Relevance-Based Automated Essay Scoring via Hierarchical Recurrent Model

In recent years, neural network models have been used in automated essay scoring task and achieved good performance. However, few studies investigated using the prompt information into the neural network. We know that there is a close relevance between the essay content and the topic. Therefore, the relevance between the essay and the topic can aid to represent the relationship between the essay and its score. That is to say, the degree of relevance between the high score essay and the topic will be higher while the low score essay is less similar to the topic. Inspired by this idea, we propose to use the similarity of the essay and the topic as auxiliary information which can be concatenated into the final representation of the essay. We first use a hierarchical recurrent neural network combined with attention mechanism to learn the content representation of the essay and the topic on sentence-level and document-level. Then, we multiply the essay representation and the topic representation to get a similarity representation between them. In the end, we concatenate the similarity representation into the essay's representation to get a final representation of the essay. We tested our model on ASAP dataset and the experimental results show that our model outperformed the existing state-of-the-art models.