The dataset is based on 81,446 WikiArt (WikiArt is an online, user-editable encyclopedia of visual art) artworks, and contains 439,121 comments from more than 6,377 human users, indicating the emotion conveyed by a painting (including 8 emotions: anger, disgust, fear, sadness, amusement, awe, satisfaction, and excitement), and explaining why they chose a certain emotion. The comments form a corpus with a total of 36,347 different words.

The ArtEmis dataset aims to enable computers to understand in detail the interaction between visual and expressive content, their emotional effects, and to explain why the corresponding emotional effects occur.

Based on this data, the dataset publisher trained and demonstrated a series of annotation systems that can express and interpret emotions from visual stimuli. The annotations generated by these systems for artistic paintings generally successfully reflect the semantic and abstract content of the paintings, far exceeding the systems trained on existing datasets.

The dataset was published by Stanford University, École Polytechnique, and King Abdullah University of Science and Technology.

"ArtEmis: Affective Language for Visual Art"