Cross-Language Offensive Speech Detection Using the mBERT Model

Aggressive speech can be detrimental to social stability. However, due to a lack of annotated data for aggressive speech, current automatic detection of aggressive speech focuses primarily on a few high-resource languages, making detection for low-resource languages difficult. We propose a cross-language, unsupervised, aggressive transfer detection method to address this. Firstly, we utilize a multilingual BERT (mBERT) model to learn aggressive features on a high-resource English dataset, resulting in an original model. Following that, by analyzing the linguistic similarity between English and low-resource languages like Danish, Arabic, Turkish, and Greek, we transfer the original model to these four low-resource languages, enabling automatic detection of aggressive speech. Experiment results show that, when compared to four other methods—BERT, linear regression (LR), multilayer perceptron (MLP), and support vector machine (SVM), our proposed method improves accuracy and F1 score by nearly 2 percentage points in Danish, Arabic, Turkish, and Greek. This method, which combines cross-language model transfer learning and transfer detection, shows promise in achieving unsupervised aggressive detection in low-resource languages, with performance comparable to current supervised detection methods.