Research Paper#Educational Assessment, Natural Language Processing, Machine Learning🔬 ResearchAnalyzed: Jan 3, 2026 15:58
Separating Student Content from Teacher Bias in Open-Response Scoring
Published:Dec 30, 2025 02:06
•1 min read
•ArXiv
Analysis
This paper addresses a crucial problem in educational assessment: the conflation of student understanding with teacher grading biases. By disentangling content from rater tendencies, the authors offer a framework for more accurate and transparent evaluation of student responses. This is particularly important for open-ended responses where subjective judgment plays a significant role. The use of dynamic priors and residualization techniques is a promising approach to mitigate confounding factors and improve the reliability of automated scoring.
Key Takeaways
- •Proposes a framework to separate student content from teacher grading biases in open-ended responses.
- •Uses dynamic priors and residualization to mitigate confounding factors.
- •Demonstrates improved performance when combining teacher priors with content embeddings.
- •Provides a practical pipeline for creating learning analytics that can be used for reflection by teachers and researchers.
Reference
“The strongest results arise when priors are combined with content embeddings (AUC~0.815), while content-only models remain above chance but substantially weaker (AUC~0.626).”