Using Comments for Predicting the Affective Response to Social Media Posts

Yi-Chia Wang, Jane Dwivedi-Yu, Robert E. Kraut, Alon Y. Halevy

Abstract: What people see on social media influences their affective state. Predictions of the affective reaction of an audience to a post could help posters creating content and viewers searching for it. This paper examines the value of both real comments and artificially generated ones in predicting the affective responses of an audience. We built an affect prediction model based on Facebook anonymized public posts to predict affective responses (anger, amusement, and sadness affect) as indicated by three Facebook reaction clicks (Angry, Haha, and Sad). Using the content of the original post can predict reactions well (.71 to.87 F1-scores). Adding the text of real post comments improves F1-score by up to 11%. Surprisingly, generated comments improve predictions as much as real comments. These artificial comments were produced using a pre-trained sequence-to-sequence, BART natural language generation model given a post as input. Using artificial comments means that one can predict affect reactions early in the history of a discussion, before anyone has actually commented on a post.