Zifei Fay Chen (Communication Studies, CAS), June Y. Lee (Entrepreneurship, Innovation, Strategy, and International Business, SOM), Shan Wang (Data Science, CAS), Diane Woodbridge (Data Science, CAS)
Since the outbreak of the COVID-19 global pandemic in early 2020, the Asian American and Pacific Islander (AAPI) communities have experienced an uptick of anti-AAPI discrimination, racism, and hate incidents. These hate incidents range from individual acts of shunning, verbal harassment, and physical attacks, to civil rights violations including refusal of service and workplace discrimination. According to Stop AAPI Hate, a total of 10,905 anti-AAPI incidents were reported from March 19, 2020 to December 31, 2021, causing significant detrimental impact on AAPI persons’ mental health. Importantly, scholars and activists have noted that these anti-AAPI hate incidents are not only associated with the anti-AAPI rhetoric during the COVID-19 pandemic, but also have their historical roots of anti-AAPI discrimination and racism, such as the “perpetual foreigners” and “Yellow Peril” stigmas, as well as the “model minority” myth that was used to delegitimaze and silence concerns from the AAPI communities.
To cope with racial trauma, many AAPI persons have turned to social media to view related content, share information, participate in online communities/forums, and join discussions. Besides being a coping tool, social media can also be a tool to advocate for counter-hate messages and facilitate social movement. Previous social media research has highlighted the role of emotion, where it was suggested that by engaging with emotion-carrying content on social media, people can better regulate their emotions and reconstruct their emotional episodes. In this Interdisciplinary Action Group project supported by CRASE, we set out to explore if and how emotion may drive engagement in counter-hate content on Twitter during the #StopAAPIHate movement.
We drew insights from the emotion theories, social media engagement literature, and used machine learning and computational methods to analyze data. To delineate a more nuanced understanding, we focused on two types of frameworks in our categorization and analyses of emotion: (1) the valence approach where emotion was categorized into positive, negative, and neutral, and (2) the discrete approach where emotion was further categorized into joy, sadness, anger, fear, surprise, and disgust.
Using Twitter API for Academic Research, we collected tweets between January 1, 2020 to August 31, 2021. The retrieval search criteria included 1) tweets written in English and 2) tweets with at least one of the following hashtags: #StopAAPIhate, #StopAsianhate, #IAmNotAVirus, #WashTheHate, #RacismIsAVirus, #IAmNotCovid19, #BeCool2Asians, and #HateIsAVirus, resulting in a total of 1,773,683 tweets.
To identify sentiment (negative, positive, and neutral) and the existence of six discrete emotions (joy, sadness, anger, fear, disgust, and surprise) in the tweets, we used the RoBERTa model, the robustly optimized Bidirectional Encoder Representations from Transformers (BERT) Pre-training Approach.
We then applied the developed models for sentiment/valence analysis and discrete emotion classification. Further, we developed a regression model and applied feature importance to understand the valence and discrete emotions affecting the level of engagement (likes, retweets, and replies) towards a tweet.
Emotional valence reflected in the counter-hate content on Twitter
Among all tweets collected, about 22.7% were negative, 25.3% were positive, and 52% were neutral. In this analysis, one tweet may only include one valence.
Discrete emotions reflected in counter-hate content on Twitter
Among the tweets collected, about 21.5% contained anger, 17.4% contained sadness, 11.9% contained joy, 5% contained disgust, 2.1% contained fear, and 1% contained surprise. In this analysis, a single tweet may include multiple emotions or no emotion.
The impact of emotional valence and discrete emotions on social media engagement
Emotional valence was a moderate predictor of the number of favorites, retweets, and replies, along with other tweet features including hashtag counts, referencing another tweet, multimedia attachment, and replying to other users. For discrete emotions, the emotion of anger, sadness, and disgust were predictors of the number of favorites, retweets, and replies, along with other tweet features including hashtag count, referencing another tweet, multimedia attachment, and replying to other users. Particularly for replies, joy was also shown as a predictor in driving the volume of replies.
Next Steps and Implications
The current stage of the project has demonstrated the features that drive engagement in counter-hate content on Twitter. For the next steps of the study, we will continue building models that inform the direction and magnitude of the effects specifically from each emotional valence and discrete emotion. Using these research insights, we hope to achieve a more in-depth and nuanced understanding of the effects from emotional valence and discrete emotions in driving further engagement in counter-hate content. And in order to achieve this engagement, we will provide greater empirical communication evidence from the large data set to further support the #StopAAPIHate movement on social media.