Unconscious biases refer to relatively unconscious prejudice in favor of or against one thing, person, or group compared with another. Unconscious biases deeply affect our day-to-day behavior and decision making like accessing new people or making investment decisions. The lack of awareness about such biases will potentially lead to unwise judgment or exacerbation of stereotypes, both of which may cause problems like severe financial loss or lack of diversity.
With the increasing availability of large datasets that capture user behavior, we can approach the problem through large-scale data analysis. In this talk, I will discuss my works on understanding users’ unconscious biases from large-scale data. First, I will discuss my work on the echo-chamber effect in online investment discussion board as an example of confirmation bias. Our results show that user sentiment in the investment discussion board does not correlate with stock price changes. The majority of market sentiment is produced by a small number of community leaders. Second, I will introduce my work on detecting gender bias in language. We systematically develop a gender stereotype lexicon and compare the performance in stereotype detection between a lexicon-based approach and an end-to-end neural network approach. Our results show that the end-to-end approach significantly outperforms the lexicon approach, by overcoming some fundamental limitations of the lexicon approach.