Stylometric Detection Architectures

• Stylometric Detection Architectures are algorithmic frameworks that extract quantifiable writing style features to differentiate human- and machine-generated texts.
• They combine hand-crafted features such as lexical diversity, n-gram frequencies, and readability indices with classical, ensemble, and deep learning models to achieve high detection accuracy.
• These architectures are pivotal for authorship attribution, disinformation detection, and code provenance, ensuring authenticity and traceability across diverse text domains.

Stylometric Detection Architectures provide algorithmic frameworks for distinguishing between texts on the basis of quantifiable writing style features. These architectures are central to tasks such as AI-generated text detection, authorship attribution, code provenance, and information integrity assurance. Stylometric detection systems leverage hand-crafted or learned features spanning lexical diversity, syntactic complexity, n-gram distributions, punctuation usage, and higher-order linguistic statistics, coupled with machine learning models—ranging from linear classifiers to deep neural architectures and ensemble methods—to discriminate between human- and machine-generated texts or to identify authorial signatures.

1. Stylometric Feature Engineering and Formal Metrics

Detection architectures extract multidimensional feature vectors from input texts. The foundational families of features include:

• Lexical diversity: Type–Token Ratio (TTR) quantifies vocabulary breadth: $\mathrm{TTR} = V / N$, where $V$ is the number of unique types and $N$ the token count. Hapax Legomena Rate gauges the proportion of words appearing once: $\mathrm{HLR} = V_1 / N$.
• Vocabulary richness: Yule’s K and Honore’s R further capture lexical concentration and rarity via their respective formulas.
• Syntactic and POS statistics: Normalized counts or distributions over parts-of-speech ($p_j = C_{\text{POS}_j}/\sum_k C_{\text{POS}_k}$), morphological tags, or parse-tree depths.
• n-gram frequencies: Character $n$-gram ($n=2,3,4$) and word $n$-gram statistics.
• Readability indices: Flesch–Kincaid, Gunning Fog, and others are computed directly from word, sentence, and syllable counts, e.g.,

$$\mathrm{FKGL} = 0.39 \frac{W}{S} + 11.8 \frac{\sigma}{W} - 15.59$$

• Burstiness and variation: Coefficient of variation in sentence-level perplexity or sentence lengths, $\mathrm{Burst}_{\mathrm{CV}} = \sigma / \mu$.
• Surface cues: Punctuation entropy, connector-word and AI-specific phrase densities, averaged sentence and word lengths.

Comprehensive inventories in operational detectors often span 30–60+ engineered features, which are normalized (often by division with total token/sentence counts) to mitigate text-length bias (Al-Shaibani et al., 29 May 2025, Opara, 2024, Baidya et al., 18 Mar 2026).

2. Model Architectures: Classical, Ensemble, and Deep Learning

Detection architectures are instantiated via several canonical model classes:

• Classical classifiers: Linear SVMs and logistic regression operating on TF-IDF or stylometric feature vectors. Notably, SVMs on character $V$0-gram TF-IDF achieve strong baselines (Bitton et al., 3 Mar 2025).
• Tree ensembles: Random Forest and Gradient-Boosted Trees (e.g., XGBoost, LightGBM) are widely used with stylometric vectors. Their performance is competitive with deep models and their feature importances are directly interpretable (Ochab et al., 16 Jul 2025, Baidya et al., 18 Mar 2026).
• Neural networks: Shallow feed-forward networks (FFNN) over stylometric vectors or combined embeddings; deeper architectures are suited for large-scale or end-to-end learning but offer limited interpretability.
• Fine-tuned transformer encoders: RoBERTa, BERT, XLM-RoBERTa, and DeBERTa are commonly fine-tuned for classification with a linear head over the [CLS] embedding state. Some variants also support stylometric-feature fusion via concatenation and an auxiliary MLP (Al-Shaibani et al., 29 May 2025, Rezaei et al., 25 Nov 2025, Kumarage et al., 2023).
• Model ensembles: Architectures incorporating multiple heterogeneous classifiers (SVM, transformer head, FFNN) with unanimity voting achieve vanishingly low false-positive rates, as demonstrated by a 3-pronged ensemble with a FPR of 0.0004 (Bitton et al., 3 Mar 2025).

The table below summarizes representative architecture classes and their primary feature modalities.

