Multimodal communication on tumblr
Elli E. Bourlai, Susan C. Herring · 2014
We manually analyzed a corpus of Tumblr posts for sentiment, looking at images, text, and their combination. A dataset was constructed of posts with both text and images, as well as a dataset of posts containing only text, along with a codebook for classifying and counting the content in each. This paper reports on the construction of the overall corpus and the codebook, and presents the results of a preliminary analysis that focuses on emotion. Posts containing images expressed more emotion, more intense emotion, and were more positive in valence than posts containing only text. The study contributes a micro-level analysis of multimodal communication in a social media platform, as well as a gold standard corpus that can be used to train learning algorithms to identify sentiment in multimodal Tumblr data.