About

Ingram is a web interface for kveta, rhymeTagger, and sticho Python packages.

For any Czech or English poetic text(s) uploaded by user it provides:

• complex versification and linguistic analysis
• stylometric analysis comparing the analyzed text to the samples drawn from the Corpus of Czech Verse

Versification (and linguistic) analysis

Once user inserts poetic text or texts separated by a specified tag (see below) into the input textarea, versification analysis may be trigerred by pushing the button. Results for each poem are displayed as a separate tabs with following levels of annotation for each line:

• token: tokenization provided by MorphoDiTa for Czech and TreeTagger for English
• meter: poetic meter recognition provided by Kveta for Czech and Prosodic for English (limited to the most common meters, that is iamb, trochee, dactyl, and amphibrach in Czech; iamb, trochee, dactyl, and anapest in English)
• rhythm: distribution of stressed (1) and unstressed (0) syllables provided by Kveta for Czech and Prosodic for English
• rhyme: integer index marking rhyming lines, rhyme analysis provided by rhymeTagger
• lemma: lemmatization provided by MorphoDiTa for Czech and TreeTagger for English
• tag: morphological tags provided by MorphoDiTa for Czech and TreeTagger for English; for tagsets see [Czech] / [English]

The result is also available in machine readable JSON format.

Following parameters of versification analysis may be specified:

• separator: string used to separate patricular poems (default: <poem>)
• m-score: minimum weight a poetic meter has to achieve to be assigned to a poem; see more for Czech / English (default: 0.5)
• window: how many lines forward to look for a rhyming line; parameter of the same name in the RhymeTagger class (default: 6)
• P(sounds) / P(n-grams): minimum weights based on relation of final sounds/characters that pair of lines has to achieve to be annotated as rhyme; see here for details; probability_sampa_min & probability_ngrams_min parameters in the RhymeTagger class (default: 0.95).

Stylometric analysis

Once the versification analysis is performed the input text(s) may be compared to the samples drawn from the Corpus of Czech Verse or the Guttenberg English Poetry Corpus.

Settings

Stylometric analysis may be based on the following features (specified in the "settings" section):

• versification features:
• rhythmic types (frequencies of bitstrings representing distribution of stresses and unstressed syllables)
• rhythmic n-grams (frequencies of rhythmic types subtrings, substring length may be set from 2 to 5 positions)
• rhyme sounds (frequencies of sounds at particular positions in rhyming words)
• rhyme word-lengths (frequencies of rhymes represented as an unordered pair of rhyming word's lengths measured by number of syllables)
• rhyme stress placement (frequencies of rhymes represented as an unordered pair of distances of rhyming lines final stress from the end of line measured by number of syllables)
• sound frequencies (frequencies of particular sounds)
• lexical features:
• words (frequencies of word forms)
• lemmata (frequencies of lemmata)
• tags (frequencies of morphological tags)
• character n-grams (frequencies of character (2 to 5)-grams )

For each of the feature sets the value of MFT (most frequent types) may be specified (if MFT == 0, no limit is applied).

Two popular classifiers may be used for stylometric analysis: Burrows' Delta (nearest neitnghbour classifier) and the support vector machine (SVM). With Burrows' Delta one may specify the metric to be applied and the clustering technique that is going to be used in the dendrogram representing the results. With SVM one may specify the kernel to be used (linear/polynomial, in case of latter also its degree) and the penalty parameter.

Sampling

After pushing the "stylometric analysis" button, one may specify the size of the samples drawn from the corpus (100 / 300 lines). In next step the candidate set is defined by selecting 2 to 5 authors from the list. The number enclosed in parentheses gives the number of samples available by a given author. From these, 10 samples are then randomly selected for each author.

Result

For SVM leave-one-out cross validation is performed and accuracy is reported. Result of classification of the input text (TARGET) is given as a "VERDICT". Details are shown in the chord diagram, where nodes labeled in UPPERCASE represent the classes (authors) + input text (TARGET) and nodes labeled in lowercase represent particular samples. Edges represent particular predictions of cross-validation and the final verdict.

For Burrows' Delta accuracy is calculated as the portion of samples having sample by the same author as a nearest neighbour in a given vector space. Author of the sample that is the nearest neighbour of the input text (TARGET) is reported as a "VERDICT". Details are shown in dendrogram based on the metric specified in the "settings" section.

How to cite

Plecháč, Petr (2019). Ingram. Versification and stylometric analysis tool. Prague: Institute of Czech Literature, Czech Academy of Sciences. Available at <http://versologie.cz>.

Why "Ingram"?

John Kells Ingram (1823–1907) was an Irish mathematician and economist. In 1874 he published a pioneering article "On the weak endings of Shakspere, with some account of the history of verse-tests in general" using versification analysis in the process of authorship recognition.

read more at wikipedia