Presented at NAACL 2024, this research introduces a mathematically rigorous and computationally efficient framework for stripping social biases from Natural Language Processing (NLP) models. By leveraging the statistical theory of "sufficient dimension reduction," the authors provide a unified method that works across various embedding types, addressing both intrinsic word-level associations and extrinsic downstream task fairness.
The Challenge of Model Bias
Pre-trained models like BERT or GloVe inadvertently learn human prejudices—such as gender, race, and religion—from their training corpora. Researchers typically evaluate this bias through two lenses:
- Intrinsic Bias: Measuring whether words like "man" and "doctor" cluster more closely than "woman" and "doctor."
- Extrinsic Bias (Fairness): Assessing whether downstream applications, such as resume screening, perform differently based on sensitive attributes.
A significant hurdle in debiasing is that these two types of bias are often weakly correlated; mitigating one does not guarantee the mitigation of the other. Furthermore, existing methods are often computationally expensive (adversarial training), limited to binary categories, or overly aggressive, potentially damaging the model's underlying semantic knowledge.
The Sufficient Universal Projection (SUP) Framework
The core contribution of this work is the Sufficient Universal Projection (SUP) algorithm. Unlike previous methods that assume a strong linear relationship between bias and vector representation, SUP adopts a "weak" linearity assumption. This allows for non-linear relationships between sensitive attributes (e.g., race) and the vector space, provided the bias resides within a specific subspace.
As shown in the figure above, the framework utilizes Sufficient Dimension Reduction (SDR) techniques—specifically Sliced Inverse Regression (SIR) and Pooled Marginal Slicing (PMS). These methods identify the "central subspace," which is the minimal set of directions capturing all information about a sensitive attribute. Once identified via eigenvector decomposition, the algorithm projects the original vectors onto an orthogonal "clean" subspace.
Key Insights and Theoretical Foundations
- Theoretical Bridge: The authors provide a mathematical proof (Theorem 6.1) demonstrating that identifying and projecting out the minimal subspace for a sensitive attribute creates a representation that is both intrinsically debiased and extrinsically fair.
- Continuous Attributes: Because SUP is grounded in regression-style statistics, it can handle continuous sensitive variables. This allows gender or other attributes to be treated as a spectrum rather than a binary toggle.
- Closed-Form Efficiency: Unlike iterative adversarial methods, SUP involves calculating a covariance matrix and its eigenvectors, making it a "plug-and-play" solution for existing models.
Experimental Results
The authors evaluated SUP across static (GloVe) and contextual (BERT) embeddings using benchmarks like WEAT, SEAT, and various classification tasks (BIOS, MOJI, Toxic).
As detailed in the results tables, SUP consistently outperforms state-of-the-art baselines like INLP and RLACE:
- Superior Bias Reduction: In the MOJI sentiment task, SUP reduced the "discrepancy gap" (a measure of unfairness) from 31.0 to 10.5.
- Utility Preservation: On word similarity tasks (e.g., WordSim-353), SUP maintained or slightly improved correlations with human judgments, ensuring semantic integrity.
- Extreme Speed: In the BIOS classification task, SUP took only 6.76 seconds to compute the projection, compared to 271 seconds for INLP and over 4,300 seconds for RLACE.
- Intersectional Bias: SUP proved uniquely effective on the "Toxic" dataset, which involves comments belonging to multiple sensitive groups (e.g., race and sexual orientation).
Limitations and Future Directions
While SUP represents a significant advancement, the authors note that evaluations were primarily English-centric due to a lack of established fairness benchmarks in other languages. Additionally, while the framework handles weak linearity, the final transformation remains a linear projection. Future work may explore fully non-linear transformations to further refine the debiasing process.
Summary updated