Investigating Hate Speech Beyond Detection and Classification: Uncovering Complex Intensities and Targets
Keywords: hate speech, offensive, intensity, target, social media
TL;DR: This paper presents a new hate speech dataset for three tasks namely; detecting its category, targeted groups, and rating intensities of hatefulness and offensiveness in tweets utilizing Africentric transformer models.
Abstract: Despite the complex nature of hate speech, studies focus primarily on detecting its binary categories, often overlooking the continuous spectrum of offensiveness and hatefulness inherent in the message. This study presents benchmark datasets for Amharic, comprising 8,258 tweets annotated for three distinct tasks: category classification, identification of hate targets, and rating of offensiveness and hatefulness intensities. Our study highlights that a considerable majority of tweets belong to the less offensive and less hateful intensity levels, underscoring the need for early interventions by stakeholders. The prevalence of ethnic and political hatred targets, with significant overlaps in our dataset, emphasizes the complex relationships within Ethiopia's sociopolitical landscape. This study revealed that hate and offensive speech cannot be addressed by simplistic binary classification methods. Instead, they manifest themselves as variables across a continuous range of values. The Afro-XLMR-large model exhibits the best performance, achieving F1 scores of 75. 30\%, and 70. 59\% for the category and target classification tasks, respectively. The 80.22\% correlation coefficient of the Afro-XLMR-large and Afro-XLMR-large-with-active-learning models exhibits strong alignments in the regression tasks.
Submission Number: 57
Loading