Automated hate speech detection in a low-resource environment

Abstract

The problem of hate speech on social media is a growing concern. Much work has been done to tackle online hate including work into the automated detection of hate speech. The problem of automated hate speech detection at scale, however, remains by and large unsolved. This is in part due to the difficulty of classifying short texts without contextual information, difficulties in ensuring consistent annotation quality, contextual differences in different regions and social settings, and the informal and nuanced language used on social media. Automated detection of hate speech is made all the more difficult in low-resource regions for which large existing hate speech corpora are unavailable. Here, I present a sampling framework to tackle some of these challenges. The framework uses sequential data annotation phases, each allowing for the training of a hate speech filter that further refines our ability to collect useful data in subsequent phases. This framework is implemented for two phases on Twitter data collected around discourses in South Africa, and its efficacy assessed through a cross-dataset analysis between phases, as well as an analysis to determine the classification performance of decision tree-based methods on relatively small datasets. I conclude that this framework is a viable approach for curating hate speech corpora for automated hate speech detection in a low-resource setting.