Detecting Rhyme and Meter in Hungarian Poetry: From Algorithms to Web Tools and Research

4.1 Main Principles of Detecting Hungarian Quantitative Meter

The input for the program detecting Hungarian quantitative meter consists of the TEI XML files from the ELTE Poetry Corpus, which contain automatically annotated syllable lengths for each line, represented as a sequence of 0s and 1s (e.g., 01010101). Syllable lengths in Hungarian poetry can be determined by applying a set of general rules, which are the following (see Szepes and Szerdahelyi 1981).

1. A syllable is short if it contains a short vowel and is not followed directly by a consonant, or if it is followed by only a short consonant.

2. A syllable is long if it contains a long vowel, or if a short vowel is followed by a long consonant or two or more consonants.

3. Word-initial consonant clusters (e.g. in krákog, strigula) do not lengthen the preceding syllable ending in a short vowel.

P. Horváth et al. (2022) presented an evaluation of syllable length detection. The percentage of incorrectly annotated lines in the sample was 2.33%. However, the error rate is slightly influenced by the time of the poems. For poems written between 1505 and 1701, it was 3.5%, while for those from 1772 to 1854 and 1855 to 1909, it was 1.5% and 2.0%, respectively.

The algorithm can recognize four types of quantitative meter: dactylic, anapestic, iambic, and trochaic. In addition to the label of the meter type, the poems also get a regularity score between 0 and 1. The regularity score reflects how consistently the rhythm realizes the meter in the poem. The higher this score, the more regularly the poem’s meter is realized. By using regularity scores, it is possible to grasp meter as a scalar phenomenon instead of as a binary category (yes or no). Since poems can realize quantitative meter types in varying degrees, it seemed reasonable to design the algorithm to indicate this degree of regularity in the output. The regularity score allows for the use of a threshold to exclude poems with low regularity from the analysis.

The program tests the poems for all four meter types and gives the label of the meter category for which it has the highest regularity score. To achieve this goal, the algorithm first divides the sequences of 0s and 1s, which represent short and long syllables, into metrical feet. This segmentation is based on different principles for dactylic and anapestic and for iambic and trochaic meters. While in the case of dactylic and anapestic meters, the segmentation is based on the number of moras, for iambic and trochaic meters, the segmentation is based on the number of syllables. When testing dactylic and anapestic meters, the algorithm segments lines into four-mora feet, allowing the last foot to be incomplete with fewer than four moras. Mora is the abstract time unit in Greek, Latin, or Hungarian quantitative meter systems based on syllable length. Short syllables represent one mora, while long syllables represent two moras. In the case of dactyls, anapests, and spondees, the sum of the syllables’ moras in the feet is four. When testing iambic and trochaic meters, the program segments lines into two-syllable feet, permitting the last foot to be incomplete with only one syllable.

The regularity score of the poems is determined based on a scoring system. Each complete foot gets a score from 0 to 4, depending on the tested meter type. The base feet of the tested meter type receive 4 points, the primary substitute feet receive 2 points, and the secondary substitute feet receive 1 point. The opposite feet of the tested meter get 0 points. For instance, if the tested meter is iambic, iamb receives 4 points as the base foot of the meter type, spondee gets 2 points as the primary substitute foot, pyrrhus gets 1 point as the secondary substitute foot, and trochee does not get any points because it is the opposite foot of the tested meter. Table 3 shows the scoring system for the four meter types.

In calculating the regularity score of the poem, the algorithm first calculates the regularity score of each line for the tested meter type. The regularity score of the line is the mean score of the complete feet in the line. Based on the scoring system presented above, the formula for calculating the regularity score of a line is as follows (where b is the number of base feet, p is the number of primary substitute feet, s is the number of secondary substitute feet, and n is the total number of complete feet in the line):

$\text{line score}=\frac{4b + 2p + s}{n}$

However, for a line to receive a regularity score greater than 0 for a given meter type, it must meet certain minimum conditions. These conditions, based on the norms of Hungarian versification system, are the following.

1. Minimum conditions for anapestic and dactylic lines: (1) All the complete feet consist of four moras.⁴ (2) There is no amphibrach among the feet.

2. Minimum conditions for iambic lines: (1) If the last foot of the line is complete, then it is an iamb or a pyrrhus. (2) If the last foot is incomplete, then the penultimate foot (the last complete foot) is an iamb.

3. Minimum conditions for trochaic lines: (1) If the last foot of the line is complete, then it is a trochee or a spondee. (2) If the last foot is incomplete, then the penultimate foot (the last complete foot) is a trochee.

