Essay Scoring Model Based on Gated Recurrent Unit Technique

Abstract

Educational evaluation is a major factor in determining students’ learning aptitude and academic performance. The scoring technique that relies solely on human labour is time consuming, costly, and logistically challenging as this rating is usually based on the opinion of “biased” human. Several studies have considered using machine learning techniques with feature extraction based on Term Frequency (TF) - Part of Speech (POS) Tagging without consideration to global vectorization (GloVe). These solutions require the process of selecting deterministic features that are directly related to essay quality which is time-consuming and needs a great deal of linguistic knowledge. Gated Recurrent Unit (a variation of Recurrent Neural Network) deep learning technique with focus on morphological analysis of essays for content-based assessment has therefore shown the capability of addressing the challenges posed by other AES techniques by building more abstract and complete linkages among features. Deep learning algorithms such as Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) were used to learn the model with performance evaluation on metrics such as validation accuracy, training time, loss function, and Quadratic Weighted Kappa. The performance results showed that MLP, LSTM and GRU had average Quadratic Weighted Kappa (QWK) values of 0.65, 0.86 and 0.88 respectively with each algorithm having an average training time of 61.4, 62.68 and 67.86 seconds respectively. The loss functions for MLP, LSTM and GRU were 0.296, 0.24 and 0.126. This meant that GRU had the best estimate of the difference between the actual and forecasted scores. MLP, LSTM, and GRU had average validation accuracy of 0.48, 0.537, and 0.511 respectively. GRU was shown to be the optimal classifier and was used in the development of the essay scoring model.