Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics

This article introduces two novel measures for authorship attribution—Rank-Turbulence Delta and Jensen–Shannon Delta—which generalize Burrows’s classical Delta by employing distance functions derived from probabilistic distributions, thereby providing a more interpretable framework for stylistic analysis.

The gist: Rank-Turbulence Delta attains attribution accuracy comparable with Cosine Delta; Jensen–Shannon Delta consistently matches or exceeds the performance of canonical Burrows’s Delta.

Theoretical Foundation and Vector Representation

The study begins by re-casting uncentred word-frequency vectors as probability distributions, which allows for the application of distance measures from information theory and complex-systems analysis. The canonical Burrows’s Delta compares standardized (z-scored) word-frequency profiles across texts, defined by the Manhattan (L1) distance between these z-vectors. However, uncentred z-vectors can be normalized to represent probability distributions where each component is the relative weight of a token within the stylistic profile of the text. This transformation opens the possibility of employing divergence measures such as Jensen–Shannon divergence and Rank-Turbulence divergence.

Probabilistic Distance Measures

The paper develops probabilistic distance measures based on these probability distributions. The Jensen–Shannon divergence compares two texts not directly to each other, but to their weighted average, where the mixture weights are chosen such that the sum equals one. The Kullback–Leibler divergence is also introduced as an element of the Jensen–Shannon divergence. Furthermore, for rank-based representations, the rank-turbulence divergence is proposed to compare hierarchical ordering of tokens using a formula involving ranks and a tuning parameter, where smaller values of alpha amplify the contribution of lower-ranked tokens.

Token-Level Decomposition and Interpretation

A systematic token-level decomposition is developed to render every Delta distance numerically interpretable by identifying the token-level contributions. For canonical Burrows’s Delta, the contribution of a token is defined as:

(12) (1) (2) p p    = −

This decomposition allows researchers to visualize which individual words exert the greatest influence on stylistic differences between texts. For Cosine Delta, contributions arise from products rather than sums and can be positive or negative, with specific interpretations based on whether components exceed or fall below their corpus means. Similarly, for Jensen–Shannon Delta and Rank-Turbulence Delta, token contributions are extracted in a manner analogous to the canonical Burrows’s Delta formula.

Experimental Validation and Robustness

The effectiveness of these methods is assessed across four literary corpora in English, German, French, and Russian. Experiments on clustering quality demonstrate that Rank-Turbulence and Cosine Delta show consistently high performance across languages and mfw ranges, suggesting stable representations of stylistic structure. The study also tests the robustness of these measures under temporal and stylistic variation using the SOCIOLIT corpus. Robustness checks include:

1. Variation of most frequent words (mfw) size, showing that Jaccard overlap between adjacent mfw settings remained consistently substantial across metrics.

2. Document-level bootstrap resampling, where the mean Jaccard overlap of the top-30 tokens ranged between approximately 0.37 and 0.50, indicating a stable signal rather than sampling noise.

3. A removal experiment confirming that removing top-contributing tokens led to a consistent reduction in inter-author distance across all four metrics, confirming their functional role in document separation.

Discussion on Stylistic Significance

The findings suggest that mid-frequency words play a particularly important role in stylistic discrimination, as clustering quality is maximized at lower values of the rank parameter alpha. The research further reveals an asymmetry between lexical presence and absence: while centred representations show strongly negative deviations tend to occupy lower positions in the rank hierarchy, the reorganization of prominent lexical signals appears more structurally informative for document separation. This indicates that mid-frequency vocabulary carries substantial signal, reinforcing that neither exclusively dominant nor exclusively rare items suffice to capture stylistic identity. The proposed framework provides a structured quantitative map of lexical contrasts that guides subsequent qualitative analysis.

Limitations and Future Directions

Limitations include the context-dependent nature of parameter selection, specifically for alpha, which may require corpus-specific stability analysis. While the methods cover multiple languages and corpora, further testing on heterogeneous historical datasets is suggested. Additionally, the transformation required for Rank-Turbulence divergence involves an explicit reordering of features with asymptotic complexity comparable to sorting operations. The proposed interpretative framework is noted as a tool to guide qualitative analysis rather than replacing close reading.

Data and Code Availability

Code and data used in this study are publicly available at https://github.com/DDPronin/Rank-Turbulence-Delta, with metadata and derived representations provided in the repository. The preprocessing pipeline involved standardizing term-frequency matrices using the 20,000 most frequent word types across corpora, followed by computing both centred and uncentred z-standardized representations.

