Detecting Semantic Reuse in Ancient Greek Literature: AComputational Approach.

This paper introduces the first step towards a computational method for detecting semantic textual reuse in Ancient Greek literature. While existing tools focus primarily on exact or nearlexical matching, our approach leverages the semantic capabilities of contextual LLMs, aiming to finetune a pretrained encoder via contrastive learning to recognize textual reuse even when expressions are paraphrased and/or morphologically altered. To build a suitable dataset, we developed an automatic pipeline that generates positive samples by extracting paraphrases for each sentence using the Ancient Greek Wordnet and a customtrained morphological re-inflection model. Negative samples, or “confounders”, are selected through topic modeling to ensure thematic relevance while preserving semantic dissimilarity. The model is evaluated through a curated case study on Homeric formulae. We retrieve the top ten most similar sentences in a corpus of Ancient Greek authors from the classical age, assessing model outputs using both standard metrics and comparison with established philological studies. The outcomes demonstrate that contrastive fine-tuning, paired with linguistically informed data augmentation, offers promising directions for identifying non-literal textual reuse in historical corpora. This work contributes a framework for philological discovery, combining deep learning with interpretive scholarship in classical studies.