Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics

summary

The gist

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

In short

The study introduces two new measures for authorship attribution: Rank-Turbulence Delta and Jensen–Shannon Delta. These methods generalize classical Burrows’s Delta using probabilistic distance functions derived from word frequency distributions. The results show that Rank-Turbulence Delta performs comparably to Cosine Delta, while Jensen–Shannon consistently matches or beats the traditional Burrows’s Delta, offering a more interpretable way to analyze stylistic differences.

Key concepts

Rank-Turbulence Delta: This is a novel authorship attribution metric based on comparing word frequency profiles using a divergence measure derived from rank-based representations. It uses a formula that incorporates token ranks and a tuning parameter to compare the hierarchical ordering of words between texts, allowing for more nuanced stylistic comparisons.
Jensen–Shannon Delta: This is another new metric that compares two texts based on their probability distributions. It measures the divergence between two distributions by comparing each text to their weighted average, where weights sum to one. It is built upon the Kullback–Leibler divergence and provides a robust way to compare stylistic profiles.
Token-Level Decomposition: This technique breaks down the final Delta distance calculation to show exactly which individual words contribute most significantly to the difference between two texts. This helps researchers visualize which specific vocabulary items are driving the stylistic distinction, making the results more meaningful.
Probabilistic Distance Measures: These are mathematical tools used to quantify how different two word frequency distributions are. Instead of simple subtraction, they use information theory concepts like divergence to measure stylistic distance based on how likely certain words are to appear in each text.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 24 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes
Sample Complexity — explained in 15 episodes
Federated Learning — explained in 14 episodes
Generalization — explained in 13 episodes

This episode discusses

Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics · Paper Radio

The paper

Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics · Read on arXiv

Dmitry Pronin, Evgeny Kazartsev

HSE University

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics".

Tom: 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,

Jane: First, who's behind it and why it matters.

Paper summary: Tom: Essentially, this paper is proposing Rank-Turbulence Delta and Jensen–Shannon Delta as new tools for authorship attribution because they generalize Burrows’s classical Delta by using distance functions based on probabilistic distributions. The authors claim these methods offer a more interpretable framework for stylistic analysis, which is a big deal when you need to understand *why* two texts are different.

Jane: They start by re-casting uncentred word-frequency vectors as probability distributions, which then lets them apply distance measures from information theory and complex systems analysis instead of just using the standard Manhattan distance on z-scored vectors, which is what Burrows’s Delta typically uses.

Lu: The paper goes into developing probabilistic distance measures, like Jensen–Shannon divergence and Rank-Turbulence divergence, and they also introduce a token-level decomposition that makes every Delta distance numerically interpretable by showing the "token-level contributions."

Meng: That token-level decomposition sounds promising for practical implementation because it lets researchers pinpoint exactly which individual words are driving the stylistic variation between texts.

Lalam: I think that ability to see the specific tokens contributing allows for a much deeper level of textual understanding, moving beyond just saying "Text A is different from Text B."

Tom: Exactly. They show that Rank-Turbulence Delta achieves attribution accuracy comparable to Cosine Delta, and Jensen–Shannon Delta consistently matches or even exceeds the performance of canonical Burrows’s Delta. The authors test this across four literary corpora in English, German, French and Russian.

Jane: Their experimental validation is quite thorough; they assess clustering quality and attribution performance on these diverse datasets while also testing the robustness of their methods under temporal and stylistic variation using the SOCIOLIT corpus.

Lu: The results suggest that Rank-Turbulence Delta shows consistently high performance across languages and word frequency ranges, which points toward a stable representation of stylistic structure in these metrics.

Conclusion: Tom: So, wrapping up this discussion on "Rank-Turbulence Delta and Interpretable Approaches to Stylometric Delta Metrics," the main point is that these new metrics offer a structured quantitative map of lexical contrasts, which helps guide qualitative analysis after the initial statistical comparison.

Jane: The authors have really expanded on Burrows’s original concept by introducing probabilistic measures that allow for a clearer look at how word frequencies actually contribute to stylistic separation in a way that's more transparent.

Lu: What I find particularly interesting is the finding about mid-frequency words; the study suggests these play a particularly important role in stylistic discrimination, especially when you look at lower values of the rank parameter alpha.

Meng: From an engineering standpoint, knowing which specific tokens are driving that separation means we can build systems that are more sensitive to those key lexical signals without just treating all word frequencies equally.

Lalam: If this framework helps us understand the asymmetry between lexical presence and absence, it could help us design AI models that better capture nuanced cultural expression rather than just surface-level vocabulary counts.

Tom: It really shows that neither exclusively dominant nor exclusively rare items are sufficient to capture stylistic identity; mid-frequency vocabulary carries substantial signal in these new approaches. This has big implications for how we model human creativity and language use.

Jane: Ultimately, the implication is a more sophisticated way to analyze authorship attribution because it provides not just an accuracy score, but also a pathway to understanding the underlying linguistic structure driving that score.

More episodes

← Home