Introduction

In recent years, the use of rhythm-based features in authorship recognition of poetic texts has been studied thoroughly in several languages, including Czech, German, Spanish (Plecháč; Plecháč & Birnbaum), Portuguese (Mittmann), Latin (Nagy), Old English (Neidorf et al.), and Russian (Šeļa; Orekhov). In this contribution we expand our analysis of the relationship between versification type and authorship recognition accuracy provided in Plecháč & Birnbaum by performing an experiment with poetic texts in six languages: Czech (CS), German (DE), English (EN), Spanish (ES), Italian (IT), and Russian (RU).

Method

From each corpus we extract one or more subcorpora containing 11-syllable lines (except for EN, where we use 10-syllable iambic pentameter) by five authors born in a specific time span (see Table 1). Ten 100-line samples are drawn at random for each author and each sample is represented by bitstrings that encode the stressed and unstressed syllables in particular lines (e.g., The curfew tolls the knell of parting day ~ 0101010101). Leave-one-out cross-validation is performed to evaluate Support Vector Machine models (linear kernel) for bitstring-based authorship recognition. The entire procedure is repeated 10 times, resulting in 10 accuracy estimations for each subcorpus.

Results

Figure 1 shows the results. All subcorpora significantly outperform the random baseline (0.2), yet there are substantial differences across languages.² We hypothesize that one of the key factors may be the versification type: while in syllabic versification only the number of syllables in a line is constrained and the distribution of stress is left to the author’s preferences, in accentual-syllabic versification both syllable count and stress placement are subject to constraints, which leaves considerably less space for authors to individualize their rhythm.³

Figure 1: Rhythm-based authorship recognition accuracy; 30 random samplings per subcorpus; leave-one-out cross-validation; linear SVM. Boxplots are constructed in a following way: the box gives the interquartile range (IQR) of the 30 samplings; the horizontal line gives the median value; whiskers are plotted at the 1.5 IQR values; data beyond the whiskers are plotted as individual points.

Interpretation

At first glance, this hypothesis fails to explain the differences. Accentual-syllabic CS ranks among the best scoring languages and accentual-syllabic EN outperforms part of syllabic ES. We need, however, to keep in mind that the traditional categorical notion of versification types is misleading-it is, rather, a continuous scale (cf. Gasparov). To formalize the degree of accentual regularity, we measure the entropy of bitstrings in particular subcorpora:

As Figure 2 shows, there is some sort of association between accentual regularity and the accuracy of authorship recognition as we proceed from strictly organized RU and DE; across semi-organized CS, EN, and ES; to loose IT, the accuracy tends to increase (linear regression, R ² = 0.5).

Figure 2: Relationship between rhythm-based authorship recognition accuracy and bitstrings entropy; linear regression (R 2 = 0.5)

Conclusions

The experiment described here yields two types of results. The first, which extends our earlier work in Plecháč & Birnbaum, is a demonstration that rhythmic organization (and, specifically, accentual regularity) can function as a meaningful feature for authorship recognition. The second is our observation that the categorical identification of verse traditions as either syllabic or accentual-syllabic masks the actually continuous nature of the feature, and we propose a formula that remedies that limitation by quantifying the accentual regularity of verse corpora as a continuous property.