Abstract

The proliferation of online hate speech requires a rigorous examination of the datasets used to train detection models. In this work, we analyze six Brazilian Portuguese datasets annotated for hate speech or toxicity to investigate how their lexical "anatomy" and domain characteristics affect cross-domain generalization. We combine HurtLex-based lexical profiling with cross-dataset evaluation in a feature-based transfer-learning setup, using BERTimbau embeddings and an XGBoost classifier. Our analysis shows that, although the datasets share a broadly similar macro-level focus, they diverge substantially in how specific terms are used and labeled across platforms and topics. Results indicate that lexical breadth and annotation practices strongly predict transferability: datasets with broader and more heterogeneous toxic vocabulary yield better cross-domain performance, whereas resources with narrow, profanity-centered labeling lead to severe generalization gaps, even when lexical overlap is high. These findings underscore the impact of collection and labeling strategies on the curation and evaluation of Portuguese hate speech datasets. Warning! This work and the referenced datasets contain examples of offensive and hateful language.