Automated Essay Scoring Incorporating Multi-level Semantic Features

Abstract

Essay writing might reveal the language proficiency of a student. Utilizing intelligent technology to automatically grade essays is an effective method of saving significant manpower and time resources, and improving the accuracy of score. The present models typically rely on shallow semantic features, deep semantic features and multi-level semantic features. Existing models, however, struggle to be superior in both scoring accuracy and generalization performance. As a result, we propose a model that incorporates multi-level semantic features. Specifically, we manually define and automatically extracted the shallow semantic features; we use the BERT pre-training model, convolutional neural networks and recurrent neural networks to extract the deep semantic features; and last, feature fusion is used to score essay automatically. The proposed model outperforms three state-of-the-art baseline methods, according to experimental results on two datasets. Additionally, the generalization of the model has been greatly enhanced. The study has a significant impact on automated essay scoring theoretical investigations and practical applications.