Enhanced Authorship Verification for Textual Similarity with Siamese Deep Learning

Abstract

The internet is filled with documents written under false names or without revealing the author’s identity. Identifying the authorship of these documents can help decrease the success rate of potential criminals for financial or legal consequences. Most previous research on authorship verification focused on general text, but social media texts like tweets are more challenging since they are short, improperly structured, and cover a wide range of subjects. This paper proposes a new approach to determining textual similarity between these challenging messages. Inspired by the popularity of the Siamese networks in determining input similarity, four deep learning models based on this architecture were developed: a long-short-term memory (LSTM), a convolutional neural network (CNN), a combination of the two and a BERT model. These models were evaluated on a Twitter-based dataset, and the results show that the Siamese CNN-LSTM similarity model achieved the best performance with 0,97 accuracy.