CorPipe 25: CRAC 2025 Winning System for Multilingual Coreference Resolution

This repository will contain the source code and the pre-trained models of CorPipe 25, the CRAC 2025 winning system for multilingual coreference resolution. Both will be published before the CODI-CRAC 2025 workshop and are described in the following paper:

---

CorPipe at CRAC 2025: Evaluating Multilingual Encoders for Multilingual Coreference Resolution

Milan Straka
Charles University
Faculty of Mathematics and Physics
Institute of Formal and Applied Lingustics
Malostranské nám. 25, Prague, Czech Republic

Abstract: We present CorPipe 25, the winning entry to the CRAC 2025 Shared Task on Multilingual Coreference Resolution. This fourth iteration of the shared task introduces a new LLM track alongside the original unconstrained track, features reduced development and test sets to lower computational requirements, and includes additional datasets. CorPipe 25 represents a complete reimplementation of our previous systems, migrating from TensorFlow to PyTorch. Our system significantly outperforms all other submissions in both the LLM and unconstrained tracks by a substantial margin of 8 percentage points.

---

Content of this Repository

• The directory data is for the CRAC 2025 data, and the preprocessed and tokenized version needed for training.

  • The script data/get.sh downloads and extracts the CRAC 2025 training and development data, plus the unannotated test data of the CRAC 2025 shared task.

• The corpipe25.py is the complete CorPipe 25 source file.

• The corefud-score.sh is an evaluation script used by corpipe25.py, which

  • performs evaluation (using the official evaluation script from the corefud-scorer submodule),
  • optionally (when -v is passed), it also:
    • runs validation (using the official UD validator from the validator submodule) on the output data,
    • performs evaluation with singletons,
    • performs evaluation with exact match.

• The res.py is our script for visualizing performance of running and finished experiments, and for comparing two experiments. It was developed for our needs and we provide it as-is without documentation.

The Released Models

Three pretrained CorPipe 25 models have been released under the CC BY-NC-SA 4.0 license:

• corpipe25-corefud1.3-base-251101 also available on LINDAT/CLARIAH-CZ;
• corpipe25-corefud1.3-large-251101 also available on LINDAT/CLARIAH-CZ;
• corpipe25-corefud1.3-xl-251101 also available on LINDAT/CLARIAH-CZ.

See the respective model repositories for their performances and training hyperparameters.

Training a Multilingual mT5-large-based CorPipe 25 Model

To train a multilingual model on all the data using mT5 large, you should

1. run the data/get.sh script to download the CRAC 2025 data,

2. create a Python environments with the packages listed in requirements.txt,

3. train the model itself using the corpipe25.py script.

   For training a mT5-large variant with square-root mix ratios and without corpus ids, use

   tbs="ca_ancora cs_pcedt cs_pdt cu_proiel de_potsdamcc en_gum en_litbank es_ancora fr_ancor fr_democrat grc_proiel hbo_ptnk hi_hdtb hu_korkor hu_szegedkoref ko_ecmt lt_lcc no_bokmaalnarc no_nynorsknarc pl_pcc ru_rucor tr_itcc"
   
   python3 corpipe25.py --train --dev --treebanks $(for c in $tbs; do echo data/$c/$c-corefud-train.conllu; done) --batch_size=8 --learning_rate=6e-4 --learning_rate_decay  --adafactor --encoder=google/mt5-large --exp=corpipe25-corefud1.3-large --compile

Predicting with a CorPipe 25 Model

To predict with a trained model, use the following arguments:

corpipe25.py --load ufal/corpipe25-corefud1.3-large-251101 --exp target_directory --epoch 0 --test input1.conllu input2.conllu

• instead of a HuggingFace identifier, you can use directory name – if the given path name exists, the model is loaded from it;
• the outputs are generated in the target directory, with .00.conllu suffix;
• if you want to also evaluate the predicted files, you can use --dev option instead of --test;
• optionally, you can pass --segment 2560 to specify longer context size, which very likely produces better results, but needs more GPU memory.

