Demographic Metadata as Construct-Irrelevant Noise in DistilBERT-Based Automated Essay Scoring

Technical Summary: Demographic Metadata as Construct-Irrelevant Noise in DistilBERT-Based Automated Essay Scoring

Problem Statement
Automated Essay Scoring (AES) systems are increasingly deployed to support educators and ensure reliability in large-scale assessments. While human grading is known to be influenced by student demographic characteristics (e.g., socioeconomic status, disability, ethnicity), the efficacy of integrating such demographic metadata into AES models remains underexplored. A critical gap exists regarding whether a "naïve" multimodal fusion strategy—directly concatenating demographic metadata with tokenized textual input—improves predictive accuracy, aids training convergence, or mitigates bias. There is a concern that treating demographic characteristics as scoring signals without sophisticated architectural handling could inadvertently amplify existing educational inequities.

Methodology
This study empirically evaluated the impact of naïve metadata concatenation on a DistilBERT-based AES model using the ASAP 2.0 dataset, which contains approximately 24,278 essays with comprehensive demographic labels (economic disadvantage, disability status, ELL status, gender, and race/ethnicity).

• Experimental Design: Two models were compared:
  1. Baseline Model: Trained solely on tokenized textual input.
  2. Experimental Model: Trained on tokenized textual input concatenated with demographic metadata (formatted as "key: value" strings) using an early fusion strategy.
• Data Preprocessing: The study employed anonymization, text normalization, and specific imputation strategies where missing categorical values were replaced with the token "Unknown" rather than numerical NaNs to maintain token compatibility. To address class imbalance in essay scores, inverse class weighting was applied.
• Architecture & Training: Both models utilized a DistilBERT architecture with the first two layers frozen to preserve pre-trained linguistic patterns. The models were trained using ordinal cross-entropy loss and AdamW optimization. A 10-fold group-based cross-validation strategy was used to prevent data leakage from essay segmentation.
• Evaluation Metrics: Performance was assessed using Quadratic Weighted Kappa (QWK) and Linear Weighted Kappa (LWK). Training convergence was measured via validation loss. Fairness was evaluated using the Mann-Whitney U test on prediction residuals across demographic subgroups to determine score parity.

Key Results
The study found that the naïve integration of demographic metadata significantly degraded model performance and fairness:

• Predictive Accuracy: The baseline model achieved a QWK of 0.727. Upon integrating demographic metadata, the experimental model's QWK dropped to 0.656 (a decrease of 0.071). Similarly, the Linear Weighted Kappa (LWK) deteriorated from 0.564 to 0.481.
• Training Convergence: The experimental model exhibited higher validation loss (1.29) compared to the baseline (1.25), indicating poorer convergence and increased overfitting.
• Score Parity: The baseline model demonstrated parity in 15 out of 19 demographic subgroup tests. In contrast, the experimental model showed exacerbated bias, achieving parity in only 12 out of 19 tests. Specifically, demographic features that were unbiased in the baseline (e.g., socioeconomic status, ELL status, gender) became biased in the experimental model.
• Error Analysis: Subgroup-level mean prediction shifts revealed that the experimental model introduced a systematic upward score inflation disproportionately affecting historically marginalized groups (e.g., students with disabilities, economically disadvantaged students, and English Language Learners).
• Entropy Analysis: Shannon entropy analysis of the classification head showed a statistically significant increase in predictive uncertainty for the experimental model (mean entropy 1.5041 vs. 1.4781), suggesting a modality mismatch between structured categorical metadata and unstructured text.

Key Contributions

1. Empirical Evaluation of Naïve Fusion: This study provides the first empirical assessment of naïve demographic metadata concatenation in a DistilBERT-based AES model trained on the ASAP 2.0 dataset. It demonstrates that this specific strategy simultaneously degrades both predictive accuracy and fairness.
2. Identification of Construct-Irrelevant Noise: The findings characterize demographic metadata, when naively concatenated, as "construct-irrelevant noise." The study posits that this noise creates a sparse feature space and modality mismatch, disrupting the model's ability to extract reliable linguistic patterns.
3. Warning on Systemic Bias: The research highlights that naïve metadata integration does not merely fail to assist; it actively introduces systematic upward score inflation for marginalized subgroups. This serves as a critical warning that without architecturally appropriate fusion strategies, incorporating demographic data risks amplifying educational inequities rather than mitigating them.

Significance
The paper concludes that for DistilBERT-based AES models utilizing naïve early fusion, demographic metadata acts as a detrimental factor. The study argues that while demographic data is relevant to human grading contexts, its direct concatenation with text in transformer-based models creates spurious correlations and modality mismatches. Consequently, the authors suggest that practitioners and policymakers should exercise extreme caution regarding demographic-aware AES deployment. The findings do not preclude the utility of demographic metadata in more sophisticated architectures (e.g., intermediate fusion with cross-attention or gated mechanisms) but emphasize that naïve approaches are counterproductive. Future work is directed toward advanced fusion strategies and targeted data augmentation to address the specific biases identified.