Multi-Author Style Analysis

Updated 7 July 2026

Multi-Author Writing Style Analysis is the study of identifying and segmenting distinct authorial voices using traditional stylometry, supervised classification, and neural boundary detection.
It leverages classic function-word analysis alongside hierarchical neural models to address challenges in disentangling topic and style signals.
The research enables applications in forensic authorship, digital humanities, and academic integrity through accurate attribution and stylistic boundary detection.

Multi-Author Writing Style Analysis is the stylometric study of how multiple authorial voices are detected, separated, compared, or profiled within and across documents. In contemporary work, it spans closed-set authorship attribution, document-level classification of single versus multi-authored texts, intrinsic detection of positions where writing style changes, author-presence recognition in collaborative documents, and unsupervised segmentation of a text into latent stylistic regions. The field therefore connects classic function-word stylometry, supervised text classification, representation learning, contrastive retrieval, and recent Transformer- and LLM-based boundary detectors, with granularity ranging from chapters and paragraphs to adjacent sentence boundaries (Iyer et al., 2019, Kucukkaya et al., 2023, Fifield et al., 2015, Bevendorff et al., 9 Feb 2026).

1. Problem space and task formulations

The central object of study is a document that may contain one or more authors, where “writing style” is treated as a proxy for authorship. In the now-standard PAN formulation, the task is intrinsic: systems inspect only the document itself and must identify all positions where the writing style changes. Earlier and parallel formulations are broader. Some treat the problem as closed-set authorship attribution over a fixed author set; others reduce it to binary same-author versus different-author decisions over adjacent segments; still others ask only whether a document is single-authored or multi-authored (Bevendorff et al., 9 Feb 2026, Iyer et al., 2019, Kucukkaya et al., 2023, Zamir et al., 2023).

At sentence granularity, the boundary-detection view can be written as a binary sequence over adjacent units. For a sequence of sentences $[s_1,\dots,s_n]$ , the output is $[c_1,\dots,c_{n-1}]$ , where $c_i=1$ if there is a style change between $s_i$ and $s_{i+1}$ , and $c_i=0$ otherwise (Römisch et al., 1 Aug 2025). Paragraph-level formulations use the same logic, but only at paragraph boundaries (Kucukkaya et al., 2023). In closed-set attribution, by contrast, a document $d_i$ is assigned a single author $a_j$ from a candidate set $A$ by maximizing $P(a_j\mid d_i)$ or, operationally, a linear multi-class decision score (Iyer et al., 2019).

Variant	Output	Representative formulation
Closed-set attribution	One author label from a fixed set	Single-label multi-class text categorization (Iyer et al., 2019)
Intrinsic change detection	$[c_1,\dots,c_{n-1}]$ 0 label at each boundary	Consecutive paragraph or sentence pairs (Kucukkaya et al., 2023, Römisch et al., 1 Aug 2025)
Single vs multi-authored classification	One binary document label	Document-level provenance/authentication setting (Zamir et al., 2023, Zamir et al., 2024)
Author presence recognition	Multi-label author vector per document	Collaborative documents with 5 possible authors (Albaqami et al., 3 Jan 2026)
Unsupervised segmentation	Soft author weights over segments	Multiple shifted clusterings of one text (Fifield et al., 2015)

This plurality of task definitions is not merely terminological. It reflects different assumptions about what is known in advance. Closed-set attribution presumes training data for known authors; intrinsic boundary detection presumes only local changes in a single document; unsupervised attribution presumes neither author identities nor author exemplars, only a hypothesized number of latent authorial styles (Fifield et al., 2015). A plausible implication is that multi-author writing style analysis is best understood as a family of related stylometric inference problems rather than a single benchmark task.

