Using Authorship Embeddings to Understand Writing Style in Social Media

Abstract

With the escalation of misinformation and malicious behavior issues on social media platforms, traditional detection-based measures often fail to address the problem in time. The use of multiple accounts or the continuous creation of new accounts makes it difficult to re-detect the presence of a user who, for example, has disseminated false information. In this paper, we present a novel approach to understanding and characterizing authorship in social media using a model called PARTSCL, an improvement of the previous PART model. PARTSCL generates “authorship embeddings”, numerical representations of an author’s writing style, allowing for more accurate and earlier detection of malicious behavior. Our main contributions include the PARTSCL model itself, a new pre-training approach for authorship attribution, and the application of our model on different datasets. These advances help bridge the gap between popular Natural Language Processing techniques such as Transformers and feature engineering, providing a robust tool for the ongoing fight against online misbehavior and misinformation.