Multi-level Feature Fusion for Automated Essay Scoring

Automatic Essay Scoring (AES) is one of the significant and challenging research topics in the Natural Language Processing(NLP) area. However, existing AES models majorly consider features derived from vocabulary while failing to integrate sentence and chapter information from multi-level perspectives. In this study, we proposed a multi-level feature fusion model, which was used to capture the multi-level features from different perspectives to improve the accuracy of scoring. Our model consisted of three components to respectively capture the vocabulary-level, sentence-level, and chapter-level features, which were then fed into a CNN and BiLSTM network for the final essay scor-ing. The results show that the proposed model outperforms a set of state-of-the-art AES models on the dataset of the Kaggle Automated Student Assessment Prize (ASAP) com-petition. The average Quadratic Weighted Kappa (QWK) value reaches 0.816, which verifies the efficacy of the model in the task of automated essay scoring. © 2023, Taiwan Ubiquitous Information CO LTD.