Model	Feature Modality	Interpretability
SVM, LR	TF-IDF, stylometric	High
RF, XGBoost	Explicit stylometric	High
FFNN	Stylometric, embeddings	Moderate
Transformer	Raw text (optionally fused)	Low–Moderate
Ensemble	Mixed (lexical/syntactic/deep)	High

Best-in-class systems achieve F1-scores $V$1 in-domain; ensemble voting and explicit stylometry enhance cross-domain and adversarial robustness (Bitton et al., 3 Mar 2025, Baidya et al., 18 Mar 2026).

3. Detection Pipelines, Training Protocols, and Preprocessing

Standard stylometric detection pipelines comprise:

1. Preprocessing: Text normalization (Unicode, case folding), sentence splitting, tokenization, POS- and morphology-tagging, and (optionally) orthographic normalization (important for under-resourced languages like Arabic (Al-Shaibani et al., 29 May 2025)).
2. Feature Extraction: Programmatic computation of all stylometric, syntactic, and readibility features; postprocessing may include scaling (z-score, min-max) and feature selection (frequency filtering, L1 regularization).
3. Classifier Training: Supervised learning with train/validation/test splits suited to textual domain balance (commonly $V$2 or $V$3); early stopping and hyperparameter tuning via cross-validation or held-out splits.
4. Evaluation: Reporting class-wise and macro-averaged Accuracy, Precision, Recall, F1, and ROC-AUC; confusion matrices are analyzed to assess false-positive/negative rates (Opara, 2024, Al-Shaibani et al., 29 May 2025, Ochab et al., 16 Jul 2025).

Domain adaptation and multi-task learning—combining formal and informal datasets, or multiple prompt-generation strategies—are critical for generalization in low-resource and cross-domain settings. Adversarial robustness is evaluated with paraphrasing, domain-shift, and active learning strategies (Baidya et al., 18 Mar 2026, Al-Shaibani et al., 29 May 2025).

4. Generalization, Robustness, and Explainability

Empirical studies consistently show that:

• In-distribution performance: Fine-tuned transformers and boosted tree ensembles yield $V$4 F1 and ROC-AUC (Baidya et al., 18 Mar 2026, Ochab et al., 16 Jul 2025).
• Cross-domain/LLM shift: All models degrade under domain or LLM-source shift; stylometric and tree-ensemble models typically see smaller drops ($V$50.90–0.93 macro-F1/AUC) than deep-only models which may drop $V$6 points (Baidya et al., 18 Mar 2026, Li et al., 14 Oct 2025).
• Paraphrasing/humanization defense: Ensemble and stylometric-feature based detectors maintain higher robustness (AUC drop $V$75%) than pure neural; they leverage surface and syntactic cues less sensitive to lexical paraphrase (Li et al., 14 Oct 2025, Baidya et al., 18 Mar 2026).
• Abstention and coverage: Unanimous-vote ensembles prefer to abstain on unfamiliar distributions, conferring extremely low false-positive rates (0.0004) with high reliability at modest coverage (Bitton et al., 3 Mar 2025).
• Interpretability: SHAP and TreeSHAP are used to rank stylometric feature importances; key cues include sentence perplexity CV, function-word and connector densities, and rare word statistics. Feature ablation confirms that a small subset (e.g., UniqueWordCount, TTR, Hapax) accounts for most discrimination (Opara, 2024, Baidya et al., 18 Mar 2026, Bitton et al., 3 Mar 2025).

Explainable AI techniques, such as Integrated Gradients over transformer inputs, attribute classifier decisions to specific tokens or style markers, increasing transparency for high-stakes applications (Rezaei et al., 25 Nov 2025, Li et al., 14 Oct 2025).

