When the same person writes a news article and a tweet, almost everything on the surface changes: vocabulary, topic, length, even sentence structure. Authorship models trained on one kind of text therefore collapse on the other. On the CROSSNEWS benchmark, which pairs journalists' articles with their own tweets, non-transformer verifiers fall from about 70% accuracy on article-article pairs to about 50% on article-tweet pairs (CROSSNEWS). The usual remedy is to hide the words that carry topic (names, content words, rare keywords) and let the model fall back on grammar. We argue this fix is incomplete. Once topic is masked, what remains is mostly the grammar of the platform. Telegraphic tweet syntax and nested article syntax differ far more between genres than between authors, and documents are known to cluster by genre rather than by author (When Writing Style Drifts).
This proposal tests a specific hypothesis. Author identity survives a change of platform not as an absolute syntactic profile but as a deviation from that platform's norm. A journalist who uses more subordinate clauses than the average journalist should also use more of them than the average tweeter, even though both counts fall sharply on Twitter. We build an author representation from delexicalized syntax, meaning text in which entities, platform markers and content words are replaced by category labels. Each profile is standardized against a population of other authors on the same platform. We then ask whether this recovers cross-platform authorship signal that masked neural embeddings miss.
Core Claim
Masking entities and topic words removes the topical shortcut but leaves a second shortcut in place: platform syntax. Standardizing delexicalized syntactic statistics per platform removes it, and what is left transfers across platforms because the author's idiosyncrasy is rank-preserved. Concretely, let $f_{a,g}$ be the smoothed frequency vector of delexicalized n-grams for author $a$ in genre $g$. Here n-grams are part-of-speech tags interleaved with literal function words and punctuation. Let $\mu_g$ and $\sigma_g$ be the per-feature mean and standard deviation over a disjoint reference population of authors in genre $g$. The platform-relative profile is
$$d_{a,g} = \frac{f_{a,g} - \mu_g}{\sigma_g},$$
and a cross-genre verification score for a candidate pair is $s(a,b) = \cos(d_{a,\text{article}}, d_{b,\text{tweet}})$. The claim is falsifiable at two levels. First, for the same author, per-feature deviations should correlate across genres well above what shuffled author pairings produce. Second, adding $d$ to a strong neural verifier should raise article-tweet accuracy most on pairs built so that topic misleads.
Relation to Prior Work
Two papers motivate the masking half of the idea, and both stop short of the platform confound. Learning Universal Authorship Representations trains contrastive authorship embeddings across Amazon, Fanfiction and Reddit and masks up to 60% of subwords at random, gaining about 2.6 Recall@8 points on Amazon-to-Reddit transfer. Its masking is unguided and contains no syntactic structure. Can Authorship Representation Learning Capture Stylistic Features? masks content words (PertLE) or high-TF-IDF words (TertLE) and finds that Reddit mean reciprocal rank barely moves, from 0.360 to 0.329. That study evaluates each dataset separately and never asks whether the surviving signal is authorial or platform-level. We reuse PertLE and TertLE as masking operators and ask exactly that question.
The syntactic half comes from stylometry. POS n-grams stay stable when author-topic pairings are inverted in the Topic Confusion Task, while BERT-style models fall to 50–60% (The Topic Confusion Task). POS patterns also capture style in informal blogs (Improving Gender Classification of Blog Authors), and function-word transition networks carry an authorial fingerprint (Function Word Adjacency Networks). All of these work within a single formal corpus. None of them standardizes against a genre population, which is the step we add.
Several findings constrain the design rather than supply it. Function-word networks need on the order of 60,000 words per author, and CROSSNEWS Gold offers roughly 100 tweets per author. We therefore use smoothed short n-grams, not full transition matrices. Parse-path encodings reach 83.48% on IMDB62 using syntax alone (Syntax Encoding with Application in Authorship Attribution), but they depend on treebank-trained parsers that degrade on tweets. We keep them only as an article-side ablation. High-order character n-grams quietly absorb topic (The Power of Character N-grams), so we exclude them from the syntactic channel. Delexicalization can strip short social texts nearly bare (Enhancing Portuguese Variety Identification), so we pool at the author level rather than classify single tweets.
Two findings shape the masking itself. Detectors are known to exploit platform tokens such as #sarcasm (MMSD2.0), so hashtags, mentions, URLs and retweet markers get their own masking layer. Delexicalized distant supervision has already transferred across platforms, from Reddit to Silk Road, for jargon detection (JEDIS). That work concerns single tokens, not authors.
Research Program
We build a layered delexicalizer, a platform-relative syntactic profiler, and a fusion verifier.
The delexicalizer applies masking layers in a fixed order. It starts with rule-based platform markers, which need no tagger. Next it replaces named entities with type labels. It then maps content words to POS tags in the PertLE style, or alternatively applies TertLE's high-TF-IDF masking as a language-agnostic, tagger-free variant. Function words and punctuation are kept literally. Tagging runs after platform markers are masked, which removes one known source of tagger error on tweets. We also report tagger agreement between two taggers on a hand-checked tweet sample, so readers can see how noisy the syntactic channel is.
The profiler computes $d_{a,g}$. The reference populations $\mu_g, \sigma_g$ come from the CROSSNEWS Silver Set, restricted to authors disjoint from all evaluation authors, so the normalization never sees test identities.
The fusion verifier combines $s(a,b)$ with the score of a neural embedding verifier through logistic regression. The fusion weights are fit on a development split.
The analysis half measures what each masking layer removes. It extends the Topic Confusion Task to verification by using CROSSNEWS topic annotations to build topic-adversarial pairs: same-author pairs share no topic, and different-author pairs share one. It then decomposes errors into those explained by topic overlap and those explained by platform-marker density.
