Elamur Research⌕
☰
home / papers / 2022 / Joint Coreference Resolution for Zeros and non-Zeros in Arabic

Joint Coreference Resolution for Zeros and non-Zeros in Arabic

Summary

In pro-drop languages such as Arabic, a subject can be left unspoken. These gaps, called anaphoric zero pronouns (AZPs), still refer back to earlier entities. Standard Arabic coreference systems cluster only overt mentions, and AZP resolvers link a zero to a single antecedent while ignoring the rest of the chain. The authors handle both mention types together. First, they rebuilt the Arabic part of the CoNLL-2012 dataset by pulling AZPs out of the OntoNotes Normal Forms (ONF) and merging them into the CoNLL files, with all annotation layers. They then compare two systems. The pipeline runs a coreference model and an AZP detector, then attaches each detected zero to an output cluster. The joint model marks detected zeros in the input and learns full chains that include them. On joint clustering, the joint model scores 67.1% F1 and beats the end-to-end system of Chen et al. (2021). On AZP-only resolution, Chen et al. stays marginally ahead. The extended dataset is the first Arabic resource that supports studying zeros and overt mentions together.

Key points

  • The extended dataset keeps the original CoNLL-2012 split:
    • train: 359 documents, 264,589 words, 3,495 AZPs;
    • dev: 44 documents, 474 AZPs;
    • test: 44 documents, 412 AZPs.
  • Some ONF chains contain one overt mention plus one AZP. CoNLL-2012 had dropped these because removing the zero left a singleton. The authors' script restores them as new clusters, while AZPs in larger chains take the existing chain number.
  • The pipeline is built from the Arabic coreference system of Aloraini et al. (2020) and the AZP identifier of Aloraini and Poesio (2020a).
  • Joint resolution of AZPs and non-AZPs (average F1):
    • joint learning: 67.1;
    • Chen et al. (2021): 64.2;
    • pipeline: about 0.7 points below Chen et al.
  • AZP resolution F1:
    • pipeline, first mention as cluster representative: 58.08;
    • pipeline, last mention as cluster representative: 58.59;
    • joint learning: 59.33;
    • Chen et al. (2021): 59.49.

Interesting ideas

  • I1. When a convenient corpus format drops empty elements, recover them from a richer source format (here ONF) and keep the chains that would otherwise collapse into singletons. — For Mova (tree): During treebank and UD dialect conversion, keep empty nodes and implicit arguments from source annotations instead of discarding them, so downstream coreference can train on them.
  • I2. Detect zeros first, then insert them as explicit tagged tokens, so the coreference model treats them like ordinary mentions with full document context. — For Mova (frames): Implicit subjects of English controlled infinitives, imperatives and ellipsis could become placeholder mentions in prag, letting the coreference step resolve them in the same pass as overt pronouns.
  • I3. Limit candidate zero positions with a linguistic constraint (only gaps after verbs) instead of treating every gap as a possible zero, which shrinks the search space and helps detection. — For Mova (frames): Posit an implicit argument only where the UD tree shows a verb lacking its expected core argument (e.g. an xcomp without nsubj), which keeps the inference explainable and low-noise.
  • I4. Resolve a zero to a whole entity cluster rather than to one antecedent mention, and test cheap choices of cluster representative (first vs. last mention). — For Mova (world): Implicit arguments should attach directly to world entities. Comparing which mention best represents an entity is an easy ablation for the entity store.
  • I5. Score the combined task and the zero-only subtask separately, because a system can win on one and lose on the other. — For Mova (learn): Evaluations should report implicit-argument resolution apart from overall coreference, so gains on overt mentions do not hide regressions on null ones.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.wanlp-1.2 · arXiv 2210.12169

@inproceedings{aloraini2022joint,
  title = {{Joint Coreference Resolution for Zeros and non-Zeros in Arabic}},
  author = {Aloraini, Abdulrahman and Pradhan, Sameer and Poesio, Massimo},
  booktitle = {Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP)},
  year = {2022},
  doi = {10.18653/v1/2022.wanlp-1.2},
  eprint = {2210.12169},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2022.wanlp-1.2}
}

