“Twitter has evolved into a significant communication nexus, coupling personal and highly contextual utterances with local news, memes, celebrity gossip, headlines, and other microblogging subgenres. If we take Twitter as a large and varied dynamic collection, how can we predict which tweets will be interesting to a broad audience in advance of lagging social indicators of interest such as retweets? The telegraphic form of tweets, coupled with the subjective notion of interestingness, makes it difficult for human judges to agree on which tweets are indeed interesting. In this paper, we address two questions: Can we develop a reliable strategy that results in high-quality labels for a collection of tweets, and can we use this labeled collection to predict a tweet’s interestingness? To answer the first question, we performed a series of studies using crowdsourcing to reach a diverse set of workers who served as a proxy for an audience with variable interests and perspectives. This method allowed us to explore different labeling strategies, including varying the judges, the labels they applied, the datasets, and other aspects of the task. To address the second question, we used crowdsourcing to assemble a set of tweets rated as interesting or not; we scored these tweets using textual and contextual features; and we used these scores as inputs to a binary classifier. We were able to achieve moderate agreement (kappa = 0.52) between the best classifier and the human assessments, a figure which reflects the challenges of the judgment task.”