Multilingual Authorship Attribution

• Multilingual authorship attribution is a technique that determines a text’s true author across various languages and scripts by analyzing stylistic and linguistic patterns.
• It leverages diverse methodologies including character n-grams, structural syntax, and multilingual encoder models to address challenges like script mismatch and cross-lingual transfer.
• Empirical studies report high macro-F1 scores and robust performance using contrastive learning and multi-task frameworks across multiple languages and author classes.

Multilingual authorship attribution is the task of identifying the precise author of a text across multiple languages and writing scripts. In current research, the term covers several closely related problems: closed-set authorship attribution among human and machine classes, cross-lingual transfer from source languages to unseen target languages, pairwise authorship verification, untranslated cross-language attribution for bilingual authors, and source attribution for LLM-generated text. It therefore extends beyond monolingual stylometry in two directions at once: the author space may include humans and multiple generators, and the linguistic space may span typologically diverse languages, scripts, and domains (Cava et al., 3 Aug 2025, Habib et al., 21 May 2025, Rao et al., 19 Aug 2025).

1. Task definitions and formal problem settings

Authorship attribution (AA) is a multi-class identification problem in which a model assigns a text to one author from a candidate set. Authorship verification (AV) is a binary decision problem that asks whether two texts were written by the same author. The literature further distinguishes multilingual AA/AV, where methods are trained and evaluated across multiple languages, from cross-lingual transfer, where a model learned on one language or language subset is applied to another. Open-world or open-set settings are recognized as more realistic, but several recent multilingual studies explicitly remain closed-set (Habib et al., 21 May 2025, Cava et al., 3 Aug 2025).

A recent formalization for machine-generated multilingual AA defines a text set $X = X_h \cup X_m$, where $X_h$ is human-written and $X_m$ machine-generated; the author label is $y \in \{y_h\} \cup V_m$, with $V_m$ containing one label per generator. In that setting, a multi-class classifier $f: X \to \{y_h\} \cup V_m$ is trained over $K = 8$ classes, and multilingual suitability is evaluated by training on all 18 languages jointly, whereas cross-lingual AA trains on $L_{\text{train}} \subset L$ and tests on disjoint $L_{\text{test}} \subset L$ (Cava et al., 3 Aug 2025).

For multilingual LLM forensics, detection and attribution are separated but related tasks. Detection asks whether a text $x$ is human-written or machine-generated; attribution identifies which specific LLM produced it. One joint formulation uses a dataset

$X_h$0

with binary detection labels $X_h$1 and attribution labels $X_h$2, optimized under a black-box assumption in which only model outputs are available (Rao et al., 19 Aug 2025).

Pairwise multilingual AV adopts a different abstraction. In Stylomech, the input is an anonymous text $X_h$3 and a suspect author’s corpus $X_h$4, and the task is to decide whether they were written by the same person. Rather than predicting from a fixed roster of authors, the model learns same-author versus different-author discrimination from pairwise stylometric similarity, so it can be applied to unseen authors without author-specific retraining (Faumi et al., 16 Jan 2025).

2. Stylometric, lexical, and structural representations

The multilingual AA literature retains a strong stylometric component. Recurrent feature families include character n-grams, function word frequencies, average word length, average sentence length, POS n-grams, punctuation patterns, TF–IDF, topic distributions, Burrows’ Delta, and contextual embeddings. The review literature emphasizes that character- and byte-level representations are repeatedly robust to scripts, morphology, orthographic variation, and out-of-vocabulary effects, particularly in low-resource and microtext settings (Habib et al., 21 May 2025).

Stylomech exemplifies a feature-rich pairwise design for English and Romanized Sinhala. It extracts language-agnostic signals such as character and word n-grams, punctuation and whitespace patterns, sentence and word length distributions, and lexical richness measures, while adding Romanized Sinhala-specific features such as English-to-Sinhala token ratio via Swa-Bhasha reverse transliteration and Levenshtein distances between similar romanized word forms. Given stylometric vectors $X_h$5 and $X_h$6, the pair representation is constructed as

$X_h$7

optionally with standardized features $X_h$8 (Faumi et al., 16 Jan 2025).

A different path to multilingual comparability is structural stylometry. DT-grams represent local dependency-tree neighborhoods using Universal Dependencies and universal POS tags, so features are directly aligned across languages without translation. The method defines ancestor chains, sibling windows, and combined $X_h$9 or inverted $X_m$0 patterns over a UD tree, then represents each document as a TF–IDF-weighted bag of such grams. POS-only node labels yield the strongest results overall on untranslated cross-language AA, and $X_m$1 and $X_m$2 outperform simpler ancestor-only or sibling-only variants (Murauer et al., 2021).

