ÚFAL CorPipe: CRAC 2022 Winning System for Multilingual Coreference Resolution

This repository contains the source code of CorPipe, which is available under the MPL-2.0 license. The architecture of CorPipe is described in the following paper:

ÚFAL CorPipe at CRAC 2022: Effectivity of Multilingual Models for Coreference Resolution

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

Abstract: ÚFAL CorPipe is a winning submission to the CRAC 2022 Shared Task on Multilingual Coreference Resolution. Our system first solves mention detection and then coreference linking on the retrieved spans with an antecedent-maximization approach, and both tasks are fine-tuned jointly with shared Transformer weights. We report results of fine-tuning a wide range of pretrained models. The center of this contribution are fine-tuned multilingual models. We found one large multilingual model with sufficiently large encoder to increase performance on all datasets across the board, with the benefit not limited only to the underrepresented languages or groups of typologically relative languages.

Content of this Repository

• The directory data is for the CorefUD 1.0 data, and the preprocessed and tokenized version needed for training.

  • The script data/get.sh downloads and extracts the CorefUD 1.0 training and development data, plus the unannotated test data of the CRAC 2022 shared task.

• The corpipe.py is the complete CorPipe source file.

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

  • runs validation (using the official UD validator from the validator submodule) on the output data,
  • perform evaluation (using the official evaluation script from the corefud-scorer submodule), both without and with singletons.

• 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.

Training the CorPipe “Multilingual” Variant

To train a single multilingual model on all the data, you should

1. run the data/get.sh script to download the CorefUD data,
2. create a Python environments with the packages listed in requirements.txt,

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

   For training the large variant using RemBERT variant, we used

   tb="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"
   corpipe.py $(for c in $tb; do echo data/$c/$c; done) --resample=6000 4 5 5 1 2 3 1 4 4 3 2 4 3 --epochs=20 --lazy_adam --learning_rate_decay --crf --batch_size=8 --bert=google/rembert --learning_rate=1e-5 --segment=512 --right=50 --exp=large-rembert

   To train a base variant using XLM-R base, we used

   tb="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"
   corpipe.py $(for c in $tb; do echo data/$c/$c; done) --resample 6000 4 5 5 1 2 3 1 4 4 3 2 4 3 --epochs=30 --lazy_adam --learning_rate_decay --crf --batch_size=8 --bert=jplu/tf-xlm-roberta-base --learning_rate=2e-5 --segment=512 --right=50 --exp=base-xlmr

Currently no model is saved; instead, the script performs and saves predictions on both the development and test sets after every epoch.