Introduction
Automated Essay Scoring (AES) systems are typically designed to emulate human raters. However, human judgment is inherently susceptible to subjective biases—ranging from cultural background to personal mood—that are unrelated to the objective quality of a text. When machine learning models are trained on these biased ratings, they inadvertently learn to replicate these flaws, leading to systematic unfairness. This research addresses the challenge of identifying and mitigating rater bias by analyzing the linguistic subjectivity of the justification comments provided alongside numerical scores.
Methodology: Quantifying Subjectivity
The researchers propose a "rectification" process to clean training data of biased samples. This process relies on five specialized Portuguese lexicons designed to capture different dimensions of subjectivity:
- Argumentation: Logical connectors (e.g., "consequently").
- Presupposition: Words implying assumed truths (e.g., "obviously").
- Modalization: Indicators of the writer's stance.
- Sentiment: Emotional markers.
- Valuation: Intensifiers (e.g., "highly").
By mapping rater comments into a five-dimensional "subjectivity space" using Word Mover’s Distance (WMD), the authors can define a "prototypical" subjectivity profile for each score level.
As shown in Figure 4, different subjectivity markers exhibit distinct distributions across the score spectrum. For instance, argumentation markers tend to increase with higher scores, while other markers like modalization show different trends at the extremes. This suggests that the subjectivity distribution within a rater's comment is a strong proxy for the score given.
Bias and Model Misalignment
A key insight of this work is that human subjectivity is a primary driver of disagreement between models and raters. When an AES model's prediction deviates significantly from the human score (misalignment), the human's justification comment often displays a spike in subjective language.
Figure 6 illustrates this correlation: as misalignment increases (moving away from 0 on the x-axis), the distribution of various subjectivity markers shifts. This suggests that the model, by focusing on objective linguistic features, may actually be "more objective" than the human rater in cases of high disagreement.
Results: The "Less is More" Principle
The authors evaluated several models, including Support Vector Regression (SVR) and Multi-Layer Perceptrons (MLP), against a gold-standard test set of 50 essays graded by objective experts.
The results demonstrate that removing the most "biased" training samples—those whose comments deviate significantly from the norm—actually improves model performance. For the MLP model, removing the most subjective 30% of the training data increased the Quadratic Weighted Kappa ($\kappa$) from 0.521 to 0.571. This statistically significant improvement across all model types confirms that filtering out subjective noise allows models to better capture the underlying features of essay quality.
Conclusion and Future Directions
This study provides a robust framework for debiasing AES by treating rater comments as proxies for subjectivity. While the current approach requires textual justifications, it highlights the potential for more fair and objective automated grading. Future work will focus on expanding these lexicons to other languages and exploring methods to identify bias in datasets where only numerical scores are available.
Summary updated