Are Large Language Models able to grasp Latin semantics? A study on preverbed motion verbs

Motion verbs offer a rich field for studying how languages express spatial and metaphorical meanings, and how these meanings evolve over time. In Latin, preverbed motion verbs are a particularly informative domain for investigating lexicalisation, as preverbs exhibit varying degrees of compositional transparency. This paper investigates whether Large Language Models (LLMs) can capture such semantic nuances and diachronic variation. We evaluate GPT-based models on a manually annotated Latin corpus spanning Early to Late Latin. We assess the models’ ability to identify preverbed verbs, separate preverb and base, and detect cases of lexicalisation. Results are promising and show that the models are able to identify the relevant instances in the majority of cases. Based on the results from the models, we analyse lexicalisation patterns in our corpus, finding evidence of known diachronic trends, with later Latin showing increased non-compositional usage, particularly for verbs with in- and per-. These results suggest that LLMs can support quantitative semantic research on Latin at scale, complementing traditional philological analysis.