A Data-driven Approach to Studying Given Names and their Gender and Ethnicity Associations

A Data-driven Approach to Studying Given Names and their Gender and Ethnicity Associations Shervin Malmasi Centre for Language Technology Macquarie University Sydney, NSW, Australia shervin.malmasi@mq.edu.au Mark Dras Centre for Language Technology Macquarie University Sydney, NSW, Australia mark.dras@mq.edu.au Abstract Studying the structure of given names and how they associate with gender and eth- nicity is an interesting research topic that has recently found practical uses in vari- ous areas. Given the paucity of annotated name data, we develop and make avail- able a new dataset containing 14k given names. Using this dataset, we take a data- driven approach to this task and achieve up to 90% accuracy for classifying the gender of unseen names. For ethnicity identifica- tion, our system achieves 83% accuracy. We also experiment with a feature analysis method for exploring the most informative features for this task. 1 Introduction The study of the structure and phonology of given names and how they associate with gender and ethnicity is a topic that has been investigated over the last several decades (Cassidy et al., 1999). Such research generally aims to identify discrim- inative features that are predictive of gender and how they can be applied in other areas. In addition to linguistic and psychological re- search, gender discrimination and inference has recently found practical uses in various areas. These include applications in NLP and text min- ing, as outlined in §2. These applications generally rely on name databases which can be costly to construct. How- ever, maintaining an exhaustive database is not always feasible for a number of reasons. Exact matching may not be possible due to spelling vari- ations, portmanteaus and self-created nicknames. Names found on the Web can also be combined or shortened, particularly in usernames. The main objective of this work is to evalu- ate a data-driven approach to inferring gender and ethnicity using machine learning and probabilis- tic models. Such methodology can help inform researchers about the structure of names and to build more robust name-gender inference systems for classifying unseen names. In line with these goals, we also develop a pub- licly available dataset of names annotated with gender and ethnicity information. 2 Related Work Linguistic research in name-gender relations has yielded some interesting results and demonstrated that phonology is quite informative about the gen- der of a name (Cassidy et al., 1999). It has also suggested that humans can often make gender at- tribution errors about other people due to the pho- netic characteristics of their names. In fact, through manual analysis, researchers have identified a slew of features strongly asso- ciated with gender (Barry Jr and Harper, 1995; Cutler et al., 1990), often from a linguistic view- point. One such example is that English female names are far more likely to end in a schwa vowel than male names (Slater and Feinman, 1985), most likely due to their Latin etymology. It is believed that humans implicitly learn these phonological cues and gender associations through exposure. Names are also a useful demographic indica- tor …

Content truncated to 500 words