A search engine approach to estimating temporal changes in gender orientation of first names

This paper presents an approach for predicting the gender orientation of any given first name over time based on a set of search engine queries with the name prefixed by masculine and feminine markers (e.g., "Uncle Taylor"). We hypothesize that these markers can capture the great majority of variability in gender orientation, including temporal changes. To test this hypothesis, we train a logistic regression model, with time-varying marker weights, using marker counts from Bing.com to predict male/female counts for 85,406 names in US Social Security Administration (SSA) data during 1880-2008. The model misclassifies 2.25% of the people in the SSA dataset (slightly worse than the 1.74% pure error rate) and provides accurate predictions for names beyond the SSA. The misclassification rate is higher in recent years (due to increasing name diversity), for general English words (e.g., Will), for names from certain countries (e.g., China), and for rare names. However, the model tends to err on the side of caution by predicting neutral/unknown rather than false positive. As hypothesized, the markers also capture temporal patterns of androgyny. For example, Daughter is a stronger female predictor for recent years while Grandfather is a stronger male predictor around the turn of the 20th century. The model has been implemented as a web-tool called Genni (available via http://abel.lis.illinois.edu/) that displays the predicted proportion of females vs. males over time for any given name. This should be a valuable resource for those who utilize names in order to discern gender on a large scale, e.g., to study the roles of gender and diversity in scholarly work based on digital libraries and bibliographic databases where the authors? names are listed.