2. Stylometric foundations and representational choices

A recurring premise is that writing style is multi-level. One line of work characterizes style as “a combination of consistent decisions associated with a specific author at different levels of language production, including lexical, syntactic, and structural,” while another organizes it into lexical, syntactic, and semantic perspectives that are intended to be task-independent (Jafariakinabad et al., 2019, Verma et al., 2019). In classic legal stylometry, the most stable signals are function words, whose usage is treated as relatively content-independent; the Supreme Court study uses 63 function words and derives a normalized variability statistic $[c_1,\dots,c_{n-1}]$ 1 from their frequencies across opinions (Rosenthal et al., 2011).

Feature engineering in authorship identification makes this multi-level view explicit. A widely cited framework combines word unigrams, word bigrams, POS bigrams, word/POS pairs, and 23 stylometric features grouped into phraseology analysis, punctuation analysis, and lexical usage analysis. The phraseology features include lexical diversity, mean word length, mean sentence length, standard deviation of sentence length, mean paragraph length, and document length; punctuation features count commas, semicolons, quotations, exclamations, colons, hyphens, and double hyphens; lexical-usage features count items such as “and,” “but,” “however,” “if,” “that,” “more,” “must,” “might,” “this,” and “very” (Iyer et al., 2019). In that framework, lexical diversity is defined as $[c_1,\dots,c_{n-1}]$ 2, where $[c_1,\dots,c_{n-1}]$ 3 is total tokens and $[c_1,\dots,c_{n-1}]$ 4 is the number of unique stems (Iyer et al., 2019).

Another consistent finding is that preprocessing choices affect stylometric signal. In document-level multi-author classification and paragraph-level change detection on PAN-style data, raw text often outperforms more aggressively cleaned text because stop words, short words, punctuation, contractions, and other special characters encode author-specific habits. Both the document-provenance study and the later multi-task stylometry analysis report that removing such material can reduce performance, especially for tasks that hinge on subtle style variation rather than topic words (Zamir et al., 2023, Zamir et al., 2024).

Neural representation learning has extended rather than replaced these signals. The Style-HAN architecture uses parallel lexical and syntactic channels: GloVe word embeddings in one channel, POS-tag embeddings in the other, CNNs to capture local lexical or POS $[c_1,\dots,c_{n-1}]$ 5-grams, then a BiLSTM with attention to model document structure and emphasize stylistically salient sentences (Jafariakinabad et al., 2019). VADES moves further toward explicit style-aware representation learning by learning Gaussian embeddings for both authors and documents under a Variational Information Bottleneck objective, while aligning embedding dimensions with 300 predefined stylistic features including letters, numbers, structural features, punctuation, function words, POS tags, named entities, and readability indexes (Terreau et al., 2024). This suggests that the field increasingly treats style as a structured latent representation, not merely a bag of sparse counts.

3. Methodological paradigms

A first paradigm is supervised high-dimensional classification. In the Reuters-50-50 setting, authorship identification is cast as a 50-way single-label problem over 5,000 balanced news stories, and a LibLINEAR SVM is selected after comparison with logistic regression, SMO, Naïve Bayes, and J48. The final linear model uses 60,000 selected features after chi-squared ranking and delivers robust multi-author discrimination from sparse textual and stylometric inputs (Iyer et al., 2019). A closely related document-level paradigm applies merit-based late fusion: BERT, DistilBERT, RoBERTa, XLM-RoBERTa, and ALBERT are fine-tuned separately and then combined through weighted score fusion, where weights are optimized with PSO, Nelder–Mead, or Powell’s method (Zamir et al., 2024, Zamir et al., 2023).

A second paradigm is hierarchical or style-aware neural encoding. Style-HAN learns author-sensitive document representations by jointly modeling lexical, syntactic, and structural information, while VADES learns author and document embeddings in a shared low-dimensional style space with an explicit stylistic constraint (Jafariakinabad et al., 2019, Terreau et al., 2024). These models are naturally suited to many-author analysis because authors become vectors or distributions in representation space, enabling attribution, comparison, and clustering.

