Enhanced Hierarchical Structure Features for Automated Essay Scoring

Automated Essay Scoring (AES) aims to evaluate the quality of an essay automatically. In practice, an essay is usually organized in a hierarchical structure, which means that the writer needs to organize the main ideas into different paragraphs, and organize coherent sentences and appropriate words for the essay. Therefore, it is crucial to model the hierarchical structure of essays for AES. For addressing this issue, most of the existing works used neural network-based architectures (e.g., CNNs and LSTMs) to model the hierarchical structure of essays. Different from previous studies, we propose a novel hierarchical graph structure based on graph convolutional networks (GCN) to encode the hierarchical structure of essays and hope to obtain those structured coherence and discourse information from the graph aggregation. We conduct several experiments on ASAP dataset and the experimental results demonstrate the effectiveness of our method.