesis, evidence, concession, elaboration, conclusion-give students actionable structural feedback that holistic quality scores cannot provide. This paper applies an RST-Transformer model to discourse-move detection across 24,000 academic essays annotated with 172,000 move-level labels (five categories, mean six moves per essay), achieving a macro-averaged F1 of 0.864, outperforming a BERT sequence classifier (F1 0.804; +7.5%), a CRF sequential tagger (0.724; +19.3%), and a rule-based keyword matcher (0.584; +48.0%). Essay-level structural completeness scores derived from the detected moves correlate at r = 0.848 with expert completeness ratings, and the model identifies missing structural sections with 84.4% recall. The quantities of student essays with all of the required structural moves increased from 48.4% to 74.8% (+54.5% relative) after one academic term of deployment and mean revision cycles per essay decreased from 3.8 to 2.4 (−36.8% [(3.8-2.4)/3.8]). There were 142 false positive flags (0.6% of essays), and the model had reduced F1 of 0.724 compared to 0.864 (−16.2%) for non-linear and reduced F1 of 0.804 compared to 0.864 (−6.9%) for highly technical discipline-specific argumentation.

Hali tarjima qilinmagan