As a female professor, implicit gender bias is highly relevant to me. I am motivated to advocate for gender equality. My personal experience inspires me to work on this project.
Online reputation systems are essential to creating trust in the virtual world. However, the assumption that online reputation systems are fairer due to its anonymity dangerously overlooks the robustness and prevalence of implicit bias in ordinary language usage.
Bias in models:
Implicit gender bias is more subtle and prevalent. When machines learn from human language use, the algorithm needs to pay attention to unstructured data (text, images, video, etc.) involving social groups that are more likely to be victims of prejudice. To ensure “fairness” in AI, detecting implicit bias is as important as other data cleaning procedures. We do not want machines/models to acquire biased media frames and biased choices of words.
Bias in data:
The data imbalance issue has impacted the analysis in several ways. First, comments about male professors were almost three times as comments about female professors. Second, the imbalance between genders was further complicated by the imbalance between positive vs. negative comments. There were less than 10,000 negative comments about female STEM professors. As a result, the sentiment-topic cross-categorization with the coherence measurement recommended much fewer topics for negative comments on the female than the other categories. Third, due to the same imbalance issues, the lack of word-gender dissociation may be partially attributed to insufficient data. Lastly, our data were not evenly distributed across disciplines. There were more professors in biology, biological statistics, computer science, engineering, and mathematics in the sample.Example: EduKC