Bertsification: Language modeling fine-tuning for Spanish scansion

Since ancient times, cultures have maintained oral traditions by sharing stories in poetic form (Francis 2017). Although different traditions treat poetry according to their own thinking frameworks, most of them perform some sort of analysis that requires the extraction of stress patterns of lines or verses. Scansion, as this procedure is broadly referred to, is usually language dependent. On the assumption that modern context-dependent language models are able to retain the structural properties of text, this paper explores whether metrical patterns are part of the information encoded in the embedding spaces of text