If the line does not satisfy the minimum conditions for the tested meter type, the regularity score of the line will be zero. In the second step, the algorithm calculates the regularity score of the whole poem for the tested meter by dividing the sum of the regularity scores of the lines by the number of lines. In the final step, the program compares the four regularity scores obtained for the four meter categories and labels the poem with the meter category having the highest regularity score.

4.2 Main Principles of Detecting Hungarian Qualitative Meter

Since the Hungarian qualitative meter system is based on different principles than the quantitative one, a second algorithm had to be developed. The output of the algorithm is the bar structure of the poem. The algorithm can recognize two main types of qualitative meter. The first one is when all of the lines in the poem have the same bar structure. For instance, the output 4-4 means that each line consists of two four-syllable bars. Naturally, lines can have more than two bars. For example, when the output is 4-3-2, each line of the poem has three bars, the first one contains four syllables, the second one has three syllables, and the third bar has two syllables. The second main type of bar structure detected by the program is when the odd and even lines of the stanzas have different bar structures. This means that two bar structures alternate from line to line. For instance, the output 4-4_4-3-2 indicates that the odd lines of the stanzas contain two four-syllable bars, while the even lines consist of three bars: one with four syllables, one with three, and one with two. There are other, more complex types of Hungarian qualitative meter as well; however, these two main types cover the majority of Hungarian poems with qualitative meter.

A bar can consist of only one syllable. In the program’s analysis, the maximum number of syllables per bar is six. This means that if the poem’s lines have fewer than seven syllables, the lines themselves can form a bar without any inner segmentation. For instance, an output of 6 indicates that each line of the poem has six syllables, and these six-syllable units cannot be further divided into smaller bars.

The main principle behind the algorithm is the use of matrices to define the bar structure. The rows of the matrix represent the lines, and the columns represent the syllable positions of the lines. In the matrix, only two values are used: 1 and 0. The value 1 represents syllables that are the first syllables of words (stressed syllables), while the value 0 represents all other syllables (unstressed syllables). The algorithm determines the first syllables of the bars by calculating the mean values of the columns. If this value is at least 0.75, the syllable position is analyzed as the beginning of a new bar. The 0.75 threshold reflects that there can be lines that do not follow the general bar structure of the poem. In the algorithm, there is a restriction on the second syllable position: this cannot be the beginning of a new bar, regardless of its mean value. The reason for this restriction is that defining single-syllable bars at the beginning of lines would be inconsistent with Hungarian versification. Not to mention that these single-syllable words at the beginning of the lines are usually unstressed function words.

4.3 Frequency and Regularity of Meter Types in Canonical Hungarian Poetry

These two algorithms implemented in a Python program called Hunpoem_meter_analyzer have been used to annotate the quantitative and qualitative meters of the poems in the ELTE Poetry Corpus (see level3 and level4 XML files in the repository of the corpus). Table 4 shows the number of poems with different meters. Only those quantitative meter poems are counted that have a regularity score higher than 0.5. The third column indicates the number of poems that have qualitative and quantitative meters simultaneously. For example, 1370 poems follow qualitative and iambic meters at the same time. The total number of these simultaneous poems is 2554, using the threshold of 0.5. Table 5 shows the same, but with a threshold of 0.3.

The regularity scores of the poems’ rhythm allow for the analysis of changes in rhythmic regularity throughout the history of canonical Hungarian poetry. Figure 1 shows the median regularity scores of iambic poems for authors from the 19th century and the first half of the 20th century, arranged chronologically. Only those poets who have at least 100 poems have been included in the analysis. Figure 2 presents the median scores for poems written in trochaic meter. A threshold of 0.3 was applied in both cases. Using black and gray bar colors, the bar charts differentiate the median regularity scores of authors born before and after Endre Ady (1877–1919), considered the first poet of Hungarian classical modernism at the beginning of the 20th century. As the bar charts indicate, the rhythmic regularity of both iambic and trochaic poems declined in the first half of the 20th century.⁵ This decline appears to align with the rise of classical modernism in Hungarian poetry.

Figure 1: Median regularity scores of iambic poems over time (Threshold = 0.3).

Figure 2: Median regularity scores of trochaic poems over time (Threshold = 0.3).

4.4 Evaluation of the Algorithms Detecting Hungarian Quantitative and Qualitative Meters

Unfortunately, there is no gold-standard corpus of Hungarian poems with manually annotated meter. Therefore, to evaluate the efficiency of the algorithms detecting meter, an alternative method had to be developed. Instead of building a costly gold-standard corpus, a database was created that lists the example poems from the book by Szepes and Szerdahelyi (1981). This is one of the most comprehensive handbooks on Hungarian poetry meters, and the two algorithms developed for detecting Hungarian quantitative and qualitative meters are largely based on its approach. The book provides numerous examples for each meter type. I have recorded the titles of the poems, along with their meter types, for those that are also included in the ELTE Poetry Corpus. In other words, the book’s examples, which illustrate different meter types, serve as a quasi-gold-standard corpus. In this way, 78 poems with quantitative meter and 55 poems with qualitative meter were collected. Only poems with an unambiguous meter, as identified by the authors, were recorded. For example, poems in which different stanzas follow different meters were not included in the list.

