Intent to Hate

A study on understanding motivations behind hate speech on Twiter/X during the COVID-19 pandemic

Disclosure: In regard to the sensitive nature of the data, we have chosen to keep the data used in this study private.

Introduction

The COVID-19 pandemic ushered in an influx on social media. As our world entered lockdown, our society tuned in online, and with the good, the bad elements of societal behaviours followed suit. Research [1,2] followed the rise of short content format, with the increase in the number of users of platforms such as TikTok, Instagram Reels, and Twitter. One of the major drawbacks of this short content format presentation was that creators had a few seconds to grab the attention of an audience that could swipe up on them in no time. The need to stand out and pander to the biased attitudes of those who were forced to stay in or be unemployed led to an increase in the spread and availability of misinformation and hate. 

Hate during the COVID-19 pandemic was primarily targeted towards Asians. The hate which originated online, continued like the pandemic, with physical assaults towards members of the asian community. Important research [4,6,7,8,9] has been conducted, identifying sources of hate on text-based platforms, studying the linguistic, and lexical features of the hate speech, and conducting temporal, network, and geopolitical analysis, to further understand and curb this spread on social media. 

With the advent of large language models, the aim of this research article is to understand another dimension of hate, the intent. We have seen possible reasons for systemic hate, that is hate that is powered by the platform and organisation, we wish to dive deep into the intrinsic motivators behind hate. We want to understand the method, mode and target behind the hate speech. To do so, we chalk three categories that can help us understand the intent behind hate speech. We study this in the context of hate speech on X (formerly Twitter)

Codebook

Who does that hate speech target? 

We ask this question to understand whether hate is communal or individualistic. When a person posts a hate tweet, is the intent primarily to hurt an individual, and incidentally offend a community, or is the hate principally intended to endanger the community? 

How is the tweet targeting?

While reading through tweets, one of the biggest sticklers was how tweets that contained disinformation about the origin of COVID-19, or vaccines, were often encouraged with comments laden with hate speech. On the other end, this disinformation morphed into stereotypes that attacked the identity and culture of these communities[5]. Intent to spread hate came from this question, is the intent to spread disinformation or is the intent to regurgitate stereotypes? 

Why does the tweet contain hate?

In the paper “Why We Hate” [3], the authors shed light on the origin of hate. They classify the intent of hating, broadly into categories, either to be offensive or to retaliate and be defensive. Was hate speech on Twitter born in defence of one’s community or livelihood, or was it perpetrated to offend another community, for no personal sake? 

The advent of large language models has made research more accessible and available to solve tasks without the need for large data. We use GPT4 to build a few shots and a chain-of-thought prompting model. We provide the model with the definition of three categories: method, mode and target, and the options that it has to classify each of them, as shown below:

Experiments and Results

We give the model a few examples (n=10), to train it to adapt to the codebook we defined, and also output the results in a particular format. Additionally, we ensure it is a chain-of-thought modelling, that is the model produces reasoning for its decisions to assign a class for each feature. This reasoning is to ensure that the model doesn’t hallucinate and is able to produce decisions made on logic and the training data. We collected data from Twitter (now X) using the redundant academic API and found the following result. 

Regarding the model performance, the model is accurate with its responses. First, n = 100, tweets that we tested on, 6 of them had to be discarded because of hallucinations by the model in the reasoning. These hallucinations were primarily the model reusing an explanation, word to word, for two different tweets. While parsing through the results, we also noticed that even though the model was given exactly two classes for each feature, it contributed more, and added a class of its own: ‘Others’, for Target and Method. In an effort to maintain integrity for the model’s reasoning, we don’t discard these outputs.  

Two human annotators verified the results of the model and had a disagreement score of 0.12 with the model. Most disagreements between the human annotators and the model were in the inability of the model to identify the target correctly.

The results are as follows:

Conclusion

Hate targetted communities. Very little hate was intended to question individuals and those that did, hated political leaders and people in positions of power. Hate was equally internalised, harping on stereotypes, and reducing one’s culture and traditions, and equally intended to spread disinformation and fake news. Hate was often offensive, with the aim of blaming a community, for causing indirect harm to their livelihood and lifestyle. 

While the classifier is able to inform us of the motivations behind hate, research should also focus on finding solutions and interventions to reduce and stop this propagation and spread of hate, irrespective of the context and situation. 

References

[1] Feldkamp, J. (2021). The Rise of TikTok: The Evolution of a Social Media Platform During COVID-19. In: Hovestadt, C., Recker, J., Richter, J., Werder, K. (eds) Digital Responses to Covid-19. SpringerBriefs in Information Systems. Springer, Cham. https://doi.org/10.1007/978-3-030-66611-8_6

[2] Kim, Antino and Dennis, Alan R., Says Who? The Effects of Presentation Format and Source Rating on Fake News in Social Media (August 16, 2018). MIS Quarterly, Vol. 43, No. 3, pp. 1025–1039 (2019), Available at SSRN: https://ssrn.com/abstract=2987866 or http://dx.doi.org/10.2139/ssrn.2987866

[3] Fischer, A., Halperin, E., Canetti, D., & Jasini, A. (2018). Why We Hate. Emotion Review, 10(4), 309-320. https://doi.org/10.1177/1754073917751229

[4] He, Bing, et al. "Racism is a virus: Anti-Asian hate and counterspeech in social media during the COVID-19 crisis." Proceedings of the 2021 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining. 2021.

[5]Kumar, Srijan & Shah, Neil. (2018). False Information on Web and Social Media: A Survey.

[6] Mathew, B., Kumar, N., Ravina, Goyal, P., & Mukherjee, A. (2018). Analyzing the hate and counter-speech accounts on Twitter. ArXiv, abs/1812.02712.

[7] Gover, A.R., Harper, S.B. & Langton, L. Anti-Asian Hate Crime During the COVID-19 Pandemic: Exploring the Reproduction of Inequality. Am J Crim Just 45, 647–667 (2020). https://doi.org/10.1007/s12103-020-09545-1

[8]Horse, Aggie J. Yellow. "Anti-Asian racism, xenophobia and Asian American health during COVID-19." The COVID-19 crisis: Social perspectives. Taylor and Francis, 2021. 195-206.

[9] Chen, H. Alexander, Jessica Trinh, and George P. Yang. "Anti-Asian sentiment in the United States–COVID-19 and history." The American Journal of Surgery 220.3 (2020): 556-557.