Running the Model on Plain Text

To run the model on plain text, first the plain text needs to be tokenized and converted to CoNLL-U (and optionally parsed if you also want mention heads), by using for example UDPipe 2:

curl -F data="Eve came home and Peter greeted her there. Then Peter and Paul set out to a trip and Eve waved them off." \
  -F model=english -F tokenizer= -F tagger= -F parser=  https://lindat.mff.cuni.cz/services/udpipe/api/process \
  | python -X utf8 -c "import sys,json; sys.stdout.write(json.load(sys.stdin)['result'])" >input.conllu

Then the CoNLL-U file can be processed by CorPipe 25, by using for example

python3 corpipe25.py --load ufal/corpipe25-corefud1.3-large-251101 --exp . --epoch 0 --test input.conllu

which would generate the following predictions in input.00.conllu:

# generator = UDPipe 2, https://lindat.mff.cuni.cz/services/udpipe
# udpipe_model = english-ewt-ud-2.17-251125
# udpipe_model_licence = CC BY-NC-SA
# newdoc
# global.Entity = eid-etype-head-other
# newpar
# sent_id = 1
# text = Eve came home and Peter greeted her there.
1	Eve	Eve	PROPN	NNP	Number=Sing	2	nsubj	_	Entity=(c1--1)
2	came	come	VERB	VBD	Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	0	root	_	_
3	home	home	ADV	RB	_	2	advmod	_	Entity=(c2--1)
4	and	and	CCONJ	CC	_	6	cc	_	_
5	Peter	Peter	PROPN	NNP	Number=Sing	6	nsubj	_	Entity=(c3--1)
6	greeted	greet	VERB	VBD	Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	2	conj	_	_
7	her	she	PRON	PRP	Case=Acc|Gender=Fem|Number=Sing|Person=3|PronType=Prs	6	obj	_	Entity=(c1--1)
8	there	there	ADV	RB	PronType=Dem	6	advmod	_	Entity=(c2--1)|SpaceAfter=No
9	.	.	PUNCT	.	_	2	punct	_	_

# sent_id = 2
# text = Then Peter and Paul set out to a trip and Eve waved them off.
1	Then	then	ADV	RB	PronType=Dem	5	advmod	_	_
2	Peter	Peter	PROPN	NNP	Number=Sing	5	nsubj	_	Entity=(c4--1(c3--1)
3	and	and	CCONJ	CC	_	4	cc	_	_
4	Paul	Paul	PROPN	NNP	Number=Sing	2	conj	_	Entity=(c5--1)c4)
5	set	set	VERB	VBD	Mood=Ind|Number=Plur|Person=3|Tense=Past|VerbForm=Fin	0	root	_	_
6	out	out	ADP	RP	_	5	compound:prt	_	_
7	to	to	ADP	IN	_	9	case	_	_
8	a	a	DET	DT	Definite=Ind|PronType=Art	9	det	_	Entity=(c6--2
9	trip	trip	NOUN	NN	Number=Sing	5	obl	_	Entity=c6)
10	and	and	CCONJ	CC	_	12	cc	_	_
11	Eve	Eve	PROPN	NNP	Number=Sing	12	nsubj	_	Entity=(c1--1)
12	waved	wave	VERB	VBD	Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	5	conj	_	_
13	them	they	PRON	PRP	Case=Acc|Number=Plur|Person=3|PronType=Prs	12	obj	_	Entity=(c4--1)
14	off	off	ADP	RP	_	12	compound:prt	_	SpaceAfter=No
15	.	.	PUNCT	.	_	5	punct	_	SpaceAfter=No

How to Cite

@inproceedings{straka-2025-corpipe,
  title = "{C}or{P}ipe at {CRAC} 2025: Evaluating Multilingual Encoders for Multilingual Coreference Resolution",
  author = "Straka, Milan",
  editor = "Ogrodniczuk, Maciej and Novak, Michal and Poesio, Massimo and Pradhan, Sameer and Ng, Vincent",
  booktitle = "Proceedings of the Eighth Workshop on Computational Models of Reference, Anaphora and Coreference",
  month = nov,
  year = "2025",
  address = "Suzhou, China",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2025.crac-1.11/",
  doi = "10.18653/v1/2025.crac-1.11",
  pages = "130--139",
}