Authorship attribution means deciding who wrote a text by its style rather than its content. It works well on novels and poorly on tweets. The problem gets worse when the system is deployed, because new messages keep arriving and their topics keep shifting. An investigator may hold only a handful of reference articles per suspect and then face a stream of unlabeled short posts whose subject matter changes week to week. A model trained once will gradually confuse "talks about the election" with "writes like author X." A model that keeps updating itself on its own guesses tends to reinforce those confusions. This proposal asks one question: can an authorship model adapt to a drifting stream of short texts without labels, and without forgetting, if every update is tied to features we already know carry style and not topic?
The core idea is simple. While the stream runs, the model trains itself on a single auxiliary task: reconstructing masked pivot features, meaning character patterns built around punctuation, function words, and author-unique "signatures," which earlier stylometry work showed to be largely topic-independent. A second component is an adversary, conditioned on the predicted author, that removes whatever information identifies the current time window. The bet is that a self-supervised signal anchored to style lets the encoder track drift safely. A generic signal, such as reconstructing arbitrary masked characters, would let topic leak back into the representation. The method is a small trainable head that can sit on a compact character encoder or on frozen large-language-model embeddings, so it competes directly with the strongest static systems rather than replacing them.
Core Claim
Label-free continual adaptation for authorship is safe only if two things hold. The adaptation target must be topic-independent, and the alignment pressure must be conditioned on author identity. We claim that pivot-anchored test-time training combined with author-conditioned adversarial alignment will beat both a frozen encoder with growing author prototypes and generic self-training. The gain should appear mainly on topics that recur late in a long stream, which is where existing methods either drift or forget. If generic masked-character reconstruction does as well as pivot reconstruction, the claim is false.
What Exists and What Is Missing
The supplied literature covers every pair of these directions but never the combination that matters for deployment. On the stylometry side, short-text attribution is solved best by local character patterns. Convolutional Neural Networks for Authorship Attribution of Short Texts runs multi-width character convolutions with max-pooling over tweets and stays stable when training data shrinks from 500 to 50 tweets per author. Authorship Attribution of Micro-Messages shows that 81% of users have at least one k-signature, a pattern present in their tweets and absent from everyone else's. Both assume a fixed, closed set of authors and no topic drift. We reuse both as the source of our pivots.
Three works address topic and genre shift, and all three are static. Domain Adaptation for Authorship Attribution: Improved Structural Correspondence Learning selects punctuation-typed character n-grams as topic-independent pivots and learns linear correspondences from unlabeled target text, but it does so offline on Guardian articles. Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help? shows that pooling labeled data from many topics pushes a support vector machine toward invariant character 3-grams, but this needs labeled multi-topic data. CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark introduces SELMA, which extracts frozen embeddings from an instruction-tuned large language model and is the strongest cross-genre system in the file, but it never updates. Taken together, these papers motivate the proposal: pivots work, diversity helps, and strong frozen encoders exist. What none of them does is adapt as the stream moves.
The continual-adaptation work supplies our methods and also warns us about failure modes. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts updates a shared encoder on an auxiliary self-supervised task at test time, but its task (image rotation) has no text analogue. Pivot reconstruction is our text analogue. Continual Test-Time Domain Adaptation prevents runaway drift by randomly resetting 1% of weights to their source values at each step, and it documents the "death spiral" of self-reinforcing pseudo-labels. Its augmentation-averaged teacher depends on image flips and crops, so we keep the restoration step and drop the augmentations. Incremental Adversarial Domain Adaptation breaks a large shift into a sequence of small adversarially aligned steps, but its authors state that it does not prevent forgetting of earlier domains. That warning constrains our evaluation, which explicitly revisits old topics.
Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring shows that a topic adversary on a small parameter budget (0.39% of BERT-base) can make essay representations topic-invariant. Unconditional invariance has a known cost in stylometry, however: it can erase stylistic choices that correlate with topic. We expect this cost to be large here, because in a stream the active set of authors also changes from window to window. Continual Learning for Text Classification with Information Disentanglement Based Regularization splits text representations into a protected generic space and a plastic task-specific space. We borrow the split, but our protected space is defined by pivot reconstruction rather than next-sentence prediction, which carries topic.
