Transformer Based Essay Generation and Automatic Evaluation Framework

The purpose of Automated Essay Grading (AEG) systems is to evaluate and assign grades to essays efficiently, thereby reducing manual effort, time, and cost. The traditional AEG system mainly focuses its efforts on extractive evaluation rather than abstractive evaluation. The objective of this research is to explore the differences in the grading system of traditional and grammar schools. This research develops a transformer-based system that combines extractive and abstractive essay generation and evaluation. We utilize the Bidirectional Encoder Representations from Transformers (BERT) model for extractive essay generation and Quillbot for abstractive paraphrasing, and design a framework that evaluates both types of essays. To achieve this objective, we created the Long Essay Poets (LEP) dataset and evaluated this across four modes using four models. We compare the performance of four models: Random Forest, Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a combined approach of CNN and LSTM. After performing the experiment, it is concluded that 46% of grades declined in Mode 3 and 44% of grades improved in Mode 4, and in the context of essay evaluation, the Random Forest model performs better in extractive and merging scenarios, and the Long Short-Term Memory (LSTM) Model outperforms in abstractive essay evaluation.