Offensive language detection has been well studied in many languages, but it is lagging behind in low-resource languages, such as Hebrew. In this paper, we present a new offensive language corpus in Hebrew. A total of 15,881 tweets were retrieved from Twitter. Each was labeled with one or more of five classes (abusive, hate, violence, pornographic, or none offensive) by Arabic-Hebrew bilingual speakers. The annotation process was challenging as each annotator is expected to be familiar with the Israeli culture, politics, and practices to understand the context of each tweet. We fine-tuned two Hebrew BERT models, HeBERT and AlephBERT, using our proposed dataset and another published dataset. We observed that our data boosts HeBERT performance by 2% when combined with D_OLaH. Fine-tuning AlephBERT on our data and testing on D_OLaH yields 69% accuracy, while fine-tuning on D_OLaH and testing on our data yields 57% accuracy, which may be an indication to the generalizability our data offers. Our dataset and fine-tuned models are available on GitHub and Huggingface.

翻译：攻击性语言检测在多种语言中已得到充分研究，但在希伯来语等低资源语言中仍相对滞后。本文提出了一个新的希伯来语攻击性语言语料库。我们从Twitter平台收集了15,881条推文，由阿拉伯语-希伯来语双语标注者对其逐一标注（每一条可能包含辱骂、仇恨、暴力、色情或无攻击性等五个类别中的多个标签）。标注过程具有挑战性，要求每位标注者熟悉以色列的文化、政治和社会实践才能理解每条推文的语境。我们基于所提出的数据集及另一公开数据集，对两个希伯来语BERT模型（HeBERT和AlephBERT）进行了微调。实验发现，将我们的数据集与D_OLaH结合使用时，HeBERT的性能提升了2%。当在O_数据上微调AlephBERT并在D_OLaH上测试时，准确率达到69%；而在D_OLaH上微调并在本数据集上测试时，准确率仅为57%，这或许表明我们的数据具有更强的泛化能力。本数据集及微调后的模型已发布于GitHub和Huggingface平台。