HonestBait: Forward References for Attractive but Faithful Headline Generation

“In this work, we propose making real news intriguing by learning what fake news is good at. We seek to learn what makes fake news eye-catching instead of simply mimicking the titles of fake news. Quantity-wise, the many circulating fake news articles serve as learning materials by which we can learn to generate more attractive headlines; stylewise, fake news is deliberately written to attract attention. To learn such attractive writing styles, we adopt the forward-reference (FR) writing technique (Blom and Hansen, 2015), which draws from psychology and journalism, and is frequently used to create attractive headlines. Specifically, FR creates an information gap between readers and the news content with the headline, motivating the reader’s curiosity (Loewenstein, 1994) to investigate the news content, and hence provoking the desire to click on the headline. One example is the headline “Wanna be an enviable couple? 12 things a happy couple must do... It’s that simple!”, which drives readers to find out what those things are.

Here, to understand the relation between veracity, attractiveness, and FR types in news headlines, we conducted a preliminary user study to investigate the attractiveness of fake and real news, and analyzed the FR types used in headlines in terms of veracity. Given these results and observations, we propose HonestBait, a novel framework by which to generate attractive but faithful headlines. In this framework, we use FR to remove the need to learn directly from the click-based dataset. To ensure the faithfulness of the generated headlines, we design a lexical-bias-robust textual entailment component on the generated headline and its original content to confirm that the content infers the headline. In addition, we propose PANCO, an innovative dataset which consists of pairs of fake and verified news headlines, their content, and their FR types.”