Literary bilingualism has also been studied through targeted lexical and morphosyntactic probes. In French texts by bilingual Russian–French authors, frequent-word classifiers and chi-squared word-profile distances were used to separate bilingual and monolingual groups. Adjectives in $X_m$3, the noun gouttelette(s), and adverbs in $X_m$4 were examined, but the most consistent interference signal was a relative predominance of demonstratives over definite articles, especially for cette and ces (Makarova, 2023).

3. Multilingual encoder-based, contrastive, and multi-task models

Recent multilingual AA systems center on pretrained multilingual encoders and decoder-based classifiers. For multilingual machine-generated text, compared supervised baselines include RoBERTa-large, XLM-RoBERTa-large, OTBDetector with XLM-RoBERTa-large replacing Longformer, and mdok, which fine-tunes Qwen3-4B-Base via QLoRA with an 8-way classification head. These models are evaluated on 18 languages and 8 author classes, with multilingual encoders and decoders explicitly intended to capture language-agnostic signals (Cava et al., 3 Aug 2025).

A complementary research line learns authorship representations rather than direct class labels. In multilingual AR learning, an encoder $X_m$5 maps a text $X_m$6 to an embedding $X_m$7 with $X_m$8, so similarity is cosine:

$X_m$9

Training uses supervised contrastive learning with two documents per author and an InfoNCE-style loss. This framework introduces two language-agnostic techniques: Probabilistic Content Masking (PCM), which masks content tokens more often than function tokens, and Language-Aware Batching (LAB), which forms each contrastive batch from one language at a time to reduce cross-lingual interference (Kim et al., 20 Sep 2025).

Multilingual LLM-source attribution has also been cast as multi-task learning. DA-MTL uses a shared encoder plus two task-specific heads, one for detection and one for attribution. If $y \in \{y_h\} \cup V_m$0 is the shared representation, then the detection head is

$y \in \{y_h\} \cup V_m$1

and the attribution head is

$y \in \{y_h\} \cup V_m$2

The optimization combines binary cross-entropy for detection and categorical cross-entropy for attribution with a weighted update

$y \in \{y_h\} \cup V_m$3

with default $y \in \{y_h\} \cup V_m$4 (Rao et al., 19 Aug 2025).

The reported motivation for joint training is asymmetrical task difficulty. Detection is described as coarser and easier to learn, with rapidly decreasing loss, whereas attribution is subtler and more variable. This suggests that shared representations regularized by the easier task can stabilize the harder one, particularly when human-written text is retained in the unified label space (Rao et al., 19 Aug 2025).

4. Corpora, languages, scripts, and evaluation protocols

Recent multilingual AA studies have broadened both language coverage and task diversity. MULTITUDE v3 is a central benchmark for machine-generated multilingual AA: it combines human-written news from MassiveSumm with machine-generated counterparts produced from headlines under consistent generation configurations, covers 18 languages, 8 author classes, and maintains near-uniform class distributions with aggregate counts of 143,114 train and 42,943 test samples (Cava et al., 3 Aug 2025).

The language coverage in MULTITUDE v3 spans multiple families and scripts: Germanic, Romance, Slavic-Latin, Slavic-Cyrillic, Hellenic, Uralic, Semitic, and Sino-Tibetan. This diversity is important because script and family boundaries strongly affect transfer, and statistical zero-shot detectors degrade sharply on Cyrillic, Arabic, and Hanzi (Cava et al., 3 Aug 2025).

LLM-source attribution benchmarks use different corpora. DA-MTL is evaluated on nine datasets, including MGTBench, MULTITuDE, and M4. MGTBench contains three English domains—Essay, WritingPrompt, and Reuters—with 1,000 samples per generator from six LLMs. MULTITuDE contains 74,081 samples across 11 languages and focuses cross-lingual analyses on EN, ES, RU, UK, CS, DE, NL, CA, and PT. M4 spans Wikipedia, Reddit, WikiHow, PeerRead, and arXiv abstracts with multiple generators and 2,344–3,000 samples per generator and domain (Rao et al., 19 Aug 2025).

Human-authorship studies remain heterogeneous. Stylomech uses English blogs and tweets from 100 authors and Romanized Sinhala WhatsApp chats from 50 volunteers. The untranslated DT-grams study uses bilingual Reddit comments for EN+DE, EN+ES, EN+PT, EN+NL, and EN+FR, with a minimum document length threshold of 3,000 characters. The Russian–French interference study uses 19 literary works split into 1700-word segments and evaluated with leave-one-book-out cross-validation (Faumi et al., 16 Jan 2025, Murauer et al., 2021, Makarova, 2023).

Evaluation protocols differ by task family. Multilingual AA commonly reports macro-F1 across classes per language; cross-lingual AA reports per-language macro-F1 under specific source-language settings; authorship representation learning reports Recall@8 and Mean Reciprocal Rank; pairwise AV reports accuracy, precision, recall, and F1; and untranslated cross-language AA reports macro-averaged F1 across authors. Several studies explicitly do not report statistical tests or confidence intervals, although the multilingual AR paper shows 95% confidence intervals in figures (Cava et al., 3 Aug 2025, Kim et al., 20 Sep 2025, Rao et al., 19 Aug 2025).

