The distant supervision assumes that if entity e corresponds to an attribute for user i, at least one posting from user i’s Twitter stream containing a mention of e might express that attribute.

We separately trained three classifiers regarding the three attributes. All variables are observed during training; we therefore take a feature-based approach to learning structure prediction models inspired by structure compilation (Liang et al., 2008). In our setting, a subset of the features (those based on network information) are com- puted based on predictions that will need to be made at test time, but are observed during train- ing. This simplified approach to learning avoids expensive inference; at test time, however, we still need to jointly predict the best attribute values for friends.