Modelling Social Media Popularity of News Articles Using Headline Text

The way we formulate headlines matters -- this is the central tenet of this thesis.

Headlines play a key role in attracting and engaging online audiences. With the increasing usage of mobile apps and social media to consume news, headlines are the most prominent -- and often the only -- part of the news article visible to readers. Earlier studies examined how readers' preferences and their social network influence which headlines are clicked or shared on social media. However, there is limited research on the impact of the headline text on social media popularity.

To address this research gap we pose the following question: how to formulate a headline so that it reaches as many readers as possible on social media. To answer this question we adopt an experimental approach to model and predict the popularity of news articles on social media using headlines. First, we develop computational methods for an automatic extraction of two types of headline characteristics. The first type is news values: Prominence, Sentiment, Magnitude, Proximity, Surprise, and Uniqueness. The second type is linguistic style: Brevity, Simplicity, Unambiguity, Punctuation, Nouns, Verbs, and Adverbs. We then investigate the impact of these features on popularity using social media popularity on Twitter and Facebook, and perceived popularity obtained from a crowdsourced survey. Finally, using these features and headline metadata we build prediction models for global and country-specific social media popularity. For the country-specific prediction model we augment several news values features with country relatedness information using knowledge graphs.

Our research established that computational methods can be reliably used to characterise headlines in terms of news values and linguistic style features; and that most of these features significantly correlate with social media popularity and to a lesser extent with perceived popularity. Our prediction model for global social media popularity outperformed state-of-the-art baselines, showing that headline wording has an effect on social media popularity. With the country-specific prediction model we showed that we improved the features implementations by adding data from knowledge graphs.

These findings indicate that formulating a headline in a certain way can lead to wider readership engagement. Furthermore, our methods can be applied to other types of digital content similar to headlines, such as titles for blog posts or videos. More broadly our results signify the importance of content analysis for popularity prediction.