Neural Automated Essay Scoring for Improved Confidence Estimation and Score Prediction Through Integrated Classification and Regression

Masaki Uto, Yuto Takahashi

Abstract: Essay writing questions are a type of constructed-response question commonly used in educational assessments. However, substantial scoring costs and reduced evaluation reliability due to rater biases can be problematic, especially in large-scale assessments. To overcome these challenges, automated essay-scoring models utilizing machine learning technologies have gained significant attention. In recent years, scoring models based on deep neural networks have achieved high accuracy. However, even highly accurate neural models still suffer from scoring errors, limiting their adoption in high-stakes assessments. To address this issue, recent studies have explored scoring models that provide confidence levels along with score predictions. This study proposes improvements to the latest confidence-predicting scoring model, enhancing its performance in both confidence estimation and score prediction.