Combining Contrastive Learning and Sequence Learning for Automated Essay Scoring

The objective of automated essay scoring (AES) is to employ artificial intelligence techniques to automate the scoring process and minimize the impact of subjective factors on grading. Previous works tend to treat it solely as a regression or classification task, without considering the integration of both. Additionally, neural networks trained on limited samples often exhibit poor performance in capturing the deep semantics of texts. To enhance the performance of AES, we propose a novel approach that combines contrastive learning with sequence learning, effectively integrating regression loss and classification loss. This paper employs a variety of data augmentation techniques to construct negative samples suitable for contrastive learning, aiming to alleviate the inherent sample imbalance issue in essay datasets. Additionally, we propose to utilize sequence learning for essay scoring, incorporating empirical distribution based on the general distribution characteristics of the essay dataset to address the issue of unbalanced prediction results caused by sample imbalance. Experimental results demonstrate that the proposed multi-task learning framework outperforms the single-task learning framework in enhancing the effectiveness of automatic essay scoring.