Understanding Rhyme Through Network Analysis

Understanding Rhyme Through Network AnalysisRhyme enacts numerous relationships in poetic texts: relationships between words that share similar sounds; relationships between lines of verse that end in rhyming words; and relationships between the sound and semantic meaning of words that are linked together through rhyme. This panel brings together different approaches to using network analysis to understand the relationships that rhyme enacts in poetry in different language traditions. As the papers in this panel suggest, as one moves from considering rhyme’s function within a single poem to examining larger datasets, one can also consider how rhyme connects words, documents, and/or authors within a corpus. By juxtaposing papers focusing on poetry in Czech, English, Russian, and Spanish, this panel highlights the fact that rhyme is defined differently in different linguistic and poetic contexts, due to the levels of inflection present in different languages and to the development of different poetic traditions. Some rhyme definitions focus on the shared stressed vowel of single syllables, others encompass polysyllabic rhyme, and others focus on component phonemes, rather than whole words. Thus computational approaches to rhyme must be tailored to the particular languages of the texts under study. The papers on this panel use different network analysis and graph visualization methods to examine rhyme at the level of the corpus or dataset, rather than the individual poem, in order to understand how rhyme practice changes over time, across languages, and in relation to literary canon formation. It thus contributes both to computational poetics and distant reading methodologies within the digital humanities.   Distant Reading Nineteenth-Century British Poetry With Rhyme NetworksNatalie M. HoustonUniversity of Massachusetts Lowell1. IntroductionThe expectations and assumptions that nineteenth-century English readers brought to their reading of poetry was necessarily different from that which readers today bring to the same texts. One feature of that historical difference was their familiarity with poetic rhyme, and the assumptions about poetic language that it created. By examining rhyme words and sounds in a large dataset of English poems, we can better understand how rhyme shaped poetic discourse in the nineteenth century. This paper suggests that network analysis methods are useful for understanding the semantic networks created by the relatively limited set of rhymes available in English; for examining chronological and aesthetic explanations for changes in rhyme practice; and for exploring the relationships between poems that use the same rhyme pairs. Such analyses reveal the semantic and sonic features of conventional nineteenth-century poetry, and can thereby also distinguish unconventional or distinctive uses of rhyme.2. Context The vast majority of English poems published in the nineteenth century were rhymed (95% of the 108,842 poems in the Chadwyck-Healey English Poetry corpus used for this study). Both poets and readers thus expected poetry to be rhymed (McDonald 2012, 7) and many of the rhyme sounds and rhyme words used in nineteenth-century poetry were so frequently used as to create a set of implicit conventions of poetic discourse. Uncovering such implicit conventions can help reveal the structures of the field of poetry at the large scale (Bourdieu 1993). This project identifies rhyme words, groups, and syllables through a method that operationalizes the historical rules for rhyme found in nineteenth-century British rhyme dictionaries, in order to match words according to historical pronunciation and poetics (Houston 2016, 2019). Three different kinds of network analyses are then performed: a rhyme word co-occurrence network, a rhyme pair co-printing network, and a textual coupling network (Houston 2017).4. Rhyme Networks Because rhyme word frequency in British poetry follows a power law distribution, in which a small number of rhyme words are very frequently used, followed by a long tail of additional words, the rhyme word co-occurrence network can reveal the relationship between those frequencies and the clusters of rhyme words that are most likely to occur within the same poem. The rhyme pair co-printing network, based on co-citation analysis, links specific rhyme pairs if they appear in the same poem. Together these two networks reveal the semantic and sonic patterns that structured nineteenth-century poetic discourse. The textual coupling network, based on bibliographic coupling, links two poems if they use the same rhyme pairs. This network reveals chronological and aesthetic subgroups, suggesting how rhyme practice changed through the century. The talk employs network analysis in order to explore to what extent rhyme pairs are shared across works by different authors in nineteenth-century Czech poetry. We show that the degree of recurrence is (1) comparable to Russian poetry, (2) noticeably weaker as compared to German and Spanish poetry and (3) strikingly weaker as compared to English poetry. Following the hypothesis formulated in Plecháč 2018, we offer a linguistic explanation for such findings: the size of rhyme repertory in a given language depends upon its inflection, the changes in word forms used to mark grammatical aspects such as voice, case, tense, mood, number, or gender. Roughly speaking—the more suffixes the language employs, the richer rhyme repertory gets as completely different words may rhyme only due to being followed by the same grammatical endings.As one may expect, we find that in all the languages examined the tendency to share rhyme pairs is stronger between works that come roughly from the same period than between those distant in time. We thus try to represent each work as a vector defined by the relative frequencies of rhyme pairs and use machine learning techniques (Random Forest, Support Vector Machine) in order to classify them into time periods or poetic movements. Cross-validations of the models show that depending on what classes are being used, the accuracy varies between 0.6 to 0.9 and always outperforms the random baseline.