A third paradigm is intrinsic boundary detection via pairwise classification. The PAN 2023 ARC-NLP system reformulates paragraph-level style change detection as a natural language inference problem over consecutive paragraphs, using Transformer encoders and a binary classifier over the $[c_1,\dots,c_{n-1}]$ 6 representation. Its main architectural device is transition-focused truncation: for 512-token models, it keeps the last 256 tokens of the first paragraph and the first 256 tokens of the second paragraph, thereby emphasizing the boundary region itself (Kucukkaya et al., 2023). Sentence-level PAN 2024/2025 work uses an analogous local-boundary view, but with zero-shot prompting of LLMs: the model receives a JSON list of sentences and must emit a JSON array of 0/1 boundary labels (Römisch et al., 1 Aug 2025).

A fourth paradigm is retrieval and verification. Cross-genre attribution work replaces direct classification with document embedding and nearest-neighbor ranking, using a RoBERTa-based encoder trained with supervised contrastive loss and unusually careful data presentation. Same-author positives are selected to be topically dissimilar using SBERT cosine similarity, while hard negatives are induced by batch construction, yielding a model that is explicitly pressured to learn style rather than topic (Fincke et al., 2024). A training-free verification variant combines TF-IDF character $[c_1,\dots,c_{n-1}]$ 7- to $[c_1,\dots,c_{n-1}]$ 8-grams with all-MiniLM-L6-v2 embeddings, then classifies pairs by their position within empirical same-author and different-author distance distributions rather than by a trained decision boundary (Jemama et al., 29 Sep 2025).

A fifth paradigm is unsupervised segmentation. One method divides a single long text into overlapping fixed-length fragments, performs multiple independent clusterings over shifted fragment sets, aligns label identities across clusterings by maximum-weight bipartite matching or a MAX-CUT reduction, and then averages the aligned outputs to obtain soft author weights over segments (Fifield et al., 2015). A related but historically distinct approach seeks a “chrono-divide” in a chronologically ordered text: pseudo-aggregated SVM-RFE selects stable function-word and structural features, and a linear SVM score is tracked over chapters to identify a persistent stylistic break (Hu et al., 2014). These approaches are especially relevant when external author profiles do not exist.

4. Data regimes and evaluation protocols

Benchmark construction strongly shapes what counts as “style.” Reuters-50-50 minimizes topic variation by restricting all 50 authors to Reuters newswire, with 100 texts per author and balanced development, cross-validation, and holdout splits (Iyer et al., 2019). PAN 2023 makes the topic-style interaction explicit through three difficulty levels: easy, where paragraphs in a document can cover various numbers of topics; medium, where topical variety is small; and hard, where all paragraphs are on the same topic (Kucukkaya et al., 2023). PAN 2024 and 2025 move to Reddit discussion threads and sentence-level boundaries, with easy, medium, and hard splits and a typical problem containing 2–4 authors and 1–3 style changes (Römisch et al., 1 Aug 2025).

PAN 2026 continues the intrinsic change-detection task but shifts the source domain to fanfiction, where long, topically coherent excerpts can be mixed while controlling semantic and stylistic similarity. The planned dataset variants include high topical similarity with high stylistic similarity, high topical similarity with low stylistic similarity, and low topical similarity with potentially mixed stylistic similarity (Bevendorff et al., 9 Feb 2026). This indicates a deliberate movement toward more realistic long-form documents and more controlled disentanglement of topic and style.

Evaluation is correspondingly heterogeneous. Classical attribution work reports accuracy and Cohen’s kappa (Iyer et al., 2019). PAN boundary detection emphasizes Macro F1 over the change and no-change classes because class imbalance is common (Kucukkaya et al., 2023, Römisch et al., 1 Aug 2025). Cross-genre retrieval on the HIATUS Research Set uses Success@8, asking whether at least one true same-author document appears among the top eight retrieved documents from a haystack of documents by more than 21,000 authors (Fincke et al., 2024). Verification and imitation studies report accuracy, ROC AUC, F1, confidence derived from empirical distance distributions, and, when human versus machine detectability is also at issue, perplexity under GPT-2 (Jemama et al., 29 Sep 2025). Historical stylometry additionally uses corpus-level variability measures such as $[c_1,\dots,c_{n-1}]$ 9, based on a normalized chi-square statistic over function-word counts, and agreement measures between clustered fragmentations and gold authorship segmentations (Rosenthal et al., 2011, Fifield et al., 2015).

