Seeking optimal Human/Machine collaborative practice in antisemitic terminology detection
Wendy Melillo, Jessica Emami, Solene Guarinos, Dhanush Kikkisetti, Melanie Klein, Lisa Liubovich, Raza Ul Mustafa, Nathalie Japkowicz · 2024
This study is concerned with the use of antisemitic language on loosely moderated extremist alt-right social media. It compares, contrasts, and combines human- and machinebased approaches for discovering antisemitic coded or noncoded terminology in such media. Coded terminology refers to the language used in closed-knit communities to communicate beliefs and attitudes without expressing them explicitly. In noncoded discourse, the views of the writer are expressed more explicitly. The detection and analysis of antisemitic terminology in social media is an important problem that can help uncover and monitor the evolution of societal attitudes towards the Jews, and eventually, towards other minority groups once we expand our study. Given the growing prevalence of antisemitic rhetoric in online public spaces, there is a greater need for academic studies to determine its presence and identify strategies to curb its influence. Scholars have noted how intolerant expressions can threaten democratic values. Antisemitic content, like other forms of hate speech, has the potential to radicalize people who may commit violence based on their prejudicial viewpoints. The purpose of this paper is to discuss three different approaches for discovering antisemitic terminology on social media. In particular, it describes an experiment in which three groups—human, software, and mixed—"competed” on the task of retrieving antisemitic terminology over a two-month period (January-February, 2024). The results were manually evaluated for relevance and frequency of occurrence in the next month period (March 2024). The study shows that each group discovered valuable terminology and that neither the human nor the machine group could replace the other. Instead, we advocate that both groups work together using their respective knowledge to optimize the discovery of new terminology on social media.