Transformer-based identification of stochastic information cascades in social networks using text and image similarity

Abstract

Identifying the origin of information posted on social media and how this may have changed over time can be very helpful to users in determining whether they trust it or not. This currently requires disproportionate effort for the average social media user, who instead has to rely on fact-checkers or other intermediaries to identify information provenance for them. We show that it is possible to disintermediate this process by providing an automated mechanism for determining the information cascade where a post belongs. We employ a transformer-based language model as well as pretrained ResNet50 model for image similarity, to decide whether two posts are sufficiently similar to belong to the same cascade. By using semantic similarity, as well as image in addition to text, we increase accuracy where there is no explicit diffusion of reshares. In a new dataset of 1,200 news items on Twitter, our approach is able to increase clustering performance above 7% and 4.5% for the validation and test sets respectively over the previous state of the art. Moreover, we employ a probabilistic subsampling mechanism, reducing significantly cascade creation time without affecting the performance of large-scale semantic text analysis and the quality of information cascade generation. We have implemented a prototype that offers this new functionality to the user and have deployed it in our own instance of social media platform Mastodon.