Institute of Formal and Applied Linguistics

UDPipe 2 is a popular dependency parsing, morphological analysis and lemmatization tool. In this talk, we present our on-going work towards UDPipe 3, and we aim to share our insights from winning the EvaLatin 2024 Dependency Parsing shared task with our new architecture, dubbed LatinPipe for this purpose. We describe our experiences with multi-treebank/multilingual training in practical usage and other shared tasks, the importance of annotation style harmonization, how contextualization and position encoding matter for dependency parsing, and some other tricks to boost performance in competition. Besides what we hope to be an accessible, high-level explanation of the LatinPipe architecture, we will also describe the annotation process of harmonizing the annotation styles of the publicly available Latin corpora and investigate its impact on parsing results.

*** The talk will be delivered in person (MFF UK, Malostranské nám. 25, 4th floor, room S1) and will be streamed via Zoom. For details how to join the Zoom meeting, please write to sevcikova et ufal.mff.cuni.cz ***