Gender Bias in Coreference Resolution

Coreference resolution has evolved from rule-based systems like Lee et al. (2011) to statistical models such as Durrett and Klein (2013), and more recently to neural architectures like Clark and Manning (2016). Despite these advances, issues of bias—particularly gender bias—remain underexplored. Prior research identified biases in word embeddings (Caliskan et al., 2017; Garg et al., 2018), but less attention has been paid to biases manifesting in coreference tasks. As models are increasingly deployed in sensitive domains, understanding and mitigating bias has become urgent. Existing datasets often reflect societal stereotypes, which models inadvertently learn and propagate, raising fairness concerns.

The main challenge is that coreference systems tend to reinforce gender stereotypes present in training data, leading to biased pronoun resolution—e.g., associating 'doctor' predominantly with 'he' and 'nurse' with 'she'. This bias affects both model fairness and societal perceptions. Current evaluation methods lack controlled, systematic tools to measure bias directly within coreference tasks. The difficulty lies in disentangling genuine linguistic patterns from societal stereotypes, and in quantifying the bias's strength and its correlation with real-world data. Addressing this gap is critical for developing equitable NLP systems.

The core innovation is the development of Winogender schemas—structured, human-validated sentence templates that isolate gender bias in coreference resolution. Unlike prior approaches focusing on word embeddings, this method directly probes model behavior in the context of pronoun resolution for occupations. By integrating societal gender ratios from BLS data, the approach quantifies bias strength and correlates it with real-world disparities. This schema-based evaluation provides a standardized, interpretable, and scalable tool for bias detection, enabling precise diagnosis and guiding mitigation efforts.

• �� Design 120 sentence templates inspired by Winograd schemas, each with three key mentions: OCCUPATION, PARTICIPANT, and PRONOUN.
• �� Vary PRONOUN gender (female, male, neutral) and PARTICIPANT type (specific or generic), ensuring sentences are unambiguous for humans.
• �� Validate sentences via Mechanical Turk, achieving 99.7% agreement.
• �� Evaluate three coreference models: rule-based (Stanford sieve), statistical (Durrett-Klein), and neural (Clark-Manning), on these templates.
• �� Collect model predictions, analyze bias by measuring resolution differences across pronoun genders.
• �� Correlate bias metrics with occupational gender ratios from BLS and textual data.
• �� Identify 'gotcha' sentences where model predictions conflict with societal expectations, assessing bias robustness.

The evaluation involved 120 sentence templates, each instantiated with three pronoun genders and two participant types, totaling 720 sentences. Models' coreference decisions were recorded, focusing on their consistency across gender variations. Bias was quantified by the difference in resolution rates for female versus male pronouns. The analysis included correlation with BLS occupational gender ratios (r=0.87), and assessment of performance on 'gotcha' sentences, revealing significant bias amplification. Additional ablation studies tested the impact of training data and model architecture on bias levels.

All three models displayed significant gender bias, with the rule-based system resolving differently 68% of the time, and neural models 13%. Bias aligned strongly with societal gender ratios, with correlation coefficients exceeding 0.8. The models favored male pronouns in occupations like 'manager' (38.5% female in BLS, but models predicted near 0% female). 'Gotcha' sentences showed decreased accuracy, confirming bias influence. The findings demonstrate that models learn and magnify societal stereotypes, affecting fairness in real-world applications.

This schema-based evaluation can be integrated into model development pipelines to detect and reduce gender bias early. It is applicable in domains like hiring automation, legal document analysis, and social media moderation, where fair language understanding is critical. The approach also informs dataset curation and bias mitigation strategies, promoting responsible AI deployment. Long-term, it supports the creation of models that reflect societal diversity and fairness, reducing harmful stereotypes.

The schemas focus primarily on occupational bias, limiting scope for other social biases. Sentence templates, though validated, may not capture all linguistic nuances. The reliance on societal statistics assumes their accuracy and completeness. The approach may not generalize across languages or cultural contexts. Further research is needed to extend schemas and evaluate bias mitigation effectiveness.