HatEmoTweet: Low Level Emotion Classifications and Spatio-Temporal Trends of Hate and Offensive COVID-19 Tweets

Social media platforms (like Twitter) positively and negatively impact users in diverse societies; one of Twitter’s negative effects is the usage of hate and offensive language. Hate speech fosters prejudice; it also harms the vulnerable. There are always emotions associated with hateful and offensive actions. This work addressed hate and offensive tweet detection, low-level emotional classifications using 28 labels to train transformers models in three ways (model 1-BERTG28, model 2-BERTG27 and model 3-RoBERTaG27) before predicting the hateful and offensive tweets emotions. Model 1 was trained on low-level labels, and model 2 and 3 were trained on 27 labels excluding the neutral label. This study performed topic modeling to extract the discussed theme, spatiotempo-ral trends to determine where and when these tweets occurred, and event summarization for identified hate and offensive tweets. GoEmotions and Ekman were used for direct and indirect assessment, respectively, to evaluate the model’s precision, recall, and F1-score. In terms of precision evaluation, the model 1 outperformed Google Research on GoEmo-tions. Furthermore, this study’s model 2 and 3 outperformed the Google research on both the GoEmotions and Ekman’s evaluation in terms of precision and F1-score. Generally, model 2 was the best model in the Social Network Analysis and Mining analysis for both recall and F1-score while model 3 performed better for precision. Due to the training on samples without the neutral label, model 2 obtained 27% and model 3 achieved 29% label prediction out of the 30% neutral samples that was predicted in model 1 for hate and offensive tweets. This is a significant improvement to optimize classified emotions that are not truly neutral by eliminating the false neutral classes.