Projects

Unveiling and Mitigating Occupational Gender Stereotypes by Advancing Sentiment Analysis Models for Fairer Outcomes

Collaborators: Mariam Abdullah, Kina Huang, Fanqi Cheng

Our study highlights persistent gender bias in sentiment analysis models. Traditional models, such as BoW + Logistic Regression and BiLSTM, exhibited significant biases, aligning female terms with roles like “Secretary” and “Teacher” and male terms with “Truck Driver” and “Pilot,” underscoring their reliance on biased training data. Transformer-based models, including BERT, ALBERT, RoBERTa, and GPT-4o-mini, showed substantial improvements, with non-significant Female-Male bias scores. However, residual biases persisted, particularly in professions like “Pilot” and “Nurse,” reflecting underlying societal stereotypes in pre-trained embeddings.

Discovering and Mitigating Potential Gender Bias in Machine Learning

Collaborators: Daisy Mo, Yuan Huang

This research aims to tackle gender compensation disparity in the technology industry using machine learning techniques applied to Stack Overflow’s Annual Developer Survey data. We investigate whether machine learning algorithms can identify gender disparities in compensation determination. We identify gender-based compensation differences through data analysis and modeling and explore methods to mitigate bias. Findings show significant compensation gaps between male and female developers. The best-performing model, Bayesian additive regression trees, effectively reduces gender disparities when the gender feature is blinded. However, challenges remain in addressing bias stemming from various sources. Sensitivity analysis confirms the persistence of gender disparities even under scenarios of substantial unobserved factors. Overall, this study contributes to discussions on gender bias in the tech sector and highlights the potential of machine learning to promote fairer compensation practices.

Droughts, Income Shock and Marriage Age – Empirical Evidence from CHARLS

This research studies how negative income shocks affect the age of marriage of females, particularly early marriage in China. In many places in China, there are monetary transfers that occur with marriage: bride price and dowry. Income shocks may affect the age of marriage because marriage payments are a source of consumption smoothing, particularly for a woman’s family. In this research, I use droughts as a proxy for income shock. I found that with 1 unit of decrease in precipitation in the previous year or the current year, the girl’s marriage age will be 0.3% earlier. In places that have higher bride prices, this number is 0.4%. The results indicate that the age of marriage responds to changes in aggregate economic conditions if the bride price exists. This suggests that regulating the high level of bride price and providing prompt and effective subsidies to families affected by income shock may be the solution to the phenomenon of early marriage.