PROGRAM | Financial Services Analytics

By: Swati Tyagi Chair: Rick Andrews

ABSTRACT

Word embeddings have become a cornerstone in natural language processing (NLP) and machine learning, but they often inherit and perpetuate societal biases, particularly gender bias. This limitation manifests in real-world applications such as resume screening and machine
translation, potentially leading to unfair outcomes. Our research addresses this critical issue by developing advanced techniques for debiasing word embeddings while preserving their semantic utility.

We introduce a novel framework called the Debiasing-assisted Deep Generative Model, which integrates concepts of gender equity into its modeling approach. This model comprises three key building blocks: a feature extraction layer utilizing pre-trained word embeddings and an encoder; a dimension-reduced latent space layer capturing hidden structures and patterns, coupled with a semi-supervised classification layer; and a generative decoder for adaptive re-sampling and reconstruction of debiased embeddings. This architecture employs reweighting and resampling
techniques based on latent structure understanding, coupled with a bias scoring mechanism to mitigate bias effectively while preserving semantic integrity.

Our research begins with the development and evaluation of the Debiasing-assisted Deep Generative Model framework, demonstrating its efficacy in producing debiased word embeddings. The study includes extensive empirical analyses on real-world datasets and user studies to validate the framework & performance against state-of-the-art benchmarks. We then explore the application of debiased embeddings in Neural Machine Translation (NMT) systems, addressing the challenge of gender bias when translating from gender-neutral languages. This pioneering work proposes methods for integrating debiased embeddings into NMT systems, aiming to reduce male default bias in occupations and roles, thus promoting fairer translations.

Finally, we investigate the impact of debiased word embeddings on resume job-matching scenarios. This study aims to practically examine the effect of compounding imbalances in real- world resume data and job classifications, evaluate the effectiveness of the Debiasing-assisted Deep Generative Model framework in real-time scenarios for mitigating gender bias in resume screening and job matching, and propose additional techniques to mitigate gender bias by balancing the training data distribution for both genders across different occupations. We analyze how biased representations in job assignments affect classification processes, scrutinizing the nuanced language of resumes and the presence of subtle gender biases. Our findings reveal significant correlations between gender discrepancies in classification true positive rates and gender imbalances across professions. Through these studies, we validate our proposed debiasing techniques, demonstrating theirΩsuperior performance in mitigating gender bias while maintaining the utility of embeddings for various NLP tasks. This research contributes significantly to the development of gender fairer NLP systems, offering both theoretical advancements and practical implications for promoting gender equity in machine learning-driven applications.

Back >