Improving Hate Speech Classification with Cross-Taxonomy Dataset Integration
Jan Fillies, Adrian Paschke
TL;DR
The paper tackles the fragmentation of hate speech definitions across datasets by proposing a universal taxonomy and a single multi-label classifier trained through iterative fine-tuning on two differently annotated datasets. It encodes hierarchical label relationships and transfers annotations to a unified schema, achieving improved performance on an independent test set (F1 rising from 0.73 to 0.84). The approach reduces reliance on multiple niche models and supports privacy-preserving, federated learning since only model exchanges are needed. The work provides reproducible scripts and lays groundwork for expanding to broader taxonomies and ontology-based enhancements, with implications for more robust, context-aware hate speech detection across domains.
Abstract
Algorithmic hate speech detection faces significant challenges due to the diverse definitions and datasets used in research and practice. Social media platforms, legal frameworks, and institutions each apply distinct yet overlapping definitions, complicating classification efforts. This study addresses these challenges by demonstrating that existing datasets and taxonomies can be integrated into a unified model, enhancing prediction performance and reducing reliance on multiple specialized classifiers. The work introduces a universal taxonomy and a hate speech classifier capable of detecting a wide range of definitions within a single framework. Our approach is validated by combining two widely used but differently annotated datasets, showing improved classification performance on an independent test set. This work highlights the potential of dataset and taxonomy integration in advancing hate speech detection, increasing efficiency, and ensuring broader applicability across contexts.