Generalizability of Abusive Language Detection Models on Homogeneous German Datasets

Abstract Abusive language detection has become an integral part of the research, as reflected in numerous publications and several shared tasks conducted recent years. It been shown that obtained models perform well on datasets which they were trained, but have difficulty generalizing to other datasets. This work also focuses model generalization, – contrast previous we use homogeneous for our experiments, assuming a higher generalizability. We want find out how similar be trained generalize whether generalizability depends method used obtain model. To this end, selected four German from popular tasks, three are consecutive GermEval tasks. Furthermore, evaluate two deep learning methods traditional machine derive trends based results. Our experiments show generalization is only partially given, although annotation schemes these almost identical. findings additionally solely (combinations of) training sets consistent no matter what underlying is.