Baselines
The naive reference point is an unmasked character and word n-gram verifier of the kind CROSSNEWS reports, at about 50% article-tweet accuracy. We run it on our exact pair splits.
The strongest existing method is SELMA, the instruction-tuned e5-mistral-7b-instruct style embedding that CROSSNEWS reports at about 78.5% article-tweet verification accuracy and about 85.8% article-article (CROSSNEWS). It is the strongest competitor because it is the best reported cross-genre system on the exact benchmark. We re-run it on our splits rather than quoting the figure, and we give it the same number of development-set calibration choices as our fusion model. Beating it means that SELMA plus $d$ improves article-tweet accuracy by a margin whose bootstrap 95% interval excludes zero, with the gain concentrated on topic-adversarial pairs. An improvement only on ordinary pairs would suggest the extra channel is merely adding capacity.
We also run LUAR with random subword masking (Learning Universal Authorship Representations) as the best contrastive model that uses masking. It gets the same masking layers we use, so any gap reflects the representation rather than the masking.
The key ablations isolate each ingredient. The first uses raw delexicalized profiles $f_{a,g}$ without platform standardization, which tests whether standardization is doing the work. The second applies per-genre mean-centering and whitening directly to SELMA embeddings. This is the strongest simple competitor to the whole idea and is discussed below. The third drops the platform-marker masking layer to measure how much signal comes from hashtag and mention habits.
Why not just…?
Why not just mean-center the neural embeddings per platform? This is the real threat, because it applies the same "deviation from platform norm" logic to a representation that is already strong. If genre-centered SELMA matches SELMA plus $d$, then the syntactic channel adds nothing and the contribution reduces to a normalization trick. We will report that as the finding. Our reason to expect a gap comes from Wang et al.'s paraphrase result: paraphrasing collapses authorship-embedding MRR on Fanfiction from 0.325 to 0.139 (Can Authorship Representation Learning Capture Stylistic Features?), which suggests that learned embeddings still lean on lexical form. Centering removes a platform offset, but it cannot remove topic that is correlated within an author. The topic-adversarial split is where these two accounts diverge.
Why not just fine-tune a large model on cross-genre pairs? Kiefer et al. report that LoRA-tuned Gemma reaches 0.77 F1 in mixed-domain cross-genre verification on AVShift and gains up to 0.11 F1 over prompting on CrossNews (When Writing Style Drifts). That is a strong alternative. However, fine-tuning on the benchmark's own genre pair teaches that pair, and the same paper shows sharp decay under temporal shift. We will fine-tune a smaller model under our compute budget. We then test every system on a held-out transfer direction it was not tuned on (Reddit/Amazon/Fanfiction), where a platform-standardized profile needs only a new reference population rather than retraining.
Why not just learn disentangled style and content spaces with an autoencoder? An Evaluation of Disentangled Representation Learning for Texts finds that disentanglement works only with parallel paraphrase supervision. Unsupervised or adversarial variants fail or suffer posterior collapse, a failure where the model learns to ignore its latent code. No parallel article-tweet corpus exists for the same content, so this route lacks the supervision it is known to need. The multi-task topic-suppression approach of Ordoñez et al. suppresses topic but not platform, so we include it only as a secondary comparison.
Proposed Experiments and Decisive Checks
The primary setting is CROSSNEWS Gold, with 500 authors, each contributing 100 articles and 100 tweets. We evaluate verification by accuracy and AUC in the article-article, article-tweet and tweet-tweet conditions, and attribution by average rank. We report results on standard pairs and on the topic-adversarial pairs we construct.
The first decisive check is a correlation test and needs no classifier. For each author we compute the Spearman correlation between $d_{a,\text{article}}$ and $d_{a,\text{tweet}}$ and compare it to a null distribution built from mismatched author pairs. If same-author correlations are indistinguishable from the null, the hypothesis is wrong. In that case the paper becomes a negative result showing that masked cross-platform authorship signal is lexical or platform-level rather than syntactic.
The second check is the ablation ladder described under Baselines, run on topic-adversarial pairs. It asks whether gains come from standardization, from platform-marker masking, or from the syntactic features themselves.
The third check covers generalization. We run the Reddit, Amazon and Fanfiction datasets from Learning Universal Authorship Representations, where we found no reported result on Lacuna for syntactic-profile verifiers. We also run the German, single-platform, cross-genre AVShift benchmark. AVShift tests whether standardization helps when genre shifts but platform does not, and it lets us report behaviour across the time buckets up to 21 years apart.
As a failure diagnostic, we plot verification accuracy against tweet-pool size, from 10 to 100 tweets per author, to find where short-text sparsity makes the syntactic channel useless.
References
- CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark
- When Writing Style Drifts: Benchmarking Authorship Verification under Distribution Shifts in Genre, Time and the AI-Era
- Learning Universal Authorship Representations
- Can Authorship Representation Learning Capture Stylistic Features?
- The Topic Confusion Task: A Novel Evaluation Scenario for Authorship Attribution
- Improving Gender Classification of Blog Authors
- Authorship Attribution through Function Word Adjacency Networks
- Syntax Encoding with Application in Authorship Attribution
- The Power of Character N-grams in Native Language Identification
- Enhancing Portuguese Variety Identification with Cross-Domain Approaches
- MMSD2.0: Towards a Reliable Multi-modal Sarcasm Detection System
- Covering Cracks in Content Moderation: Delexicalized Distant Supervision for Illicit Drug Jargon Detection
- An Evaluation of Disentangled Representation Learning for Texts
- Will Longformers PAN Out for Authorship Verification? Notebook for PAN at CLEF 2020