Gated Character-aware Convolutional Neural Network for Effective Automated Essay Scoring

Abstract

Automated Essay Scoring (AES) is a challenging topic in Natural Language Processing. Many current state-of-the-art approaches are based on deep learning models. However, most AES models overlook the importance of character-level information, which is important to both the performance and the fairness. The character-level information is able to provide orthographic knowledge (e.g., spelling) and help the learning of infrequent and ⟨UNK⟩ tokens. In this paper, we propose a Gated Character-aware Convolutional Neural Network (GCCNN) model for the AES task. The proposed GCCNN model incorporates character-level information by a character-level encoder and a gated fusion mechanism. First, the character-level encoder learns word embeddings from sequences of characters by a hierarchical convolutional neural network. Next, the gated fusion mechanism adaptively controls the amount of word-level and character-level information to be fused using vector gating. Then, the essay-level encoder learns an essay representation based on the fused word embeddings. Finally, the fully connected layer maps the essay representation into its corresponding score. The experimental results show that our GCCNN model outperforms the baseline deep learning models. In addition, our qualitative analysis also demonstrates the importance of character-level information for tackling the out-of-vocabulary problem in grading essays.