論文アブストラクト：Social media sites are challenged by both the scale and variety of deviant behavior online. While algorithms can detect spam and obscenity, behaviors that break community guidelines on some sites are difficult because they have multimodal subtleties (images and/or text). Identifying these posts is often regulated to a few moderators. In this paper, we develop a deep learning classifier that jointly models textual and visual characteristics of pro-eating disorder content that violates community guidelines. Using a million Tumblr photo posts, our classifier discovers deviant content efficiently while also maintaining high recall (85%). Our approach uses human sensitivity throughout to guide the creation, curation, and understanding of this approach to challenging, deviant content. We discuss how automation might impact community moderation, and the ethical and social obligations of this area.