THIVLVC
“THIVLVC: Retrieval Augmented Dependency Parsing for Latin”
LT4HALA 2026 / EvaLatin 2026, LREC 2026, Palma de Mallorca, 11 May 2026, pp. 219–225.
In 2026, I published this work with Luc Pommeret and Thibault Wagret. We asked whether previously annotated Latin sentences could help a language model correct an initial syntactic analysis. This is the principle behind THIVLVC.
Results and resources
In the EvaLatin 2026 evaluation, THIVLVC_2 achieved the best results on the poetry corpus and the best LAS without subtypes on the prose corpus. Compared with UDPipe, its CLAS score with subtypes improved by 17 points on poetry and 1.5 points on prose.
We also conducted a double-blind review of 300 disagreements with the reference corpus. Of the 167 unanimous decisions, 53.3% favoured THIVLVC.