TL;DR
This paper introduces TagDebias, a novel method for mitigating gender bias in pre-trained language models (PLMs) by tagging gender-related terms with abstract, neutral concepts. TagDebias significantly improves fairness in a job-biography ranking task without sacrificing model performance, outperforming traditional scrubbing methods.
What Problem Does This Paper Address?
Pre-trained language models (PLMs) have become ubiquitous in various applications, but they often inherit and amplify social biases present in their training data. This is particularly concerning in sensitive domains like hiring, where biased models can lead to unfair outcomes. Existing debiasing methods for PLMs have limitations: some require resource-intensive counterfactual data generation, while others, like scrubbing, can remove valuable contextual information and degrade model performance. This paper addresses the need for a more effective and less detrimental approach to debiasing PLMs, specifically focusing on gender-profession biases.
Methods
The core of this paper is the TagDebias method. Instead of removing biased terms (scrubbing), TagDebias replaces gender-specific terms (like "he," "she," "man," "woman," and gender-stereotypical adjectives/professions) with abstract, neutral tags. For instance, "John" or "Emily" might be tagged as "Person," and gendered pronouns like "he" and "she" are also mapped to "Person." This approach aims to teach the PLM that these terms represent a neutral concept, thereby reducing gender-associated stereotypes. The debiased dataset is then used to fine-tune PLMs. The authors explore several tagging strategies, including tagging proper nouns, pronouns, possessives, gender-specific terms, gender-stereotype adjectives, and stereotypical jobs.
Datasets
The primary dataset used in this research is BiasinBios (De-Arteaga et al., 2019). This dataset contains biographies related to professional lives and jobs, with over 393,000 biographies across 28 job categories. It exists in two versions: "WithGender" (original) and "WithoutGender" (scrubbed). The authors also utilize the CrowS-Pairs dataset for evaluating social bias in PLMs and the CoLA dataset for some experiments.
Key Insights and Ideas
- Tagging as a less destructive alternative to scrubbing: The paper demonstrates that replacing biased terms with neutral tags is a viable strategy that preserves more contextual information than simply removing them.
- Abstracting gender to a neutral concept: The core idea of mapping gendered terms to a single "Person" tag (or similar neutral concepts) effectively decouples gender from profession in the model's understanding.
- Tagging stereotypical jobs is less effective: While tagging gender-specific terms and adjectives showed promise, tagging stereotypical jobs themselves did not yield significant fairness improvements and sometimes even negatively impacted fairness.
- Fairness in ranking is crucial: The paper highlights the importance of evaluating debiasing methods in downstream tasks like ranking, where subtle biases can still manifest.
Key Results
- The "Gender-specific-term" TagDebias model showed the lowest bias scores in terms of FNED (False Negative Equality Difference) and FPED (False Positive Equality Difference) on the BiasinBios dataset, indicating improved fairness.
- In a job-biography ranking task, the TagDebias model consistently achieved fairer rankings compared to the initial and scrubbed models, with gender distributions in top rankings approaching the ideal 50% split. For example, in the top 20 rankings, the TagDebias model showed a 51.25% proportion of men/women, compared to 60.71% for the initial model and 58.75% for the scrubbed model.
- The TagDebias approach did not significantly harm model performance (accuracy and F1 scores) compared to the initial models.
- On the SEAT benchmark, TagDebias models fine-tuned on CoLA showed improved fairness scores, with some variations depending on the specific tagging strategy and fine-tuning dataset.
Limitations and Future Work
- Dataset and Model Dependency: The effectiveness of TagDebias can vary depending on the dataset used for tagging and fine-tuning, as well as the specific PLM architecture. Future work should explore these dependencies further.
- Limited Tagging List: The current list of gender indicator terms is derived from existing literature. Expanding this list to encompass a broader range of stereotypical terms could enhance the method's effectiveness.
- Extending to Other Biases: The tagging strategy could be extended to address other forms of social bias, such as race and religion.
- MLM Objective Integration: Future research could investigate whether integrating a Masked Language Model (MLM) objective with the tagging strategy or fine-tuning with MLM-generated examples can further reduce bias.
- Large-Scale Evaluation: Evaluating TagDebias on larger, more complex pre-trained language models is a crucial next step.
Summary updated