CorefUD 1.0: Coreference Meets Universal Dependencies

Summary

Syntax has a shared multilingual annotation standard in Universal Dependencies, but coreference resolution has had no comparable standard: corpora define the phenomena differently and use different formats, so most systems are evaluated only on English OntoNotes. CorefUD is a collection of coreference corpora in several languages converted to one scheme that sits on top of UD morphosyntax (in the CoNLL-U MISC column) and also accommodates related layers such as named entities. Corpora were chosen for open licensing, size, language and annotation-scheme diversity and documentation, then converted, with an 8/1/1 train/dev/test split where none existed. The paper compares in detail which mention properties and link types each source corpus annotates. Zero mentions such as dropped pronouns are represented as UD empty nodes. The authors argue for tying coreference to syntax, since mentions usually coincide with constituents, and present the release as a step towards cross-lingual convergence in the spirit of Universal Anaphora.

How Mova uses it

coref — coref resolve writes its output in the CorefUD format: Entity= annotations in MISC with a # global.Entity = eid-etype-head-other header, where head is the position of the head inside the span and other records the sieve that attached the mention.
Mention spans follow CorefUD conventions (nominal dependents only, preposition outside the span, apposition as a separate mention), and the gold coreference of GUM in UD 2.18 is read in the same format for evaluation.

Effectiveness in Mova

Not measured separately: CorefUD is the data and output format, not an algorithm. It makes evaluation against GUM possible, where the deterministic resolver reaches CoNLL F1 60.31 on the GUM test (gold trees, head match, without singletons) and 73.04 with singletons.

👨‍🔬💥

Topics: Coreference resolution · Universal Dependencies

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{nedoluzhko2022corefud,
  title = {{CorefUD 1.0: Coreference Meets Universal Dependencies}},
  author = {Nedoluzhko, Anna and Novák, Michal and Popel, Martin and Žabokrtský, Zdeněk and Zeldes, Amir and Zeman, Daniel},
  booktitle = {Proceedings of the Thirteenth Language Resources and Evaluation Conference},
  year = {2022},
  url = {https://aclanthology.org/2022.lrec-1.520/}
}

References

12 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

(Recasens and Martı́, 2010) CC BY-NC-SA 3.0 find
(Nedoluzhko et al., 2016) CC BY-NC-SA 4.0 find
(Hajič et al., 2020) mixture of CC licenses (none contains ND) CC BY-NC 4.0 (if TED section is omitted) CC BY-SA 4.0 CC BY-NC 4.0 (if TED section is omitted) find
(Lapshinova-Koltunski et al., 2018) find
(Landragin, 2021) (Lapshinova-Koltunski et al., 2018) (Bourgonje and Stede, 2020) CC BY-NC-SA CC BY 4.0 find
(Vincze et al., 2018) CLARIN-LT End User License CC BY 3.0 CC BY-SA 4.0 find
(Žitkus and Butkienė, 2018) find
(Ogrodniczuk et al., 2013) find
(Toldova et al., 2014) CC BY 4.0 find
(Recasens and Martı́, 2010) a proprietary license (Hendrickx et al., 2008) a mixture of proprietary licenses LDC (Uryupina et al., 2020) find
(Weischedel et al., 2011) LDC (Nedoluzhko et al., 2016) tion, discourse deixis, split antecedents and bridging. find
Polish Coreference Corpus (Ogrodniczuk et al., 2013, Ogrodniczuk et al., 2015) is a corpus (∼ 540K tokens) of Polish nominal coreference built upon the National Corpus of Polish. Mentions are annotated as linear spans, with additionally marked semantic heads. The annotation includes identity coreference, quasi-identity relations and non-identity close-to-coreference relations. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/nedoluzhko-2022-corefud-1-0-coreference-meets/