Virality Prediction for News Tweets Using RoBERTa

The virality of a tweet is essential to convey its message to a broader audience and, eventually, to generate influence. This is especially important for news outlets as they struggle to transition from traditional media to online formats. As their usual readers will not migrate directly to digital news outlets need to gather new audiences from the spaces where real-time information and discussions are happening; this is Social Media and in particular Twitter. Since the news websites and Twitter languages differ greatly news outlets need to write their tweets properly to maximize their impact on Twitter. We propose a method to predict if a tweet will be influential or not influential based on its text using a variant of Google BERT named RoBERTa, and a corpus of 5000 high-quality and automatically labeled highly-influential and non-influential tweets to train and classify tweets in these categories. Our method reaches an F1 of 0.873, improving 4 and 9 over approaches using LSTMs and n-grams respectively.