Extraction of Evaluative Elements for Cross-prompt Automated Essay Scoring

Automated Essay Scoring (AES) systems attempt to automatically evaluate student-written essays with machine learning models. Existing AES trials are mostly designed prompt-specifically with supervised learning, which has limited their applicability in real-life scenarios. We extract evaluative elements from the source set of essays as axes in the vector space, applying dimensionality reduction by Principal Component Analysis (PCA). We then transfer them to a different target set of essays for score prediction. Simplified cross-prompt binary clustering task of dividing high/low-scored groups shows an acceptable level of accuracy.