A Multitask Learning System for Trait-based Automated Short Answer Scoring

Evaluating students' responses and providing feedback in the education system is widely acknowledged. However, while most research on Automated Essay Scoring (AES) has focused on generating a final score for given responses, only a few studies have attempted to generate feedback. These studies often rely on statistical features and fail to capture coherence and content-based features. To address this gap, we proposed a multitask learning system that can capture linguistic, coherence, and content-based features with Bidirectional Encoder Representations from Transformers (BERT) sentence by sentence and generate overall essay and trait scores. Our proposed system outperformed other existing models, achieving Quadratic Weighted Kappa (QWK) scores of 0.766, 0.69, and 0.701 compared to human rater scores. We evaluated our model on the Automated Student Assessment Prize (ASAP) Kaggle and operating system (OS) data set. When compared with other prescribed models proposed to multitask learning system is a promising step towards more effective and comprehensive writing assessment and feedback.

@article{Ramesh2023,
  title     = {A Multitask Learning System for Trait-based Automated Short Answer Scoring},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {10},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Dadi Ramesh and Suresh Kumar Sanampudi},
  doi       = {10.14569/IJACSA.2023.0141048},
  url       = {https://doi.org/10.14569/IJACSA.2023.0141048}
}