--- title: "End-to-end Multilingual Coreference Resolution — Pražák 2022" description: "The paper presents the University of West Bohemia system submitted to the CRAC 2022 shared task, which addressed coreference resolution across multiple…" authors: ["Ondřej Pražák", "Miloslav Konopik"] year: 2022 venue: "Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution" source: "https://aclanthology.org/2022.crac-mcr.3/" 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/papers/prazak-2022-end-end-multilingual-coreference-resolution/" --- [Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/prazak-2022-end-end-multilingual-coreference-resolution/) [home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2022](https://md.research.elamur.ai/free/papers/2022/) / End-to-end Multilingual Coreference Resolution with Mention Head Prediction # End-to-end Multilingual Coreference Resolution with Mention Head Prediction **Authors:** [Ondřej Pražák](https://md.research.elamur.ai/free/authors/ondrej-prazak/), [Miloslav Konopik](https://md.research.elamur.ai/free/authors/miloslav-konopik/) · **Year:** 2022 · **Venue:** Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2022.crac-mcr.3/) · [PDF](https://aclanthology.org/2022.crac-mcr.3.pdf) · [OpenAlex](https://openalex.org/W4297433895) **Citations:** 4 citations (openalex-snapshot · 2026-10-14) · [11](#cited-by) in our library, of the 140673 papers whose bibliographies we read ## Summary The paper presents the University of West Bohemia system submitted to the CRAC 2022 shared task, which addressed coreference resolution across multiple languages. The task uses CorefUD, which contains 13 coreference datasets in ten languages, all built on UD treebanks. The authors start from the official transformer baseline, which follows the end-to-end span-ranking approach of Lee et al. (2017). They replace mBERT with XLM-R large and also test the best monolingual model for each language. Because XLM-R adds about 50M randomly initialized parameters, they first train one model on all datasets combined and then fine-tune it per dataset. This greatly helps small corpora. The official scorer counts a mention as correct if it contains the gold head. To exploit this, they add a classifier that predicts the head word(s) of each span while clustering still uses full spans. An attempt to add dependency paths to ROOT lowered scores, probably because of an implementation bug. The system placed 3rd overall and was best on two of the 13 datasets. The results show that joint multilingual pretraining is valuable for low-resource coreference. ## Key points - **Score and ranking:** The system averaged 67.64 CoNLL F1, compared with 58.53 for the baseline and 70.72 for the winning system (straka). It achieved the top result on German-ParCorFull (73.9) and Hungarian-SzegedKoref (65.01). - **Effect of joint pretraining:** Joint pretraining (80k steps on all data, then about 30k fine-tuning steps per dataset) raised the average from 62.03 (XLM-R per dataset) to 66.21. Adding dev data raised it further to 66.94. German-ParCorFull improved from 33.75 to 68.91. - **Head prediction:** Two variants were tested, one predicting head positions with sigmoid outputs and one classifying span–head pairs. Both performed about equally. They gave clear gains on three datasets but did not raise the average. - **Monolingual models:** Polish (HerBERT large) was the only case where a monolingual model outperformed the joint model. - **Training setup:** Training ran on NVIDIA A40 GPUs with one document per batch and up to 6 × 512-token segments. Each run took 8–20 hours. *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/shelves/linguistics/) Topics: [Coreference resolution](https://md.research.elamur.ai/free/topics/coreference-resolution/) ## Cited by Cited by 11 papers of our library cite it (of the 140673 whose bibliographies we read); 4 of them have a page here, 7 not yet. - [Findings of the Second Shared Task on Multilingual Coreference Resolution](https://md.research.elamur.ai/free/papers/abokrtsky-2023-findings-second-shared-task-multilingual/) - [Investigating Multilingual Coreference Resolution by Universal Annotations](https://md.research.elamur.ai/free/papers/chai-2023-investigating-multilingual-coreference-resolution-universal/) - [ÚFAL CorPipe at CRAC 2022: Effectivity of Multilingual Models for Coreference Resolution](https://md.research.elamur.ai/free/papers/straka-2022-fal-corpipe-crac-2022-effectivity/) - [ÚFAL CorPipe at CRAC 2023: Larger Context Improves Multilingual Coreference Resolution](https://md.research.elamur.ai/free/papers/straka-2023-fal-corpipe-crac-2023-larger/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{prazak2022end, title = {{End-to-end Multilingual Coreference Resolution with Mention Head Prediction}}, author = {Pražák, Ondřej and Konopik, Miloslav}, booktitle = {Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution}, year = {2022}, url = {https://aclanthology.org/2022.crac-mcr.3/} } ``` - [← previous Encoding and Fusing Semantic Connection and Linguistic Evidence for Implicit Discourse Relation Recognition](https://md.research.elamur.ai/free/papers/xiang-2022-encoding-fusing-semantic-connection-linguistic/) - [next → Evidentiality-guided Generation for Knowledge-Intensive NLP Tasks](https://md.research.elamur.ai/free/papers/asai-2022-evidentiality-guided-generation-knowledge-intensive/) - [↑ back to papers of 2022](https://md.research.elamur.ai/free/papers/2022/) ## References 2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. ca_ancora cs_pcedt cs_pdt de_parcorfull de_potsdamcc en_gum en_parcorfull es_ancora fr_democrat hu_szegedkoref lt_lcc pl_pcc ru_rucor avg PlanTL-GOB-ES/roberta-base-ca Czert-B-base-cased Czert-B-base-cased deepset/gbert-base deepset/gbert-large roberta-large roberta-large PlanTL-GOB-ES/roberta-large… [find](https://scholar.google.com/scholar?q=ca%5Fancora+cs%5Fpcedt+cs%5Fpdt+de%5Fparcorfull+de%5Fpotsdamcc+en%5Fgum+en%5Fparcorfull+es%5Fancora+fr%5Fdemocrat+hu%5Fszegedkoref+lt%5Flcc+pl%5Fpcc+ru%5Frucor+avg+PlanTL%2DGOB%2DES%2Froberta%2Dbase%2Dca+Czert%2DB%2Dbase%2Dcased+Czert%2DB%2Dbase%2D) 2. BASELINE Monoling XLM-R joined +dev +S2H +Tree 63.74 70 67.27 33.75 55.44 62.59 36.44 65.98 55.55 52.35 64.81 65.34 67.66 58.53 69.61 73.74 69.81 43.04 58.81 68 25.84 60.12 56.76 59.76 66.93 75.2 69.33 61.30 66.19 73.55 70.99 33.75 59.03 66.27 36.44 67.99 55.94 60.68 64.81 73.19 77.5 62.03 68.81 73.… [find](https://scholar.google.com/scholar?q=BASELINE+Monoling+XLM%2DR+joined+%2Bdev+%2BS2H+%2BTree+63%2E74+70+67%2E27+33%2E75+55%2E44+62%2E59+36%2E44+65%2E98+55%2E55+52%2E35+64%2E81+65%2E34+67%2E66+58%2E53+69%2E61+73%2E74+69%2E81+43%2E04+58%2E81+68+25%2E84+60%2E12+56%2E76+59%2E76+66%2E93+75%2E2+69%2E3)