Assessing logical consistency in academic writing is challenging.Traditional methods relying on shallow features struggle to capture deep semantic logic and textual structure.This study proposes using spatiotemporal graph convolutional networks (ST-GCN) for this task.Texts are modelled as graphs where nodes are sentences; spatial edges represent static logical links like semantic similarity, and temporal edges model dynamic sequential dependencies.Trained on 12,000 academic texts from journals and student papers across disciplines, the model captures logical consistency from micro to macro levels.In evaluation, it achieved accuracy rates of 92.5% for global consistency, 89.3% for local consistency, 91.7% for lexical cohesion, and 93.2% for argumentative consistency.It achieved a Pearson correlation of 0.89 with human evaluation and a mean absolute error below 0.15, significantly outperforming baselines and offering an effective path for automated assessment.
Jiaqi Yin (Thu,) studied this question.