Hybrid human-machine settings introduce still another evaluation layer. In an academic-integrity framework, Gemini-generated paragraphs are injected as a fifth author into PAN-style multi-authored documents, and the system is evaluated on four tasks: human versus machine text classification, single versus multi-authored document classification, author change detection, and document-level author recognition with five possible contributors (Albaqami et al., 3 Jan 2026). A plausible implication is that multi-author writing style analysis is becoming a general framework for mixed-authorship forensics rather than a purely human-only stylometric task.

5. Empirical findings

One persistent result is that explicit stylometric information improves over purely lexical baselines. In the Reuters-50-50 framework, 10-fold cross-validation accuracy rises from 88.83% with unigrams alone to 91.29% when word bigrams, POS bigrams, word/POS pairs, and 23 stylometric features are added; on the holdout test set, accuracy improves from 79.8% to 81.6%, although that final improvement is not statistically significant (Iyer et al., 2019). Style-HAN similarly shows that combining lexical, syntactic, and structural information outperforms lexical-only or syntactic-only variants across CCAT10, CCAT50, BLOGS10, and BLOGS50, with especially large gains on the blog datasets (Jafariakinabad et al., 2019).

For document-level detection of multi-authorship, raw text is often more revealing than cleaned text. On PAN 2021 single-versus-multi-authored classification, optimized fusion on raw data reaches F1 around 0.848, exceeding prior submissions and outperforming corresponding cleaned-data settings; on the same benchmark, single boundary change detection reaches F1 around 0.782 on raw data, while the more difficult real-world multi-change task reaches F1 around 0.5535, with cleaned text slightly better there (Zamir et al., 2024). A closely related provenance study reports F1 around 0.852 for single-versus-multi-authored document classification, again with merit-based fusion and un-clean text outperforming stronger cleaning regimes (Zamir et al., 2023).

For intrinsic boundary detection, local transition modeling matters. In PAN 2023, DeBERTa v3 with warmup and transition-focused truncation achieves the best validation Macro F1 on easy and medium, at 0.987 and 0.812, respectively, while DeBERTa v3 with warmup but without transition-focused truncation is best on hard, at 0.770. Warmup is especially important on medium and hard settings, where several backbones otherwise collapse toward near-chance scores (Kucukkaya et al., 2023). This suggests that the utility of local boundary context versus broader paragraph coverage depends on how much topic variation remains in the benchmark.

LLMs are already competitive in zero-shot settings. On PAN 2024/2025 multi-author writing style analysis, Claude 3.7 Sonnet outperforms GPT-4o, DeepSeek-R1, and Meta-Llama-3.1-405B-Instruct in a zero-shot prompt-based boundary detector. On the full PAN 2025 validation splits, Claude reaches Macro F1 0.8559 on easy, 0.8182 on medium, and 0.6612 on hard, nearly matching a supervised StyleDistance + LSTM + MLP system on medium difficulty (Römisch et al., 1 Aug 2025). The same study reports weak correlations between semantic similarity and predicted boundary changes in harder settings, which the authors interpret as evidence that recent LLMs are sensitive to content-independent stylistic cues (Römisch et al., 1 Aug 2025).

Style/content separation also matters in cross-genre attribution. By selecting same-author training pairs that are maximally dissimilar in SBERT space and by constructing batches with hard negatives, SADIRI yields a 62.7% relative improvement in average cross-genre authorship attribution and 16.6% in the per-genre condition (Fincke et al., 2024). This is a direct empirical argument that training data selection and presentation can matter as much as, or more than, architectural novelty.