5. Empirical performance and cross-lingual generalization

Across recent work, multilingual supervised models substantially outperform repurposed statistical detectors. In ML-MGT on 18 languages and 8 author classes, mdok and OTBDetector achieve macro-F1 consistently $y \in \{y_h\} \cup V_m$5 across most languages; mdok has an “all” average of approximately $y \in \{y_h\} \cup V_m$6, OTBDetector approximately $y \in \{y_h\} \cup V_m$7, XLM-R-large approximately $y \in \{y_h\} \cup V_m$8, RoBERTa-large approximately $y \in \{y_h\} \cup V_m$9, StatEnsemble approximately $V_m$0, Fast-DetectGPT approximately $V_m$1, and Binoculars approximately $V_m$2. Random baseline is $V_m$3 because the classes are balanced (Cava et al., 3 Aug 2025).

Cross-lingual transfer remains substantially harder than multilingual in-language training. In the same study, mdok’s all-language macro-F1 is $V_m$4 when trained on English only, $V_m$5 on Spanish only, $V_m$6 on Russian only, and $V_m$7 on a balanced en-es-ru mix. Russian-only training yields the top performance in 6 of 8 language families, and the paper attributes this to Russian’s rich morphology and inflection, which encourages models to learn deeper linguistic signals (Cava et al., 3 Aug 2025).

Multilingual AR learning reports a related but not identical pattern. A single multilingual XLM-R-based model outperforms monolingual baselines in 21 out of 22 non-English languages, with an average Recall@8 improvement of $V_m$8 percentage points and a maximum gain of $V_m$9 points in Georgian. Relative to an English-only AR model, the multilingual model improves Recall@8 by $f: X \to \{y_h\} \cup V_m$0 points on unseen languages and by $f: X \to \{y_h\} \cup V_m$1 points on unseen English domains, while suffering only about a $f: X \to \{y_h\} \cup V_m$2 point drop on seen English domains (Kim et al., 20 Sep 2025).

Joint detection-attribution learning yields smaller but consistent gains. DA-MTL improves F1 over corresponding single-task baselines for all four backbones and all nine evaluated datasets. Average gains across all nine datasets are $f: X \to \{y_h\} \cup V_m$3 for DistilBERT, $f: X \to \{y_h\} \cup V_m$4 for mBERT, $f: X \to \{y_h\} \cup V_m$5 for XLM-R, and $f: X \to \{y_h\} \cup V_m$6 for RoBERTa. On MULTITuDE, the largest average gain is $f: X \to \{y_h\} \cup V_m$7. In nine-language multilingual evaluation, DA-MTL improves detection and attribution for both mBERT and XLM-R; average gains are $f: X \to \{y_h\} \cup V_m$8 and $f: X \to \{y_h\} \cup V_m$9 for mBERT, and $K = 8$0 and $K = 8$1 for XLM-R (Rao et al., 19 Aug 2025).

Classical multilingual stylometry remains viable, but usually at lower absolute performance. On untranslated bilingual data, DT-grams outperform the next best non-translation baseline by $K = 8$2 macro-F1 on average across five language pairs; best untranslated SVM scores are reported around $K = 8$3 for EN/DE, $K = 8$4 for EN/ES, $K = 8$5 for EN/FR, $K = 8$6 for EN/NL, and $K = 8$7 for EN/PT. In contrast, the machine-translated EN+DeepL condition allows character n-grams with SVM to reach a maximum around $K = 8$8 macro-F1 (Murauer et al., 2021).

Pairwise AV results are task-specific but informative. Stylomech reports English accuracy of approximately $K = 8$9, with precision, recall, and F1 balanced around $L_{\text{train}} \subset L$0, and Romanized Sinhala accuracy of approximately $L_{\text{train}} \subset L$1, with macro averages of precision approximately $L_{\text{train}} \subset L$2, recall approximately $L_{\text{train}} \subset L$3, and F1 approximately $L_{\text{train}} \subset L$4 (Faumi et al., 16 Jan 2025).

