TransScan - Scanning Latin Dactylic Verse Using Transformers

The scansion of Latin verse involves labelling each syllable as long, short or elided in length, and is crucial to identifying the content and style of the poem. Since it is time-consuming to implement rule-based systems for several meters, we applied the popular Transformer architecture to the scansion problem. For this paper, we preprocessed 50,000 labelled verses in dactylic hexameter and elegiac couplets written from the 2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> century BC to the 2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> century AD, available from the Pede Certo project and used them to train four Transformer models of varying size for this sequence labelling task. In the paper, we present the resulting models and compare them to previous attempts at Latin verse scansion using deep learning. Our best model outperformed the control LSTM at identifying long and short syllables comprising most of the dataset (around 98% of all syllables), achieving F1-scores of 0.9937 and 0.9926 in each. However, it performed worse in identifying elided syllables, achieving an F1-score of only 0.9380.