Paraphrase-invariant stylometric projection for authorship attribution of texts

Abstract

Classical stylometric authorship classifiers, while highly accurate on original texts, suffer substantial accuracy drops when the input is rewritten by modern neural paraphrasers. Published authorship-attribution benchmarks typically conflate paraphrasing with topic shift, leaving the isolated effect of paraphrasing on attribution accuracy poorly quantified. This article isolates and measures that effect and proposes PISP-Light (Paraphrase-Invariant Stylometric Projection). The method applies singular value decomposition to the matrix of differences between the stylometric feature vectors of original and paraphrased training texts and projects the features into the subspace orthogonal to the dominant paraphrase-noise directions. In contrast to paraphrase augmentation, PISP uses paraphrases as a diagnostic instrument for discovering the nuisance subspace rather than as additional training samples. A reproducible Paraphrase Robustness Drop (PRD) metric is introduced to compare methods under paraphrase attacks. Experiments on the Guardian corpus (13 authors, 5 random seeds) show that a character n-gram classifier preserves 34 % of its original accuracy after paraphrasing, whereas a stylometric baseline loses about 80 %. Within the stylometric feature space, PISP with a fixed explained-variance threshold τ = 0.7 mildly improves robustness (PRD from 0.797 to 0.780) at the cost of a moderate decrease in original-test accuracy. A diagnostic analysis of the cumulative explained-variance curve shows that paraphrase noise in the stylometric space has only a partly low-rank structure, which sets a principled limit on linear projection. The results are relevant to developing software for authorship and information-source identification on the Internet that is robust to deliberate text transformations.

Keywords: software engineering, authorship attribution, stylometry, paraphrasing, robustness, singular value decomposition, neural text rewriting, computational stylistics, information-source identification.