Deep Learning based Multilabel Hateful Speech Text Comments Recognition and Classification Model for Resource Scarce Ethiopian Language: The case of Afaan Oromo

In response to the phenomenon of hate speech on popular social media such as Facebook, a number of researchers have investigated and developed different automated techniques in order to detect and moderate hate speech. Reliable hate speech detection tools ranging from traditional machine learning models to deep neural network models have been developed and deployed for resource rich languages such as English, French and Arabic. On the Other hand, the topic of building automated techniques for hate speech detection in Afaan Oromo is still the subject of very limited studies. In this study, we have used pretrained word embedding approaches such as word2vec and fast text, and assessed their effectiveness for hateful speech recognition and classification in Afaan Oromo text comments on Facebook. By using the proposed model, the effectiveness of deep learning algorithms like CNN and LSTM was also assessed. Moreover, the significance of one of the recent topic modeling approaches, i.e., BERTopic, has been investigated to in extracting topics that can help the categorization of Afaan Oromo hate speech. Afaan Oromo Multilabel Hateful Speech Detection and Dataset (AOMLHSDD), which contains 30728 text comments, was created after we employed the BERTopic to extract relevant topics. Accordingly, Then, we have conducted different experiments to test the effectiveness of CNN and LSTM models by using word2vec and fast text. The result of our experiments has shown showed that LSTM has outperformed the CNN approach by achieving the accuracy of 98.32%.