Verse, Variation, and Meaning: Annotating Similarity in Greek Marginal Texts

This paper introduces a manually annotated dataset for verse-level semantic textual similarity in Byzantine Greek book epigrams. The benchmark captures graded similarity judgements as it was annotated by making use of pairwise comparison. It is furthermore designed to evaluate computational models of meaning in historical languages. Alongside the benchmark, we present a series of experiments that compare static word embeddings, contextual language models, and a semantically informed edit distance approach. Our results show that lemmatised CBOW embeddings perform slightly better than more complex transformer-based models, and that edit distance methods can be effectively adapted to capture semantic proximity. The study also highlights key challenges, including annotation asymmetry and the interpretative limits of partial data. By combining philological expertise with computational experimentation, this work lays the foundation for scalable similarity detection in Ancient Greek and contributes a reusable resource to both digital philology and ancient language processing.