Verification and imitation work further complicates the picture. A training-free same-author verifier built from TF-IDF character $c_i=1$ 0- to $c_i=1$ 1-grams plus MiniLM embeddings achieves 97.49% accuracy in same-domain and 94.48% in cross-domain settings, while few-shot or completion prompting can make LLM outputs nearly indistinguishable from a target human author by this verifier, with completion reaching 99.9% agreement. Yet these same outputs remain more statistically predictable than human essays, with human essays averaging perplexity 29.5 and matched LLM outputs averaging 15.2, implying that stylistic fidelity and statistical detectability are separable (Jemama et al., 29 Sep 2025). A related adversarial study reports up to 75% multi-author imitation success among five authors using combinatorial paraphrasing guided by classifier probabilities (Gröndahl et al., 2019).

Finally, historical stylometry shows that multi-author analysis predates Transformers by decades. Function-word variability in Supreme Court opinions distinguishes justices and suggests differing degrees of clerk involvement; in that corpus, the normalized variability statistic $c_i=1$ 2 is highest for Kennedy and lowest among the main justices for Scalia (Rosenthal et al., 2011). In Chinese literary stylometry, a chrono-divide method applied to Dream of the Red Chamber produces a sharp boundary between chapters 1–80 and 81–120 and flags Chapter 67 as stylistically anomalous, while the same method finds no comparable divide in the other three Great Classical Novels (Hu et al., 2014). These findings situate current neural work within a longer tradition of detecting heterogeneity of authorial voice through stable stylistic regularities.

6. Applications, limitations, and future directions

The application space is broad. Multi-author writing style analysis is used for document provenance and authentication, single-versus-multi-authored detection, collaborative document review, forensic linguistics, plagiarism and ghostwriting detection, digital-humanities analysis of disputed literary texts, and academic-integrity workflows that explicitly treat machine text as an additional author (Zamir et al., 2023, Zamir et al., 2024, Albaqami et al., 3 Jan 2026). In PAN-style settings it is also a foundation for source-poor plagiarism detection, where the goal is not retrieval of a source document but localization of intrinsic stylistic inconsistencies (Bevendorff et al., 9 Feb 2026).

The main limitations are equally consistent across the literature. Topic-style confounding remains fundamental, especially outside controlled benchmarks (Fincke et al., 2024). Many systems assume paragraph boundaries or fixed-size chunks, and pairwise boundary detectors often lack global consistency constraints over an entire document (Kucukkaya et al., 2023). Closed-set models assume known authors and sufficient text length, while short texts destabilize lexical diversity, punctuation counts, or sentence-structure estimates (Iyer et al., 2019, Terreau et al., 2024). LLM-based approaches can depend on dataset-specific priors such as “always assume around 3 changes,” can over-segment, and may benefit from pretraining overlap with Reddit-like domains (Römisch et al., 1 Aug 2025). High-performing style-embedding and attribution systems also raise privacy risks because they support deanonymization, imitation, and identity-conditioned generation (Gröndahl et al., 2019, Jemama et al., 29 Sep 2025).

Current research directions point toward longer-context models, richer author and document embeddings, stronger topic-style disentanglement, and more realistic mixed-authorship benchmarks. PAN 2026 explicitly emphasizes fanfiction-based long-form data and controlled topical/stylistic similarity (Bevendorff et al., 9 Feb 2026). VADES indicates one route toward interpretable author-space modeling with uncertainty estimates (Terreau et al., 2024). Cross-genre contrastive training shows that data curation can force models to rely more on style than substance (Fincke et al., 2024). Zero-shot LLM boundary detection suggests that foundation models can already act as strong baselines, but the persistent gap between style matching and robust explanation implies that future systems will need better grounding, better calibration, and better evaluation of what they are actually using as evidence (Römisch et al., 1 Aug 2025).

Taken together, the literature portrays Multi-Author Writing Style Analysis as a mature but still unsettled branch of computational stylometry. Its core insight is stable: authorship leaves measurable traces in lexical choice, function words, punctuation, syntactic structure, and higher-level discourse patterns. Its unsettled question is how to isolate those traces when topic, genre, editorial intervention, collaborative writing, and LLM-assisted generation all introduce alternative explanations for stylistic variation.