Hate Speech Detection Using Textual and User Features

Social media platforms provide users with a powerful platform to share their ideas. Using one’s right to expression to incite hatred toward a particular group of people is inappropriate. However, hate speech is pervasive in our society. Spreading hate through online social networks like Facebook, Twitter, Tiktok, and Instagram is commonplace in today’s milieu. One such case is the unprecedented COVID-19 pandemic, which engendered anti-Asian hate. In current literature, there is limited study on using user features in conjunction with textual features to detect hate. This thesis aims to combine textual features with user features to improve the state-of-the-art hate speech detection technique. To test our approach, we used four different datasets available in the public domain. We have used various tools to access Twitter APIs to extract required user information, either to use directly or further compute other features using that information. We have represented the textual features in the form of BERT embeddings and linguistic features. The 97 linguistic measures computed with a Linguistic Inquiry and Word Count (LIWC) tool quantify the text’s cognitive, affective, and grammatical processes. The user feature consisted of demographic, behavioral-based, emotion-based, personality, readability, and writing style features. Our experimental evaluation over three datasets shows that the top twenty linguistic features and the top twenty user features are the best combinations for hate speech detection. Hate speech is mostly emotionally charged. We further analyzed these user and linguistic features. Among the most intuitive and prominent results was that features like anger, negative emotion, swearing, fear, and annoyance were high in hate speech, while the happiness feature was low. We compared multiple approaches along with the existing state-of-the-art. We found that the best approach with textual features was combining LIWC features with BERT embeddings. This combination gave us the F1 of 0.82 and 0.79 on Crowd-sourced (DS1) and Kaggle (DS3), respectively. Followed by this, we identified the top LIWC and user features for hate speech detection. We found that features representing negative emotions like anger, fear, sadness, and annoyance were prominently high in hate speech. Happiness is lower in hate speech. After this, we analyzed the F1 scores with standalone LIWC and user features. We also used their combinations. We found that the combination of the top twenty LIWC and top twenty user features gives the best F1 scores of 0.74, 0.90, and 0.64 on DS1, NAACL (DS2), and anti-Asian Covid hate (DS4) dataset. Finally, we used traditional machine learning algorithms combining BERT embeddings with the top twenty linguistic features and the top twenty user features. We obtained the F1 scores of 0.78, 0.92, and 0.84 on DS1, DS2, and DS4 respectively. We also compared our approach with other studies using user and textual features