References

28 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ryu Iida and Massimo Poesio. 2011. A cross-lingual ilp solution to zero anaphora resolution. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pages 804–813. find our card
  2. Abdulrahman Aloraini and Massimo Poesio. 2020a. Anaphoric zero pronoun identification: A multilingual approach. In Proceedings of the Third Workshop on Computational Models of Reference, Anaphora and Coreference, pages 22–32. find our card
  3. Ryu Iida, Kentaro Torisawa, Chikara Hashimoto, Jong-Hoon Oh, and Julien Kloetzer. 2015. Intra-sentential zero anaphora resolution using subject sharing recognition. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2179–2189. find our card
  4. Abdulrahman Aloraini and Massimo Poesio. 2020b. Cross-lingual zero pronoun resolution. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 90–98. find our card
  5. Abdulrahman Aloraini, Juntao Yu, and Massimo Poesio. 2020. Neural coreference resolution for arabic. arXiv preprint arXiv:2011.00286. ↗ our card
  6. Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. 2019. What does bert learn about the structure of language? In 57th Annual Meeting of the Association for Computational Linguistics (ACL), Florence, Italy. find our card
  7. Olga Uryupina, Alessandro Moschitti, and Massimo Poesio. 2012. Bart goes multilingual: the unitn/essex submission to the conll-2012 shared task. In Joint Conference on EMNLP and CoNLL-Shared Task, pages 122–128. Association for Computational Linguistics. find
  8. Megumi Kameyama. 1985. Zero Anaphora: The case of Japanese. Ph.D. thesis, Stanford University, Stanford, CA. find
  9. Fang Kong and Hwee Tou Ng. 2013. Exploiting zero pronouns to improve chinese coreference resolution. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 278–288. find our card
  10. Lydia White. 1985. The “pro-drop” parameter in adult second language acquisition. Language learning, 35(1):47–61. find
  11. Jingxuan Yang, Si Li, Sheng Gao, and Jun Guo. 2022. Corefdpr: A joint model for coreference resolution and dropped pronoun recovery in chinese conversations. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:571–581. find
  12. Ryuto Konno, Shun Kiyono, Yuichiroh Matsubayashi, Hiroki Ouchi, and Kentaro Inui. 2021. Pseudo zero pronoun resolution improves zero anaphora resolution. arXiv preprint arXiv:2104.07425. ↗ our card
  13. Qingyu Yin, Yu Zhang, Weinan Zhang, and Ting Liu. 2016. A deep neural network for chinese zero pronoun resolution. In arXiv preprint arXiv:1604.05800. ↗
  14. Kenton Lee, Luheng He, Mike Lewis, and Luke Zettlemoyer. 2017. End-to-end neural coreference resolution. arXiv preprint arXiv:1707.07045. ↗ our card
  15. Qingyu Yin, Yu Zhang, Weinan Zhang, and Ting Liu. 2017. Chinese zero pronoun resolution with deep memory network. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1309–1318. find our card
  16. Baoli li. 2012. Learning to model multilingual unrestricted coreference in ontonotes. In Joint Conference on EMNLP and CoNLL2012-Shared Task. find
  17. Charles N. Li and Sandra A. Thompson. 1979. Third person pronouns and zero anaphora in chinese discourse. In Syntax and Semantics, volume 12: Discourse and Syntax, pages 311–335. Academic Press. find
  18. Qingyu Yin, Yu Zhang, Weinan Zhang, Ting Liu, and William Yang Wang. 2018. Zero pronoun resolution with attention-based neural network. In Proceedings of the 27th International Conference on Computational Linguistics, pages 13–23. find our card
  19. Ting Liu, Yiming Cui, Qingyu Yin, Weinan Zhang, Shijin Wang, and Guoping Hu. 2017. Generating and exploiting large-scale pseudo training data for zero pronoun resolution. In arXiv preprint arXiv:1606.01603. ↗ our card
  20. Kim Young-Joo. 2000. Subject/object drop in the acquisition of korean: A cross-linguistic comparison. In Journal of East Asian Linguistics 9.4, pages 325–351. find
  21. Bonan Min. 2021. Exploring pre-trained transformers and bilingual transfer learning for arabic coreference resolution. In Proceedings of the Fourth Workshop on Computational Models of Reference, Anaphora and Coreference, pages 94–99. find our card
  22. Shanheng Zhao and Hwee Tou Ng. 2007. Identification and resolution of chinese zero pronouns: A machine learning approach. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 541–550. find our card
  23. 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. Association for Computational Linguistics, Association for Computational Linguistics., pages 1–40. find our card
  24. Desislava Zhekova and Sandra Kübler. 2010. Ubiu: A language-independent system for coreference resolution. In Proceedings of the 5th International Workshop on Semantic Evaluation, page 96–99. find our card
  25. Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. 2020. Pre-trained models for natural language processing: A survey. Science China Technological Sciences, 63(10):1872–1897. find our card
  26. Tomohide Shibata and Sadao Kurohashi. 2018. Entity-centric joint modeling of japanese coreference resolution and predicate argument structure analysis. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 579–589. find our card
  27. Marcus Stamborg, Dennis Medved, Peter Exner, and Pierre Nugues. 2012. Using syntactic dependencies to solve coreferences. In Joint Conference on EMNLP and CoNLL2012-Shared Task. find our card
  28. A CoNLL-2012 Annotation Layers The CoNLL-2012 annotation layers consists of the following (Pradhan et al., 2012): find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/aloraini-2022-joint-coreference-resolution-zeros-non/