Research Program
The system has three parts. The first is an encoder, either a character convolutional network or a frozen SELMA embedding. The second is a small projection head producing a style vector $z$. The third is an author scorer that compares $z$ against stored author prototypes, so new authors can be added without retraining a classifier. Before deployment, the head is trained on the labeled reference data with three losses: a contrastive author loss, a pivot-reconstruction loss, and an adversarial loss.
The pivot vocabulary is fixed in advance from the reference set. It contains punctuation-typed character n-grams in the style of the structural correspondence learning paper, the highest-frequency function words, and each author's k-signatures. The reconstruction head masks the pivot occurrences in a message and predicts them from $z$.
During deployment, the stream is cut into windows. For each window the head takes a few gradient steps on the following objective:
$$ \mathcal{L} = \mathcal{L}_{\text{pivot}}(z) - \lambda\, \mathcal{L}_{\text{win}}\big(D(z, \hat{a}),\, t\big), $$
Here $\mathcal{L}_{\text{pivot}}$ is the masked-pivot reconstruction loss, $D$ is a discriminator that receives the style vector together with the current predicted author $\hat{a}$ and tries to name the window index $t$, and $\lambda$ weights the adversarial term. The minus sign is implemented with a gradient reversal layer, a layer that flips gradient signs so the head learns to fool the discriminator. Conditioning on $\hat{a}$ matters. The discriminator then can only exploit information that separates windows within one author, which is mostly topic. Without conditioning, it could also exploit which authors happen to be active in a window, and removing that would remove identity itself. Only messages whose predicted-author margin passes a threshold feed into $\hat{a}$. CoTTA-style stochastic restoration resets a small random fraction of head weights after each window.
The sequential-extraction direction enters as a controlled question about the encoder rather than a design commitment. The literature disagrees. Recurrent Convolutional Neural Networks for Text Classification reports that its bidirectional recurrent context layer cut errors by a third on long documents, yet it scored 47.21% against 48.70% for paragraph vectors on short sentences. Meanwhile, pure character convolutional networks lead on tweets. We therefore compare three encoders under identical adaptation: a character convolutional network, the same network followed by a bidirectional recurrent layer, and a character-level version of that recurrent-convolutional model. We predict the recurrent layer helps only on the article-length reference side.
Baselines
The naive reference point is the source model frozen at deployment, using the same encoder and head, with author prototypes fixed from the references. Every gain is measured against it.
The strongest existing system is SELMA from CROSSNEWS. That paper reports about 78.5% Article-Tweet verification accuracy and an average rank of about 5.6 for Tweet-Article attribution. No reported result exists on Lacuna for any method under a drifting, streamed protocol, so we will re-run SELMA ourselves on our stream splits. We will give it its best fair form: frozen embeddings plus prototypes that are updated with confident stream messages, with the confidence threshold and update rate tuned on the same validation stream we use for our method. Beating it means higher stream-averaged attribution accuracy and higher accuracy on revisited topics, both with confidence intervals from multiple stream orderings, when our head sits on the same frozen SELMA embeddings. If our head only wins on top of the character encoder, we will report that as a cost-efficiency result and not as an accuracy result.
The strongest short-text system trained from scratch is the character bigram convolutional network from Convolutional Neural Networks for Authorship Attribution of Short Texts, which reports 76.1% accuracy on 50 authors with 1,000 tweets each, against 71.2% for the earlier feature-engineered system. We will reproduce it on that benchmark before using it, so we know our implementation is sound. The classical reference is a character 3-gram support vector machine trained in the pooled multi-topic setting of Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help?, which reports 57.46% against 39.21% intra-topic on chat. The continual reference is CoTTA-style pseudo-label self-training with restoration but without the pivot task or the adversary.
