Demographically-aware computational humor

Abstract

Computational Humor is a subfield of humor research aimed at using computational methods to understand humor. This understanding can entail analysing large data-sets to find differences in humor appreciation, training systems that can detect and rate humor, or even generating jokes. In this thesis, we draw on findings from the broader field of Humor Research to build a large dataset of humor ratings. To capture responses to offensive humor, we also collect ratings of how offensive the funny texts were perceived to be, as well as demographic characteristics about the annotators who provided the ratings. We organised a humor and offense detection competition, and 60+ research groups submitted cutting edge computational systems to detect how humorous and offensive our data was. We then analysed our dataset’s ratings when grouped by age and gender, and found that men give higher humor ratings than women, and that both women and older people enjoy offensive humor less than other groups. Lastly, we built humor and offense detection systems which included the annotators’ demographic data, and found that incorporating demographic information as a textual description of the annotator improved how well models learned during training, but did not help the models to generalise to unseen data.