Authorship Attribution Using Multilingual BERT on Cross-Platform Social Media Texts

Abstract

Authorship attribution is a technique for identifying the author of a piece of written text based on linguistic and stylistic characteristics. In recent years, with the rise of social media platforms, authors are increasingly writing short and informal texts in multiple platforms and languages, which is causing significant challenges for authorship attribution using traditional stylometry methods. In this paper, a transformer-based authorship attribution system using multilingual Bidirectional Encoder Representations from Transformers (mBERT) is proposed for authorship attribution on cross-platform social media platforms. The proposed system utilizes contextual embeddings of mBERT for authorship attribution, as it can identify syntactic and semantic writing patterns of authors in different social media platforms. The proposed system is tested using a multilingual social media dataset collected from Twitter, Facebook, and Reddit platforms, and the results show significant improvements in accuracy and F1-score compared to traditional machine learning and monolingual deep learning-based authorship attribution methods for cross-platform social media platforms.