Decoding Online Hate in the United States: A BERT-CNN Analysis of 36 Million Tweets from 2020 to 2022

Since its inception, social media has enabled people worldwide to connect with like-minded individuals and freely express their thoughts and opinions. However, its widespread nature has not only had an immeasurable impact on society but also presented significant challenges. One such challenge is online hate speech. Consequently, the identification of hate speech has recently gained considerable attention, ranging from reactive methods, such as classifying individual posts, to proactive strategies that utilize contextual information to decipher the complex lexicon of online discussions. Despite these efforts, current research lacks a comprehensive analysis of hate speech on Twitter during the crucial 2020-2022 period, marked by significant events such as the COVID-19 pandemic. In this paper, we present a BERT-based model for classifying hate speech. To this end, we collected 36 million tweets posted in the United States on Twitter during this period. We developed, trained, and tested a BERT-based Convolutional Neural Network (BERT-CNN), using it to classify the collected tweets. The classification of this dataset revealed a high incidence of targets motivated by ethnicity, with gender and nationality as other prominent categories. This work provides insightful data on the sentiments of individuals across the United States during the events of 2020-2022.