Speech Rhythm and Syntax in Poetry and Prose

Abstract.We discuss several experiments regarding the interaction of syntax with meter and rhythm in poetry and prose. First, we determine the probability of rhythmic contours (syllable prominence) for part of speech tags (POS tags), which allows us to establish a stress hierarchy (nouns are usually stressed, while articles are seldom stressed). Second, we determine the context dependence of rhythmic contours, by investigating when and how stress ambiguous words (like adverbs) change their stress based on their context, e.g, if a pronoun or a noun is next to it. Third, we classify poetry against prose, both on a syntactic and a rhythmic level. We do this with a linear discriminant analysis so we can interpret feature weights of POS sequences and rhythmic groups. Lastly, we look at the interaction of a simple form of enjambement with POS tags and verse measures.1 IntroductionThe verbal rendering of thought requires the choice of appropriate lexical items and their orderingaccording to the rules of syntax. Syntax, however, does not fully determine word order:speakers and writers can often choose among possible syntactic constructions when formulatingtheir message. Semantic, pragmatic, as well as phonological constraints are known toaffect wording. In spontaneous language production, semantic constraints presumably controlsentence structure more immediately and to a stronger degree than phonological constraints.This follows from the logical directionality of language production, in which the semantic contentof the message governs lexical choice and the assignment of syntactic function; phonologyand rhythm can exert their role and endow the structure with sound only once a syntacticscaffold has been constructed (Levelt, 1993).Prosody thus affects the choice of syntactic constructions and the order of constituentswithin a sentence (Anttila, 2016). The influence of prosody on syntax is most obviously attestedin metered poetry where strict metrical rules and poetic license influence word orderand grammaticality (Donat, 2010). However, for German, it is largely unclear which prosodicfactors affect sentence construction, and how strong their influence is on grammatical encoding.To analyze the interaction of syntax with meter and rhythm at scale, we make use of theGerman poetry corpus in version 4 (Haider and Eger, 2019), containing the poetry of Textgrid(textgrid.de) and the German text archive (DTA: deutschestextarchiv.de). The corpus contains59k poems over 1.6M lines. It is available at github.com/tnhaider/DLK. See Table 1for a size overview.We train Conditional Random Fields (CRF) with the sklearn crf-suite to automaticallyannotate the whole German Poetry corpus for part-of-speech (POS), binarymeter (BM), and free speech rhythm with three levels of syllable prominence (TR). To thatend, two students of linguistics / literature manually annotated 3600 lines of school canonpoetry for binary meter (BM) and ternary rhythm (TR). For POS, we rely on gold annotationfrom DTA and the TIGER Corpus, according to the STTS tagset. We train and test acrossseveral genres to determine the most robust POS model for our purposes. See Table 2 for anoverview of POS models. We find that training on TIGER is not robust to tag across domains,falling to around .8 F1-score when testing against different genres from DTA. Training on thewhole DTA or on Belletristik (fiction/literature) is however sufficient to tag poetry or fiction.See example (1) for an annotated line of poetry with BM and TR. BMincludes binary syllable prominence (+/-) and foot boundaries ( ). TR segments the verseinto rhythmic groups at caesuras (:) and in these segments allows for main accents (2), sideaccents (1), and no accents (0). The example line illustrates a regular iambic pentameter, acaesura at the comma, and two rhythmic groups of the same length but with different form.(1) <l met="-+ -+ -+ -+ -+ " rhythm="01020:20102:">Ge-duck-te Hüt-ten, Pfa-de wirr ver-streut,</l>TR differs from BM such that TR operates top-down from rhythmic segments to findfree speech rhythm, while BM adheres more to a conventional metrical poetry analysis andstarts bottom-up from syllable prominence. Annotators largely followed an intuitive notionof rhythm, and incorporated philological knowledge to consider the (schema) consistency ofthe poem. They were also instructed to prefer longer feet over short ones (where applicable).Our inter-annotator agreement is substantial. Five poems were annotated by two annotators,and calculated on each syllable, Cohen kappa for metric syllables was at .95, and .84 forrhythmic syllables. For the latter, mainly side accents (1) were confused. Caesuras alone hada kappa of .92. Metric feet were more challenging and will not be discussed here, as there aremultiple reasons why foot boundaries can be ambiguous.The CRF models are trained on syllabified lines without punctuation. The features containthe syllable tokens, including two syllables to the left and the right, and also orthographicfeatures like capitalization and characters. During training, we also allow the algorithm to see the surrounding labels (which are not available during testing). We achieve 95% F1-score for POS and BM, while TR (including caesuras) still performs well with 83% F1-score. The confusion of the model issimilar to humans who also confuse main and side accents.2 Experiments2.1 Stress HierarchyTo determine the likelihood of a word belonging to a certain POS class being stressed orunstressed, we iterate over our set of 1.6 million lines of poetry, using the CRF models toannotate the Corpus for BM and POS. For our experiments, we simplify the tagset. We thencount how often a POS tag falls into a metrically stressed or unstressed syllable. Multisyllabicwords have lexically fixed stress patterns, e.g., German words with two syllables are