Three ablations isolate the key ingredient. The first replaces pivot reconstruction with masked reconstruction of arbitrary characters. The second makes the discriminator unconditional, dropping the predicted author from its input. The third removes restoration.
All methods share the same reference data, the same stream, the same per-window gradient-step budget where they adapt, and the same number of tuning trials on a held-out validation stream.
Why not just…?
Why not just use frozen SELMA embeddings with nearest-prototype matching and never adapt? This is the most serious threat. SELMA already learns embeddings that separate authors from topics, and frozen models cannot spiral. We have changed the design in response: the method is a small head on top of whatever encoder is best, and the head-to-head comparison is run on SELMA's own embeddings. If prototype updating alone matches us on revisited topics, the conclusion is that strong frozen encoders make label-free adaptation unnecessary for this task. That would itself be a useful negative result, and we will report it as such.
Why not just pool many topics into one static training set? The pooled-topic result in Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help? depends on labeled data across topics, which forensic settings with few references lack. The vision evidence also shows static models losing ground under continued drift: 76.27% against 80.21% mean average precision on night driving in Incremental Adversarial Domain Adaptation, and 46.2% against 32.5% error on CIFAR-100-C in Continual Test-Time Domain Adaptation. Whether text drift behaves the same way is exactly what the frozen-source and pooled baselines will reveal.
Why not just run pseudo-label self-training with weight restoration, as CoTTA does? Self-training on authorship guesses has no topic-independent anchor. When a topic dominates a window, wrong guesses that share that topic reinforce one another. This is the death spiral that Continual Test-Time Domain Adaptation controls in vision only by averaging over augmentations, and those augmentations do not exist for text. The CoTTA-style baseline and the generic-reconstruction ablation test this directly. If either matches the pivot version, the anchoring claim fails.
Why not just prompt a large language model? In CROSSNEWS, zero-shot prompting methods trailed embedding methods. In Open-World Authorship Attribution, end-to-end attribution reached 18.7%. Prompting is included only as a sanity row.
Proposed Experiments
The main benchmark is a stream built from the CROSSNEWS gold set: 500 authors, with 100 tweets and 100 articles each, plus fine-grained topic labels. We simulate scarcity by giving each author only five reference articles. The stream consists of that author's tweets, ordered into topic blocks, and the full topic sequence repeats ten times, mirroring CoTTA's repeated-condition protocol. This exposes both slow drift and forgetting. Because the file does not tell us whether tweet timestamps are available, the stream order is topic-based by construction.
Two secondary streams test generality. The 50-author Twitter benchmark is reduced to 50 reference tweets per author, with bot-like accounts filtered as in the convolutional-network paper. The 21-author multi-genre forensic corpus is chained across topics and genres.
The metrics are stream-averaged attribution accuracy, accuracy on each topic block's last visit against its first visit (forgetting), verification accuracy on cross-genre pairs, and added latency per message.
Three checks are decisive. First, pivot reconstruction must beat generic masked-character reconstruction on late-cycle accuracy. Second, we probe topic information in $z$ by training a linear classifier to predict the topic label. This probe should stay flat over the ten cycles under our method and rise under the generic variant. Third, the author-conditioned adversary must beat the unconditional one, with the gap growing as the composition of active authors varies more between windows, which we control by resampling the stream. The encoder comparison runs alongside and settles, for this regime, whether recurrent layers earn their cost on micro-texts.
References
- Convolutional Neural Networks for Authorship Attribution of Short Texts
- Authorship Attribution of Micro-Messages
- Domain Adaptation for Authorship Attribution: Improved Structural Correspondence Learning
- Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help?
- CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark
- Test-Time Training with Self-Supervision for Generalization under Distribution Shifts
- Continual Test-Time Domain Adaptation
- Incremental Adversarial Domain Adaptation for Continually Changing Environments
- Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring
- Continual Learning for Text Classification with Information Disentanglement Based Regularization
- Recurrent Convolutional Neural Networks for Text Classification
- Open-World Authorship Attribution