Tapping the Potential of Coherence and Syntactic Features in Neural Models for Automatic Essay Scoring

In the prompt-specific holistic score prediction task for Automatic Essay Scoring (AES), general approaches include pre-trained neural model, coherence and hybrid model that incorporate syntactic features with model. this paper, we propose a novel approach to extract represent essay NSP matches state-of-the-art (SOTA) AES achieves best performance long essays. We apply feature dense embedding augment BERT-based achieve methodology AES. addition, explore various ideas combine coherence, information, semantic embeddings, which no previous study has done. Our combined also performs better than SOTA available even though it does not outperform our syntactic-enhanced further compare pure models analyze strengths weaknesses of methodologies.