Monday, 18 November, 2019 - 14:00
Room: 

What's new with UDPipe: recent successes, growing pains and future plans

Jana Straková & Milan Straka (ÚFAL MFF UK)

UDPipe is an UFAL in-house natural language processing pipeline which performs segmentation, tokenization, POS tagging, lemmatization and parsing. We will present our evaluation of UDPipe with contextual embeddings on 54 languages of the UD treebank and our recent participation in two shared tasks. The SIGMORPHON 2019 Task 2, called "Morphological analysis and lemmatization in context", featured 66 languages and UDPipe was one of the two winning systems. The second one, the CoNLL2019 shared task, was called MRP 2019 - Cross-Framework Meaning Representation Parsing. The task involved modeling of five different semantic formalisms and our semantic extension to UDPipe placed third. We will also briefly touch our concurrent (nested) NER work which we presented at ACL 2019 and which we plan to include in UDPipe. Finally, we will describe our release plans.