Elamur Research⌕
☰
home / papers / 2023 / CAW-coref: Conjunction-Aware Word-level Coreference Resolution

CAW-coref: Conjunction-Aware Word-level Coreference Resolution

Summary

Coreference resolution groups text mentions that point to the same entity. The strongest current systems call large language models many times per document, which is too costly for large-scale information extraction. Word-level coreference (WL-coref) is much cheaper. A single pass of roberta-large produces word embeddings, a light scorer links coreferent words, and full spans are then recovered around those words. The authors trace a recurring weakness to how WL-coref turns span-annotated OntoNotes into word-level training data. Each span gets one head-word from the dependency parse, so a coordinated mention like "Tom and Mary" shares its head with "Tom". This causes conflicting links, and one of them must be discarded. It also makes span extraction for nested mentions ambiguous. Their fix is to treat the coordinating conjunction as the head-word of such spans, detected using POS tags and dependency depth. Retraining with unchanged hyperparameters yields CAW-coref. It raises OntoNotes test F1 without adding model complexity and becomes the strongest efficient coreference model reported.

Key points

  • On the OntoNotes 5.0 English test set, CAW-coref reaches 81.6% average F1, compared with 80.7% for the rerun WL-coref baseline (+0.9). This closes 34.6% of the gap to the SOTA link-append system, which scores 83.3% using a 13B mT5-XXL model.
  • Both WL-coref and CAW-coref need just a single forward pass through roberta-large (355M) and have O(n²) complexity. Autoregressive methods such as link-append, ASP and corefqa need at least O(n) LM passes per document.
  • The new head-word rule works as follows. A coordinating conjunction (and, or, plus) found via OntoNotes POS tags becomes the head-word if it lies less than two dependency steps from the span's original head. About 1.17% of spans in the train and dev splits are conjoined entities.
  • CAW-coref also beats LingMess (81.4%), the best span-based model, which relies on ensembling. The code is available at github.com/KarelDO/wl-coref.
  • Some limitations remain. Chained conjunctions like "Tom and Mary and David" have more mentions than words, so they cannot be fully handled. Conjunctions expressed with commas or hyphens are also not detected.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics: Coreference resolution · Part-of-speech tagging

Cited by

10 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 8 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.crac-main.2 · arXiv 2310.06165

@inproceedings{doosterlinck2023caw,
  title = {{CAW-coref: Conjunction-Aware Word-level Coreference Resolution}},
  author = {D'Oosterlinck, Karel and Bitew, Semere Kiros and Papineau, Brandon and Potts, Christopher and Demeester, Thomas and Develder, Chris},
  booktitle = {Proceedings of the Sixth Workshop on Computational Models of Reference, Anaphora and Coreference (CRAC 2023)},
  year = {2023},
  doi = {10.18653/v1/2023.crac-main.2},
  eprint = {2310.06165},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2023.crac-main.2}
}

References

15 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. find our card
  2. Shon Otmazgin, Arie Cattan, and Yoav Goldberg. 2022. F-coref: Fast, accurate and easy to use coreference resolution. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Proce… find our card
  3. Bernd Bohnet, Chris Alberti, and Michael Collins. 2023. Coreference resolution through a seq2seq transition-based system. Transactions of the Association for Computational Linguistics, 11:212–226. find our card
  4. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416. find
  5. Shon Otmazgin, Arie Cattan, and Yoav Goldberg. 2023. Lingmess: Linguistically informed multi expert scorers for coreference resolution. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2744– 2752. find our card
  6. Vladimir Dobrovolskii. 2021. Word-level coreference resolution. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7670–7675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. find our card
  7. Ian Porada, Alexandra Olteanu, Kaheer Suleman, Adam Trischler, and Jackie Chi Kit Cheung. 2023. Investigating failures to generalize for coreference resolution models. arXiv preprint arXiv:2303.09092. find our card
  8. Sameer Pradhan, Xiaoqiang Luo, Marta Recasens, Eduard Hovy, Vincent Ng, and Michael Strube. 2014. Scoring coreference partitions of predicted mentions: A reference implementation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pag… find our card
  9. Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Olga Uryupina, and Yuchen Zhang. 2012. CoNLL-2012 shared task: Modeling multilingual unrestricted coreference in OntoNotes. In Joint Conference on EMNLP and CoNLL - Shared Task, pages 1–40, Jeju Island, Korea. Association for Computational Linguisti… find our card
  10. Ljiljana Progovac. 1998. Structure for coordination. Glot international, 3(7):3–6. find
  11. Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A model-theoretic coreference scoring scheme. In Sixth Message Understanding Conference (MUC-6): Proceedings of a Conference Held in Columbia, Maryland, November 6-8, 1995. find our card
  12. Wei Wu, Fei Wang, Arianna Yuan, Fei Wu, and Jiwei Li. 2020. Corefqa: Coreference resolution as query-based span prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6953–6963. find our card
  13. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Ling… find
  14. Cyril Edward Zoerner III. 1995. Coordination: The syntax of &P. University of California, Irvine. find
  15. Model Step Prediction Correct WL-coref WL-coref CAW-coref CAW-coref word span word span Tom and Anna are talking. They are talking. Tom and Anna are talking. They are talking. Tom and Anna are talking. They are talking. Tom and Anna are talking. They are talking. Yes No Yes Yes WL-coref WL-coref CAW… find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/oosterlinck-2023-caw-coref-conjunction-aware-word/