5. Hybridization, Language and Code Domains, and Design Guidelines

Architectures are increasingly specialized by domain and use case:

• Low-resource and multilingual detection: Incorporation of stylometric priors and orthographic normalization are essential for languages with challenging morphology or limited resources. Multilingual or language-specific BERT-variants (e.g., XLM-RoBERTa, AraBERT) are recommended backbones (Al-Shaibani et al., 29 May 2025).
• Code stylometry: Structural ratios (comment density, identifier patterns), syntactic AST traversals, and shallow decision trees with hand-tuned heuristics have yielded resource-efficient detection of LLM-generated code, with macro-F1 up to 67.35% in cross-language benchmarks (Yotkova et al., 5 May 2026). Fusion with code-prompting and code-style embeddings further enhances vulnerability and provenance detection (Biringa et al., 29 Apr 2026).
• Ensemble/hybrid pipelines: Combining orthogonal feature modalities (discrete stylometric, BERT-based embeddings, n-gram overlaps) in a single XGBoost or transformer-fusion model consistently improves both in-domain accuracy and cross-domain transfer (Li et al., 14 Oct 2025, Ochab et al., 16 Jul 2025).
• Evaluation: Cross-model, cross-domain, and adversarially perturbed scenarios are mandatory for benchmarking; classic accuracy/precision/recall must be supplemented with abstention rates and misclassification costs (Baidya et al., 18 Mar 2026).

Design recommendations emphasize modular pipeline construction, explicit feature normalization, domain-adaptive training, conservative ensemble voting, and routine feature-importance auditing for continued robustness as LLMs and writing domains evolve (Al-Shaibani et al., 29 May 2025, Bitton et al., 3 Mar 2025, Baidya et al., 18 Mar 2026).

6. Stylometric Detection Beyond Natural Language: Authorship, Disinformation, and Stylometry Engineering

Stylometric architectures extend to tasks beyond AI versus human detection:

• Authorship attribution and similarity: Feature sets emphasizing readability indices, Burrows’ Z-scores, syntactic densities, and type-token profiles provide >0.96 F1 for binary author-similarity, often with XGBoost or SVM backbones (Kingsland et al., 2019).
• News bias and veracity: Unmasking meta-learning approaches quantify style similarity between hyperpartisan sources by training and iteratively stripping discriminative features, assessing the robustness and generalizability of style-based detectors (Potthast et al., 2017).
• Psycholinguistic stylometry: Features are mapped to cognitive dimensions (lexical retrieval, discourse planning, cognitive load, self-monitoring), and aggregated importance scores yield interpretable “psychoprofiles,” distinguishing machine and human patterns at accuracies >92% (Opara, 3 May 2025).
• Multilingual and genre-generalized stylometry: Large, language-tailored feature inventories (Polish: 172; English: 196) are normalized per token and can be combined with transformer embeddings for hate speech, genre, and topic detection in low-data or cross-lingual settings (Okulska et al., 2023).

These developments underline the flexibility and interpretability of stylometric detection architectures, which remain a crucial countermeasure to information disorder and AI-provenance ambiguity in contemporary textual ecosystems.

---

References

• (Al-Shaibani et al., 29 May 2025) The Arabic AI Fingerprint: Stylometric Analysis and Detection of LLMs Text
• (Bitton et al., 3 Mar 2025) Detecting Stylistic Fingerprints of LLMs
• (Ochab et al., 16 Jul 2025) StylOch at PAN: Gradient-Boosted Trees with Frequency-Based Stylometric Features
• (Baidya et al., 18 Mar 2026) Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions
• (Li et al., 14 Oct 2025) StyleDecipher: Robust and Explainable Detection of LLM-Generated Texts with Stylistic Analysis
• (Yotkova et al., 5 May 2026) FMI_SU_Yotkova_Kastreva at SemEval-2026 Task 13: Lightweight Detection of LLM-Generated Code via Stylometric Signals
• (Opara, 2024) StyloAI: Distinguishing AI-Generated Content with Stylometric Analysis
• (Rezaei et al., 25 Nov 2025) Generation, Evaluation, and Explanation of Novelists' Styles with Single-Token Prompts
• (Opara, 3 May 2025) Distinguishing AI-Generated and Human-Written Text Through Psycholinguistic Analysis
• (Potthast et al., 2017) A Stylometric Inquiry into Hyperpartisan and Fake News
• (Kingsland et al., 2019) Determining Individual Origin Similarity (DInOS): Binary Classification of Authors Using Stylometric Features
• (Okulska et al., 2023) StyloMetrix: An Open-Source Multilingual Tool for Representing Stylometric Vectors
• (Kumarage et al., 2023) Stylometric Detection of AI-Generated Text in Twitter Timelines
• (Biringa et al., 29 Apr 2026) VulStyle: A Multi-Modal Pre-Training for Code Stylometry-Augmented Vulnerability Detection