Cross-domain authorship verification with feature interaction networks: Evaluating no-holdout and holdout protocols

Authorship verification (AV) across communication domains is a challenging task, as models must determine whether two texts are written by the same author despite variations in style, topic, and modality. In this work, we adopt a hybrid approach that combines contextual embeddings from RoBERTa with handcrafted stylometric features to capture both high-level semantic patterns and low-level stylistic cues. We evaluate our method using a cross-domain setup, considering both same-domain and strictly cross-domain pairs, including challenging “hard negatives” where authors write in markedly different styles across domains. To explicitly assess generalization to unseen domains, we further employ a leave-one-domain-out strategy, in which entire domains are excluded during training and used only for testing. Our experiments span multiple domain pairs including emails, text messages, essays, interviews, and speech-to-text data. Results show that the adopted approach consistently outperforms state-of-the-art research on comparable domain pairs, achieving higher overall verification scores while maintaining strong performance across individual metrics, improving the overall score over PAN 2022 from 0.600 to 0.654. Our findings demonstrate that combining contextual and stylistic representations, together with hard negative sampling, enables robust generalization across heterogeneous text types. Moreover, the leave-one-domain-out evaluation highlights the model’s ability to handle previously unseen domains, while analysis of same-domain test pairs shows that these representations also transfer to within-domain verification without explicit training on such pairs. This work provides a comprehensive evaluation of cross-domain AV under realistic conditions and offers insights into building more robust authorship verification systems.

• View at Publisher
• View at Scopus

Free Full Text ( Final Version , 3mb )