Table 6 shows the number of matches and their proportion (accuracy) for the four quantitative meters. Accuracy was calculated at regularity score thresholds of 0.5, 0.4, 0.3, 0.2, and 0.0. The lower the threshold value, the greater the number of poems with a correctly recognized meter. However, it is important to highlight that different meter types respond differently to threshold values. With a threshold of 0.5, all but one of the trochaic poems and 70% of the iambic poems are recognized, whereas only about half of the anapestic and dactylic poems are identified. This suggests that for anapestic and dactylic meters, a lower threshold should be used to capture a higher number of poems. Four of the 78 poems cannot be correctly recognized, even without a threshold (threshold = 0.0). In two of these cases, there is an internal caesura after three and a half feet, which the program is unable to handle.

It is worth noting that, in this case, accuracy based on exact match is equivalent to recall. Precision is 1.0 for iambic, anapestic, and dactylic poems across all threshold values. For the trochaic meter, precision is 1.0 at the 0.5 threshold, but at the 0.4 threshold, it drops to 0.88 and further decreases to 0.79 at lower thresholds. This indicates that, within the sample used, only the trochaic label is applied more broadly than necessary. In other words, all four incorrectly annotated poems are annotated with the trochaic label. However, the low and unbalanced number of suitable example poems in the book prevents obtaining a reliable precision value. Consequently, reporting F-scores would also be misleading.

Table 7 shows the number and proportion of matches for the qualitative meter. The second column indicates the exact matches, which occur when the bar structure given by the program is the same as the one provided in the book. The third column also includes matches where the program provides a more general or more specific bar structure than the book (e.g., 6-6 instead of 4-2-4-2 or 5-3-2 instead of 3-2-3-2). In this latter case, 50 out of 55 poems (91%) have been analyzed in the same or a similar way as by the authors of the book.

5.1 Data Types in the Rhyming Dictionary

The number of rhyming pairs in the rhyming dictionary is 650,561. Besides the word forms forming the rhyming pairs, the database contains three types of data. On the one hand, it includes the grammatical and phonological features of the words in the rhyming pair, such as lemma, part of speech, morphosyntactic features, number of syllables, vowel type, and phonological structure. These properties were retrieved from the annotations of the ELTE Poetry Corpus. The grammatical features were annotated using the e-magyar tool (Indig et al. 2019; Váradi et al. 2018), while the phonological features of the words were annotated with a program specifically developed for the ELTE Poetry Corpus (P. Horváth 2020).

On the other hand, the rhyming dictionary also provides information on the position of the rhyming words. In the case of each rhyming pair, the dictionary indicates the number of lines between the rhyming words (this can be a maximum of four lines), the order of the rhyming words, and the number of those lines between the rhyming words that also rhyme with them. Finally, the dictionary includes the bibliographical data of the poem containing the rhyming pair, as well as a URL of the poem pointing to the text query interface of the ELTE Poetry Corpus.

The dictionary lists the rhyming pairs in alphabetical order. The alphabetic sorting of the rhyming pairs is based on the following principles.

1. The rhyming pairs are sorted according to lemmas.

2. In the case of rhyming pairs with the same lemmas but different parts of speech, the rhyming pairs having the same part of speech are listed consecutively.

3. Rhyming pairs with the same lemmas and the same parts of speech are sorted according to word forms.

4. From the occurrences realizing a rhyming pair, those items are listed first in which the distance between the members of the rhyming pair is smaller.

5.2 The Rhyming Dictionary’s Online Query Tool

For the rhyming dictionary, an online query tool has been developed.⁶ The backend of the tool was programmed in the FastAPI framework of the Python programming language, and it queries the SQLite relational database of the dictionary. The tool allows searches by lemma and word form, with results filterable by part of speech, author, and the position of the rhyming word within the rhyming pair. In addition to the word forms of the rhyming pairs, the query displays the lemmas, the parts of speech, the distance between the members of the rhyming pairs, and the number of other rhyming words between the members. The results also include bibliographical information about the poem in which the rhyming pair occurs. By clicking the link shown in the output, the poem opens in the text query interface of the ELTE Poetry Corpus.⁷ The search results can be downloaded in TSV format. By using the Statistics function, the number of lemmas or word forms rhyming with the specified word can be displayed in a tabular format. The table also shows how many authors have at least one poem that contains these lemmas or word forms rhyming with the specified word.

