A huge number of news artiles are distributed on soial media, and news onsumers an aess those artiles at any time from hand-held devies. This means that news providers are under pressure to nd ways of attrating the interest of news onsumers. One key fator that has a great impat on the attrativeness of a news artile is its headline as a way of guiding news onsumers to news artiles. This researh explores the hallenge of automatially generating attrative news headlines for soial media, and to this end we fous on the problem of identifying key sentenes that are useful for generating viral news headlines from a given news artile. We show that this problem an be formulated as supervised sequene labeling that utilizes user ativity on soial media as supervised information, and we propose a neural network model for this purpose. Investigations with our orpus onsisting of miroblog posts and news artiles demonstrate that lead sentenes believed to be the most suitable for news summaries do not neessarily ontribute to inreased virality whereas our proposed method an aurately identify key sentenes.