A BERT with Bidirectional Long Short-Term Memory Neural Network for Automated Essay Scoring

Grading essays are time-consuming especially when an instructor handles numerous classes. Automated essay scoring is set to solve this problem as students’ essays are being evaluated automatically by the system. The study aimed to assess the performance of BERT with the Bi-LSTM model in scoring academic essays based on five essay topics: 'Views on Censorship in Libraries,' 'Benefits of Laughter,' 'Effects of Computers on People,' 'What is Patience,' and 'Why Ourselves is our Biggest Foe.' A CRISP-DM methodology framework is applied in creating the five BERT with the Bi-LSTM models on different essay topics. Results show that among the five models created the model for ‘Why Ourselves is our Biggest Foe’ performed the best in terms of overall evaluation while the model for ‘What is Patience’ performed the least. The average kappa score for all five models is 0.98, the accuracy is 0.86, 0.154 MSE, and 0.145 MAE. The BERT with Bi-LSTM models performed better compared to other deep learning models used by other studies in terms of the average kappa score.