--- title: "MSCAW-coref: Multilingual, Singleton and — Liu 2024" description: "Strong multilingual coreference systems usually enumerate candidate spans and then connect them, so their cost grows as O(n^4) in document length." authors: ["Houjun Liu", "John Bauer", "Karel D'Oosterlinck", "Christopher Potts", "Christopher D. Manning"] year: 2024 venue: "Proceedings of the Seventh Workshop on Computational Models of Reference, Anaphora and Coreference" source: "https://aclanthology.org/2024.crac-1.4/" doi: "10.18653/v1/2024.crac-1.4" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/liu-2024-mscaw-coref-multilingual-singleton-conjunction/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/liu-2024-mscaw-coref-multilingual-singleton-conjunction/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2024](https://md.research.elamur.ai/free/en/papers/2024/page/3/) / MSCAW-coref: Multilingual, Singleton and Conjunction-Aware Word-Level Coreference Resolution # MSCAW-coref: Multilingual, Singleton and Conjunction-Aware Word-Level Coreference Resolution **Authors:** Houjun Liu, [John Bauer](https://md.research.elamur.ai/free/en/authors/john-bauer/), [Karel D'Oosterlinck](https://md.research.elamur.ai/free/en/authors/karel-d-oosterlinck/), [Christopher Potts](https://md.research.elamur.ai/free/en/authors/christopher-potts/), [Christopher D. Manning](https://md.research.elamur.ai/free/en/authors/christopher-d-manning/) · **Year:** 2024 · **Venue:** Proceedings of the Seventh Workshop on Computational Models of Reference, Anaphora and Coreference · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2024.crac-1.4/) · [PDF](https://aclanthology.org/2024.crac-1.4.pdf) · [DOI 10.18653/v1/2024.crac-1.4](https://doi.org/10.18653/v1/2024.crac-1.4) · [OpenAlex](https://openalex.org/W4404783561) **Citations:** 1 citations (openalex-snapshot · 2026-10-14) · [7](#cited-by) in our library, of the 140673 papers whose bibliographies we read ## Summary Strong multilingual coreference systems usually enumerate candidate spans and then connect them, so their cost grows as O(n^4) in document length. Word-level methods such as CAW-coref reduce this to O(n^2) by linking head words first and recovering full spans afterwards. However, they were designed for English OntoNotes and cannot output singletons, which are entities mentioned only once. The authors show that porting CAW-coref unchanged to CorefUD, a multilingual corpus that annotates singletons, gives weak results. MSCAW-coref adds a per-word classifier that scores whether a word starts a new chain. This lets the chain-building step keep mentions that have no antecedent. Multilingual coverage comes from an XLM-RoBERTa-large encoder tuned with LoRA. On CorefUD 1.1, the system reaches roughly 96% of the top CRAC 2023 shared-task score under exact span matching with singletons. It stays quadratic and uses a smaller encoder than the leader. It also transfers zero-shot to languages withheld during training. Models and code are released in Stanza. ## Key points - Mean CoNLL F1 (MUC/B³/CEAFe, exact spans, singletons included) is 0.752. This compares with 0.785 for Straka (2023), the shared-task leader with a 1.2B-parameter backbone and O(n^4) complexity. - Naive CAW-coref with a multilingual backbone scores only 0.608 (about 77% of SOTA). Pražák and Konopik (2022), which uses the same 561M XLM-RoBERTa-large backbone, scores 0.656. - Singleton handling comes from a "first mention" score per word. This score is combined with coarse bilinear and fine feed-forward antecedent scores during greedy chain construction. - In leave-one-language-out ablations (LEA metric), Romance languages transfer well to each other. For example, holding out French costs it about 9.8%. Slavic and Uralic languages gained when languages from other families were removed. - Conjunction-aware head selection brought smaller gains in the multilingual setting than it did in English. *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](https://md.research.elamur.ai/free/en/shelves/linguistics/) Topics: [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) ## Cited by Cited by 7 papers of our library cite it (of the 140673 whose bibliographies we read); 3 of them have a page here, 4 not yet. - [CorefInst: Leveraging LLMs for Multilingual Coreference Resolution](https://md.research.elamur.ai/free/en/papers/arslan-2026-corefinst-leveraging-llms-multilingual-coreference/) - [GUM-SAGE: A Novel Dataset and Approach for Graded Entity Salience Prediction](https://md.research.elamur.ai/free/en/papers/lin-2025-gum-sage-novel-dataset-approach/) - [Findings of the Fourth Shared Task on Multilingual Coreference Resolution: Can LLMs Dethrone Traditional Approaches?](https://md.research.elamur.ai/free/en/papers/novak-2025-findings-fourth-shared-task-multilingual/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{liu2024mscaw, title = {{MSCAW-coref: Multilingual, Singleton and Conjunction-Aware Word-Level Coreference Resolution}}, author = {Liu, Houjun and Bauer, John and D'Oosterlinck, Karel and Potts, Christopher and Manning, Christopher D.}, booktitle = {Proceedings of the Seventh Workshop on Computational Models of Reference, Anaphora and Coreference}, year = {2024}, doi = {10.18653/v1/2024.crac-1.4}, url = {https://doi.org/10.18653/v1/2024.crac-1.4} } ``` - [← previous MsBERT: A New Model for the Reconstruction of Lacunae in Hebrew Manuscripts](https://md.research.elamur.ai/free/en/papers/aclbulk-2024-shmidman-msbert-new-model-reconstruction-lacunae/) - [next → MT-Bench-101: A Fine-Grained Benchmark for Evaluating Large Language Models in Multi-Turn Dialogues](https://md.research.elamur.ai/free/en/papers/bai-2024-mt-bench-101-fine-grained/) - [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/3/) ## References 2 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. 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](https://scholar.google.com/scholar?q=Bernd+Bohnet%2C+Chris+Alberti%2C+and+Michael+Collins%2E+2023%2E+Coreference+resolution+through+a+seq2seq+transition%2Dbased+system%2E+Transactions+of+the+Association+for+Computational+Linguistics%2C+11%3A212%E2%80%93226%2E) [our card](https://md.research.elamur.ai/free/en/papers/bohnet-2023-coreference-resolution-through-seq2seq-transition/) 2. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… [find](https://scholar.google.com/scholar?q=Alexis+Conneau%2C+Kartikay+Khandelwal%2C+Naman+Goyal%2C+Vishrav+Chaudhary%2C+Guillaume+Wenzek%2C+Francisco+Guzm%C3%A1n%2C+Edouard+Grave%2C+Myle+Ott%2C+Luke+Zettlemoyer%2C+and+Veselin+Stoyanov%2E+2020%2E+Unsupervised+cross%2Dlingu)