E-learning is gradually gaining prominence in higher education, with universities enlarging provision and more students getting enrolled. The effectiveness of automated essay scoring (AES) is thus holding a strong appeal to universities for managing an increasing learning interest and reducing costs associated with human raters. The growth in e-learning systems in the higher education system and the demand for consistent writing assessments has spurred research interest in improving the accuracy of AES systems. This paper presents a transformer-based neural network model for improved AES performance using Bi-LSTM and RoBERTa language model based on Kaggle’s ASAP dataset. The proposed model uses Bi-LSTM model over pre-trained RoBERTa language model to address the coherency issue in essays that is ignored by traditional essay scoring methods, including traditional NLP pipelines, deep learning-based methods, a mixture of both. The comparison of the experimental results on essay scoring with human raters concludes that the proposed model outperforms the existing methods in essay scoring in terms of QWK score. The comparative analysis of results demonstrates the applicability of the proposed model in automated essay scoring at higher education level.

电子学习在高等教育中逐渐受到重视，大学扩大提供范围，越来越多的学生入学。因此，自动论文评分 (AES) 的有效性对大学管理日益增长的学习兴趣和降低与人工评分者相关的成本具有强烈的吸引力。高等教育系统中电子学习系统的增长以及对一致写作评估的需求激发了提高 AES 系统准确性的研究兴趣。本文提出了一种基于 Transformer 的神经网络模型，使用基于 Kaggle 的 ASAP 数据集的 Bi-LSTM 和 RoBERTa 语言模型提高 AES 性能。所提出的模型在预训练的 RoBERTa 语言模型上使用 Bi-LSTM 模型来解决传统论文评分方法忽略的论文中的连贯性问题，包括传统的 NLP 管道、基于深度学习的方法，以及两者的混合。论文评分实验结果与人类评分者的比较得出结论，所提出的模型在 QWK 分数方面优于现有的论文评分方法。结果的比较分析证明了所提出的模型在高等教育水平的自动论文评分中的适用性。