Introduction
Authorship verification—the task of determining if two documents were written by the same individual—is a critical component of forensic linguistics and plagiarism detection. In these high-stakes domains, the "why" behind a prediction is often as significant as the prediction itself. Currently, practitioners face a difficult trade-off: high-accuracy "black-box" neural models that lack transparency, or interpretable stylometric models that use human-traceable linguistic features but often underperform in raw accuracy. This paper introduces Residualized Similarity (RS), a framework designed to bridge this gap by achieving state-of-the-art performance while maintaining faithful, human-verifiable explanations.
The Residualized Similarity (RS) Framework
The core innovation of this work is the shift from viewing neural networks as end-to-end solvers to viewing them as "error correctors." The RS approach utilizes a two-step process to generate a final authorship prediction:
- Interpretable Base: An initial similarity score is calculated using human-understandable linguistic features (e.g., part-of-speech frequencies, sentence length) via systems like Gram2vec or ELFEN.
- Neural Correction: A neural model (specifically LUAR, a state-of-the-art authorship model) is trained to predict the residual—the difference between the interpretable score and the ground truth.
The final prediction is the sum of the interpretable score and the neural correction. This design ensures faithfulness: because the final score is a literal sum, a human expert can see exactly how much the traceable features contributed versus the black-box component. To quantify this, the authors introduce Interpretability Confidence (INTCONF), measuring the relative weight of the interpretable features in the final decision.
Experimental Evaluation
The framework was evaluated across four diverse datasets to ensure cross-domain and cross-lingual robustness:
- Reddit Comments: Short, informal social media text.
- Amazon Reviews: Mid-length consumer feedback.
- Fanfiction Stories: Long-form creative writing.
- Pikabu Comments: A Russian-language dataset to test non-English performance.
Results and Analysis
The experimental results demonstrate that the RS framework successfully matches the performance of pure neural models while providing the necessary transparency for forensic applications. Notably, on the Amazon dataset, RS improved the interpretable baseline (Gram2vec) from an AUC of 0.71 to 0.90.
Key findings include:
- Performance Parity: RS matched the performance of pure neural models (like LUAR) across nearly all datasets, proving that interpretability does not require a sacrifice in accuracy.
- High Interpretability: On the Reddit dataset, the mean INTCONF was 0.65, indicating that the interpretable component provided the majority of the reasoning for the final predictions.
- Traceability: By focusing on features that a human can manually verify in the raw text (e.g., specific word usage or punctuation patterns), the system provides a concrete starting point for forensic investigations.
Conclusion
Residualized Similarity offers a path forward for deploying AI in sensitive legal and forensic contexts. By framing neural networks as supplements to—rather than replacements for—human-understandable logic, the framework provides a system that is both high-performing and inherently accountable. Future work will focus on refining the user interface for presenting these features to human experts and exploring the language-specific nature of the neural "corrections."
Summary updated