Emojis and emoticons are widely employed in user-generated content on social media. Existing generative models for short texts do not particularly handle emojis and emoticons thereby missing the extra information conveyed by these "special" expressions. We present EmDMM, a novel Dirichlet Multinomial Mixture model for capturing emotions expressed through emojis and emoticons in social media short texts. EmDMM can automatically detect emoji clusters that reflect emotion classes providing an unsupervised tool for analyzing emotions in rapidly-emerging social media content.We apply EmDMM to COVID-19 tweets and extract public emotions on topics related to the on-going pandemic.

PDF