usuallytrochaic, where the first syllable is stressed and the second syllable is unstressed. For wordswith three syllables, we found that nouns are more likely to follow the (+,-,+) pattern, whileverbs prefer (-,+,-). Here, we only measure the prominence of monosyllabic words, whichshow the most ambiguity, as the stress of monosyllabic words is mostly determined by theircontext.Anttila et al., (2018) also determined a stress hierarchy, only for sentential stress in politicalspeeches and not on poetic lines. They are able to establish a stress hierarchy of pos-tags,such that NOUN > ADJ > VERB > FUNC. This shows that functions words (FUNC, e.g.KONJ, ART, APPR, etc) are seldom stressed, while nouns are usually stressed.Based on our corpus, we determined the following hierarchy for monosyllabic word forms(see Table 3):NOUN > VERB_modal > VERB_full > ADJ > ADV > FUNC.This hierarchy reflects the ratio r of stressed to unstressed syllables, normalized to 1. Whena POS class is equally likely to be stressed or unstressed, the r will be 1.0, or 1:1. For a ratio r= 16.0 (16:1), the word class is 16 times more likely to be stressed. We found it striking thatmodal verbs are stressed so strongly (3.8:1). We also found that monosyllabic verbs are morelikely in metrically strong positions than monosyllabic adjectives, which deviates from whatAnttila et al., (2018) found. However, the ends of the hierarchy (nouns and function words)are the same. And they do not distinguish between adverbs and adjectives and also have onlyone verb class.2.2 Context Stress AmbiguityAs words are heavily dependent on their context regarding their stress, we look at the immediateleft and right context of POS tags, i.e. which POS tag occurs next to it. For brevity,we only show the left context. We retrieve the stress ratio for particular monosyllabic wordclasses dependent on their context. Context words can be multisyllabic.See Table 4 for anoverview of nouns, modal verbs, adjectives, adverbs and demonstrative pronouns.We can see that the hierarchy from Table 3 reiterates for contextual dependence. If a wordis preceded by a conjunction (KO), then the likelihood of stress is higher. However, nouns never lose their prominence (r > 1), regardless of context. Most interestingly, adverbs, which are quite balanced, also show a balanced context dependence, while modal verbs are still mostly stressed, except when they are preceded by another modal verb. We acknowledge that this table can be problematic, such that some of these contexts seem atypical for particular word classes. Future research should investigate the frequency of particular contexts, and how significant they are.2.3 Prose vs. Poetry ClassificationTo determine features that distinguish prose from poetic writing on a syntactic and rhythmiclevel, we perform classification with a regularized linear discriminant analysis that allows us tointerpret feature loadings. Regularization is necessary, as many features are collinear, makingthe feature loadings not interpretable (as collinear features will be important for both classes,without contributing to the classification).For features, we use POS n-grams and rhythmic groups. POS n-grams are straightforward,where only subsequent POS tags are considered. For rhythmic groups, we use the CRF trainedternary rhythm (TR) with caesuras, and split the sequences at caesuras. Consequently, a line’0201:020’ yields the features ’0201’ and ’020’.First, we tried to classify 100k sentences from poetry and prose (literature) respectively.We extracted the sentences from DTA. For POS n-grams, using unigrams or trigrams did notbeat the random 50% baseline. Training the classifier on bigrams achieves 67% F1. Classifyingsentences based on rhythmic groups performed at 56%, marginally better than the randombaseline.However, when we trained with POS bigrams on whole documents, F1 is around 93%,possibly because this classification relies on rare bigrams, such that sentence structure isoften the same in poetry and prose, but when it differs, it differs strongly.See Table 5 for an overview of feature weights from this latter classification. Higher standingfeatures are more important.2.4 Versemeter vs. EnjambementOur setup also allows us to get an impression of the interaction of enjambement with versemeasures and also POS transitions between lines. Enjambement is an integral part of manypoetic lines. It typically signifies incomplete syntax at the end of a line, such that the end ofthe line encourages a pause in speech, but the sentence, or clause, or phrase, or word is notyet finished.We use the simplest form to operationalize enjambement, by assigning enj+ to lines thatdo not end on a punctuation mark, and enj- to lines that do. Beyond obvious cases (ART_NNdoes not cross clause boundaries), we could not identify clear preferences of enjambement forparticular POS transistions.We implement a set of regular expressions to detect the breadth of verse measures based onthe syllable prediction of the meter CRF. Unsurprisingly, we find that lines with fewer stressed syllables prefer enjambement. However, for measures with six stressed syllables (notation: I:stressed, o:unstressed, ?:previous syllable optional, $:end of line), the runningmeasure hexameter (Ioo?Ioo?Ioo?Ioo?IooIo$) prefers the enjambement with a probability ofp(enj+) = .41, while the alexandrine (oIoIoIoIoIoIo?$) dislikes it with p(enj+) = .16.3 Conclusion & Future Work We have shown experiments on the intersection of syntax and speech rhythm, outlining stresshierarchies with and without context, questioning previous research. Also, we have shownthat a classification of documents on POS bigrams shows clear distinctive features of poetry vs.prose, while classifying sentences is challenging. In the end, we also discussed first explorations regarding enjambement and its interaction with syntax and different verse forms.