As a fastidious and diligent researcher, I have analyzed this paper, RANK-TURBULENCE DELTA AND INTERPRETABLE APPROACHES TO STYLOMETRIC DELTA METRICS. The core contribution of this work lies in moving beyond purely predictive authorship attribution towards stylometric analysis that prioritizes transparency and interpretability.

Here are the specific improvements to AI systems that can be made using the methodologies described in this paper, along with what those improved systems can achieve:

1. Improvement: Integration of Token-Level Stylistic Fingerprinting via Rank-Turbulence Delta (RTD)

2. Improvement: Implementation of a Token Contribution Mapping layer for any text classification or attribution model trained on stylometric features.

3. What the Improved AI System Can Do:

4.1. Identify and quantify the specific lexical items (words or n-grams) that are driving a particular textual similarity score (e.g., between two texts, an author and their corpus). This moves the system from merely stating Text A is similar to Text B to explaining Text A is similar to Text B because of the high contribution of tokens X, Y, and Z.

4.2. Enable targeted stylistic probing: Researchers can use this mapping to test hypotheses about authorial style (e.g., Does the difference between Author 1 and Author 2 hinge on the usage frequency of 'pacing' verbs vs. adverbs?).

4.3. Enhance interpretability in complex attribution tasks: For unsupervised or exploratory tasks where high predictive accuracy isn't the sole goal, this layer provides a transparent bridge between high-dimensional vector distances and meaningful linguistic features, facilitating close reading and validation against philological intuition (as noted in Section 1).

2. Improvement: Adoption of Probabilistic Distance Metrics (Jensen–Shannon Delta) for Similarity Assessment

3. Improvement: Replace standard Euclidean or Cosine distance measures with Jensen–Shannon Divergence (JSD) when comparing text distributions, especially in contexts involving varying corpus sizes or when the underlying feature space is not perfectly Gaussian.

4. What the Improved AI System Can Do:

4.1. Provide a more robust measure of stylistic similarity that is intrinsically based on probability theory (Section 3). JSD measures how much information would be lost if one text were approximated by a mixture distribution of the two texts, offering a symmetric and information-theoretic view of divergence (Equation 8).

4.2. Maintain performance stability across different data regimes: By using probabilistic representations derived from uncentred z-vectors, the system maintains structural integrity even when dealing with non-negative distributions, which is crucial for consistency across multilingual corpora (English, German, French, Russian).

3. Improvement: Dynamic Feature Weighting via Rank-Turbulence Parameter Tuning

4. Improvement: Incorporate a tunable parameter (like the rank sensitivity parameter α) into the distance calculation framework to allow the model to dynamically shift its focus between high-frequency dominant features and subtle mid-frequency signals.

5. What the Improved AI System Can Do:

5.1. Optimize performance for specific stylistic nuances: The system can be tuned to prioritize structural vocabulary (high frequency, low α) or distinctive lexical signals (mid-frequency, moderate α). This allows the AI to adapt its sensitivity based on the research question—whether it seeks broad stylistic tendencies or idiosyncratic markers.

5.2. Achieve stable clustering and classification: Experimental results show that optimal performance is often found at intermediate values of α, suggesting a more nuanced understanding of authorial contrast than relying solely on extreme high-frequency items, leading to more robust clustering solutions (Section 10).

4. Improvement: Robustness Checks for Feature Selection (mfw)

5. Improvement: Implement automated robustness checks that assess the stability of the top-contributing tokens against minor perturbations in vocabulary size (mfw) and document sampling variation (bootstrap resampling).

6. What the Improved AI System Can Do:

6.1. Ensure reliable feature selection: The system will not rely on a single, narrowly defined set of most frequent words. By confirming that the top-contributing tokens remain stable across varying mfw sizes (as shown in Appendix C), the resulting attribution or clustering decisions are proven to be based on a stable structural property of the authorial contrast, not an artifact of arbitrary vocabulary cutoff choices.

6.2. Increase confidence in attribution results: The system can report a confidence score based on the Jaccard overlap with adjacent mfw settings, providing assurance that the identified fingerprint tokens are genuinely discriminative and not sampling noise (Section 19).

In summary, these improvements transform a stylometric AI from a simple distance calculator into an advanced, explainable analytical tool capable of identifying, quantifying, and interpreting the precise lexical components responsible for stylistic variance.