Methodological Challenges for Detecting Interethnic Hostility on Social Media

• In Event: Methodological Challenges of Hateful and Problematic Content Detection Online

Abstract

In this talk we share methodological details from the study on detection of ethnicity-related opinions in the Russian-language social media. We first focus on difficulties of defining ethnicity and ethnicity-related opinions to be detected by differentiating them from political conflict and international relations. We then reflect on advantages and limitations of unsupervised and supervised machine learning for such tasks as finding ethno-relevant texts, extraction of contexts in which ethnic groups are placed, and detection of various aspects of text authors’ attitudes. The latter include questions on whether an ethnic group is presented as inferior or superior, as an aggressor or a victim and others. We base our conclusions on the sample of 2.6 million user messages of which 14,000 have been hand-coded, and on our experience of applying topic modeling, classification and simpler keyword-based approaches to these data.

Authors

• Olessia Koltsova, National Research U Higher School of Economics

• Sergei Koltcov