Study	Setting	Main result
(Cava et al., 3 Aug 2025)	ML-MGT, 18 languages, 8 classes	mdok “all” $L_{\text{train}} \subset L$5 macro-F1; OTBDetector $L_{\text{train}} \subset L$6
(Kim et al., 20 Sep 2025)	Multilingual AR, 22 non-English languages	Average Recall@8 gain $L_{\text{train}} \subset L$7 points; maximum $L_{\text{train}} \subset L$8 in Georgian
(Rao et al., 19 Aug 2025)	DA-MTL, 9 datasets, 4 backbones	Average F1 gains: DistilBERT $L_{\text{train}} \subset L$9, mBERT $L_{\text{test}} \subset L$0, XLM-R $L_{\text{test}} \subset L$1, RoBERTa $L_{\text{test}} \subset L$2
(Murauer et al., 2021)	Untranslated bilingual AA	DT-grams outperform next best non-translation baseline by $L_{\text{test}} \subset L$3 macro-F1
(Faumi et al., 16 Jan 2025)	Pairwise AV	Accuracy $L_{\text{test}} \subset L$4 for English; $L_{\text{test}} \subset L$5 for Romanized Sinhala

6. Error structures, robustness, ethics, and open problems

Error analysis shows that multilingual AA is often limited less by raw classification capacity than by author similarity, script mismatch, and domain shift. In multilingual machine-generated AA, English is described as paradoxically hard even for RoBERTa-large and multilingual models, likely because many LLMs are strongest in English and therefore converge stylistically. In cross-lingual settings, internal evaluations remain strong while external unseen-language performance degrades, with persistent confusion among architecturally similar models such as Llama-2-70B and Vicuna-13B (Cava et al., 3 Aug 2025).

DA-MTL confusion matrices on MULTITuDE reveal lineage-driven clusters. A “Davinci-like” cluster—Alpaca-LoRA-30B, Vicuna-13B, GPT-3.5, and Text-DaVinci-003—often confuses internally, while an “OPT-like” family—OPT-66B, OPT-IML-MAX-1.3B, and LLaMA-65B—also shows mutual confusion. NELA-feature analysis supports these proximities: OPT and LLaMA outputs are longer and more complex, with more proper nouns, past tense verbs, and explicit dates, while “Davinci-like” models are shorter, more readable, and more sentiment-heavy. However, classical ML trained directly on NELA features remains weak, with F1 around $L_{\text{test}} \subset L$6 for SVM and around $L_{\text{test}} \subset L$7 for logistic regression, indicating that deeper encoders exploit subtler distributional cues (Rao et al., 19 Aug 2025).

Robustness to obfuscation is only partially understood. On 500 English MULTITuDE samples, Mutant-X reduces attribution accuracy most on GPT-4 by $L_{\text{test}} \subset L$8 and least on OPT-66B by $L_{\text{test}} \subset L$9, while back-translation EN$x$0FR$x$1EN affects LLaMA-65B by $x$2 and GPT-3.5/DaVinci by approximately $x$3, but only Vicuna-13B by $x$4. DA-MTL is therefore reported as generally robust to mild obfuscation, but not designed as an adversarial defense (Rao et al., 19 Aug 2025).

Human-authorship work identifies analogous vulnerabilities. Stylomech reports degradation for very short texts, heavy code-mixing, extreme romanization variance, and intentional style masking. The broader review identifies topic shift, discourse shift, open-set uncertainty, paraphrasing, and calibration as central unresolved problems, and notes that tokenization-specific features degrade across scripts, whereas character-, byte-, and POS-based representations transfer better (Faumi et al., 16 Jan 2025, Habib et al., 21 May 2025).

Bilingual interference offers a distinct source of signal. In French literary prose, bilingual Russian–French authors are separable from both French classics and contemporary French writers with best leave-one-book-out cross-validation accuracy of $x$5 using a neural network. The strongest positive evidence of interference is not lexical exoticism but determiner usage: in the bilingual group, cette and ces are relatively more frequent and definite articles le/la/les relatively less frequent, with ces acquiring weights that tilt decisions toward the bilingual class in both Ridge and SVM (Makarova, 2023).

Ethical assessments in this literature are consistently dual-use. Multilingual attribution is presented as supporting accountability, intellectual property protection, content verification, and forensic analysis, but it also raises privacy, false-positive, surveillance, and cross-language fairness concerns. Recent work specifically flags uneven privacy impact across languages, demographic or regional style bias, the risk of misattribution in sensitive settings, and the need for calibrated confidence and explicit deployment restrictions. Open problems include stronger cross-lingual attribution without target-language fine-tuning, resilience to adversarial obfuscation, principled handling of rapidly evolving LLM families whose stylistic footprints converge over time, better treatment of code-switching and temporal drift, and broader multilingual benchmarks beyond English-centric or news-centric domains (Kim et al., 20 Sep 2025, Rao et al., 19 Aug 2025, Habib et al., 21 May 2025).

Multilingual authorship attribution has therefore evolved into a technically heterogeneous field unified by one principle: authorial style must be modeled in forms that survive changes in language, script, domain, and generation process. Current evidence suggests that multilingual encoders, contrastive objectives, and multi-task coupling materially improve performance, but classical stylometry and structural syntax remain relevant where interpretability, low-resource operation, or translation-free cross-language comparison are required.