In online social networks, users openly interact, share content, and endorse each other. Although the data is interconnected, previous research has primarily focused on modeling the social network behavior separately from the textual content. Here we model the data in a holistic way, taking into account connections between social behavior and content. Specifically, we define multiple decision tasks over the relationships between users and the content generated by them. We show, on a real world dataset, that a learning a joint embedding (over user characteristics and language) and using joint prediction (based on intra- and inter-task constraints) produces consistent gains over (1) learning specialized embeddings, and (2) predicting locally w.r.t. a single task, with or without constraints