Exploring Boundaries and Intensities in Offensive and Hate Speech: Unveiling the Complex Spectrum of Social Media Discourse

TL;DR

This work argues that hate and offensive speech on social media exist along a continuum rather than as binary labels, and it presents a large Amharic benchmark dataset of 8,258 tweets annotated for category, hate targets, and intensity via Likert scales. Using a 70:15:15 data split and multiple transformer models, the study shows that modeling intensity as a regression task yields strong correlations (e.g., Pearson r up to $0.8022$) and that Afro-XLMR-large consistently outperforms baselines across category, target, and intensity prediction. The key contributions include comprehensive annotation guidelines, a multi-task annotation scheme, and the release of dataset, models, and code under a permissive license, enabling nuanced hate-speech analysis in a low-resource language and highlighting the practical value of continuous-spectrum detection for early interventions and conflict monitoring. The findings demonstrate that intensity-based approaches capture subtle variations in hatefulness and offensiveness, with implications for policy, moderation, and sociopolitical understanding in Ethiopia and similar sociocultural contexts.

Abstract

The prevalence of digital media and evolving sociopolitical dynamics have significantly amplified the dissemination of hateful content. Existing studies mainly focus on classifying texts into binary categories, often overlooking the continuous spectrum of offensiveness and hatefulness inherent in the text. In this research, we present an extensive benchmark dataset for Amharic, comprising 8,258 tweets annotated for three distinct tasks: category classification, identification of hate targets, and rating offensiveness and hatefulness intensities. Our study highlights that a considerable majority of tweets belong to the less offensive and less hate 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. We build classification and regression models and investigate the efficacy of models in handling these tasks. Our results reveal that hate and offensive speech can not be addressed by a simplistic binary classification, instead manifesting as variables across a continuous range of values. The Afro-XLMR-large model exhibits the best performances achieving F1-scores of 75.30%, 70.59%, and 29.42% for the category, target, and regression tasks, respectively. The 80.22% correlation coefficient of the Afro-XLMR-large model indicates strong alignments.

Figures (4)

• Figure 1: Potato GUI for the three types (1 - category, 2 - intensity, and 3 - target) of annotation tasks.
• Figure 2: Mapping the dataset in an 11-point Likert rating scale.
• Figure 3: Distributions of 0-10 rating labels.
• Figure 4: Confusion matrix from Afro-XLMR-large.