CorefUD 1.0: Coreference Meets Universal Dependencies

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4859–4872 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4859 CorefUD 1.0: Coreference Meets Universal Dependencies Anna Nedoluzhko1, Michal Nov´ak1, Martin Popel1, Zdenˇek ˇZabokrtsk´y1, Amir Zeldes2, Daniel Zeman1 1 Charles University, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, Prague, Czech Republic {nedoluzhko,mnovak,popel,zabokrtsky,zeman}@ufal.mff.cuni.cz 2 Georgetown University, Department of Linguistics, 1421 37th St. NW, Washington, DC, 20057, USA amir.zeldes@georgetown.edu Abstract Recent advances in standardization for annotated language resources have led to successful large scale efforts, such as the Universal Dependencies (UD) project for multilingual syntactically annotated data. By comparison, the important task of coreference resolution, which clusters multiple mentions of entities in a text, has yet to be standardized in terms of data formats or annotation guidelines. In this paper we present CorefUD, a multilingual collection of corpora and a standardized format for coreference resolution, compatible with morphosyntactic annotations in the UD framework and including facilities for related tasks such as named entity recognition, which forms a first step in the direction of convergence for coreference resolution across languages. Keywords: coreference, anaphora, corpus, treebank, multilinguality, coreference resolution 1. Introduction Recent years have seen a tremendous growth in the amount and quality of annotated resources available in comparable formats and using compatible guidelines for a variety of languages. Following the introduction of Google’s universal part-of-speech tags (Petrov et al., 2012), the Universal Dependencies (UD) project (de Marneffe et al., 2021) has made remarkable progress in standardizing both the inventory of morphosyntac- tic labels as well as the guidelines for assigning them across a wide range of datasets and languages, with the result that practitioners in a variety of fields relying on syntactic analysis can have a reasonable expectation of the format and nature of automatic and manual syntac- tic analyses based on the framework. The same cannot be said for coreference resolution, the task of clustering together multiple mentions of the same entity in a text (e.g. ‘Joe Biden’, ‘the U.S. Pres- ident’ and ‘he’), as well as other tasks concerned with anaphoric relations (e.g. event coreference resolution, bridging resolution, mention detection). As our survey below will show, even for a single, high-resource lan- guage such as English, datasets diverge broadly in the phenomena covered, the way they are analyzed and the formats and information made available by datasets. As a result, state-of-the-art methods, e.g. (Joshi et al., 2020, Kirstain et al., 2021, Dobrovolskii, 2021), are usually benchmarked only on one type of English data focusing on a limited scope of phenomena (most often OntoNotes (Weischedel et al., 2011)). Our motivation for changing the current situation is the following. First, testing the methods on other lan- guages is crucial, as the properties of how anaphoric re- lations are expressed may vary considerably (Kunz and Lapshinova-Koltunski, 2015, Nov´ak and Nedoluzhko, 2015). For instance, languages exhibit various extents of pronoun dropping (which is very rare in English), different rules of …

Content truncated to 500 words