5.3 Changes in the Use of Inflectional Rhymes in Hungarian Poetry

Using the data of the rhyming dictionary, various aspects of rhyming in Hungarian canonical poetry can be explored. To demonstrate the dictionary’s potential for research, a small case study on inflectional rhymes is presented here. According to the norms of Hungarian poetry, inflectional rhymes should be avoided. However, as some observations suggest, this rule has not always been (strictly) followed. The acceptance of inflectional rhymes has varied across different periods. For example, Maróthy et al. (2021) and Seláf and Plecháč (2023) quantitatively demonstrated that inflectional rhymes were commonly used in 16th-century Hungarian historical songs. However, until now, it has been difficult to provide a precise overview of how their use has changed in Hungarian poetry in general. The rhyming dictionary makes it possible to identify and visualize this trend using quantitative data.

In Hungarian, inflections convey the morphosyntactic features of words. The same morphosyntactic features are signified by the same type of inflections. This allows for a quantitative assessment of changes in the use of inflectional rhymes by identifying rhyming pairs in which both words share the same part of speech and morphosyntactic features. In this investigation, those rhyming pairs were queried in which both rhyming words are nouns, adjectives, or verbs sharing the same morphosyntactic features. Cases where the rhyming words appear in their basic, uninflected forms were excluded. This means that the following were not considered: singular nouns in the nominative case, singular adjectives in the nominative case and positive degree, and singular third-person indicative indefinite verbs in the present tense.

By plotting the relative frequencies of the queried inflectional rhyming pairs for each author over time, the bar chart shown in Figure 3 was obtained. Using gray and black bar colors, the relative frequencies of inflectional rhymes are differentiated for authors born before 1770, between 1770 and 1860, and between 1860 and 1910. The chart reveals a decline in the frequency of inflectional rhymes during the 19th century, and the frequency remains low throughout the 20th century. In other words, the chart indicates that inflectional rhymes were widely used in Hungarian poetry before the 19th century, but in the 19th century, they gradually became stigmatized and fell out of favor.

Figure 3: Changes in the relative frequencies of inflectional rhymes over time.

6.1 Attraction between Trochaic and Qualitative Meters

To illustrate the usefulness of the Poem Form Searcher, this section presents a brief case study on the interaction between trochaic and qualitative meters. In Hungarian literary studies, there is a prevailing assumption that qualitative and trochaic meters are inherently drawn to each other (e.g. Vargyas 1966, 148–150; J. Horváth 1969, 128–133; Szepes and Szerdahelyi 1981, 267–269). In other words, it is believed that poems with trochaic meter are likely to also feature qualitative meter. This attraction between the two meters is explained by the phonological characteristics of the Hungarian language. However, to date, there has been no quantitative research testing the hypothesis that the two meters attract each other. By retrieving data using the Poem Form Searcher, this assumption can now be quantitatively examined.

To answer this question, mutual information scores were calculated for poems with iambic and qualitative meters and poems with trochaic and qualitative meters, respectively. In linguistics, association measures, such as mutual information, are usually used to extract collocations (Church and Hanks 1989). However, these measures can also be used for textual co-occurrence to measure whether two textual features co-occur in the same texts more often than would be expected by chance (see Evert 2009). Table 8 shows the mutual information scores for the pairings of iambic and qualitative meters as well as trochaic and qualitative meters. If the score is greater than 0, the two meters occur together more often than would be expected by chance, which means that they attract each other. If the value is less than 0, the two meters occur together less often than would be expected by chance. In other words, they repel each other. The MI scores were calculated for the entire corpus as well as for each author individually. In the latter case, the mean value was computed. The table also shows the number of poets with an MI score greater than 0 and less than 0. In the calculations by author, only those authors were included who had written at least one poem that is both iambic and qualitative, or trochaic and qualitative.

As Table 8 indicates, while iambic and qualitative meters occur together less often than expected by chance (the values are below 0), trochaic and qualitative meters occur together more often than expected by chance (the values are above 0). The number of authors with positive and negative MI scores differs considerably between the two cases. In the case of qualitative-trochaic meter, most authors have a positive MI score, while for qualitative-iambic meter, the majority of authors have a negative MI score. This means that the observation of literary scholars that trochaic and qualitative meters attract each other can be quantitatively validated in the context of canonical Hungarian poetry. However, the corpus does not allow for statistical inferences about Hungarian poetry in general. The results also clearly show that in canonical Hungarian poetry, iambic and trochaic meters tend to repel each other.