Transformers analyzing poetry: multilingual metrical pattern prediction with transfomer-based language models

Abstract The splitting of words into stressed and unstressed syllables is the foundation for scansion poetry, a process that aims at determining metrical pattern line verse within poem. Intricate language rules their exceptions, as well poetic licenses exerted by authors, make calculating these patterns nontrivial task. Some rhetorical devices shrink length, while others might extend it. This opens door interpretation further complicates creation automated algorithms useful automatically analyzing corpora on distant reading fashion. In this paper, we compare identification systems available Spanish, English, German, against fine-tuned monolingual multilingual models trained same Despite being initially conceived suitable semantic tasks, our results suggest transformers-based retain enough structural information to perform reasonably Spanish setting, outperforms both English German when using model three languages, showing evidence benefits cross-lingual transfer between languages.