Combining style and semantics for robust authorship verification

Abstract

Authorship Verification is a key task in Natural Language Processing, essential for applications like plagiarism detection and content authentication. This paper analyzes the use of deep learning models for Authorship Verification, focusing on combining semantic and style features to enhance model performance. We propose three models: the Feature Interaction Network, Pairwise Concatenation Network, and Siamese Network, which aim to determine if two texts are written by the same author. Each model uses RoBERTa embeddings to capture semantic content and incorporates style features such as sentence length, word frequency, and punctuation to differentiate authors based on writing style.

Our results confirm that incorporating style features consistently improves model performance, with the extent of improvement varying by architecture. This demonstrates the value of combining semantic and stylistic information for Authorship Verification. While limitations such as RoBERTa’s fixed input length and the use of predefined style features exist, they do not hinder model effectiveness and point to clear opportunities for future enhancement through extended input handling and dynamic style feature extraction.

In contrast to prior studies such as Bevendorff et al., (2020) and Kestemont, et al., (2022), which relied on balanced and homogeneous datasets with consistent topics and well-formed language, our work evaluates models on a more challenging, imbalanced, and stylistically diverse dataset, better reflecting real-world Authorship Verification conditions. Despite the increased difficulty, our models achieve competitive results, underscoring their robustness and practical applicability.

These findings support the value of combining semantic and style features for real-world Authorship Verification.