Generative Coreference Resolution

This repository contains the code implementation from our project "Generative Coreference Resolution".

Our code is based upon the work of the paper "Coreference Resolution without Span Representations" [1].

• Set up
  • Requirements
  • Download the official evaluation script
  • Prepare the dataset
• Evaluation
• Training

Set up

Requirements

Clone "Coreference Resolution without Span Representations"'s repository:

git clone https://github.com/yuvalkirstain/s2e-coref.git

Copy the required changes for running:

cp -r src/* s2e-coref/

Install the requirements:

cd s2e-coref
pip install -r requirements.txt

Download the official evaluation script

Run (from inside the repo):

git clone https://github.com/conll/reference-coreference-scorers.git

Prepare the dataset

This repo assumes access to the OntoNotes 5.0 corpus. Convert the original dataset into jsonlines format using:

export DATA_DIR=<data_dir>
python minimize.py $DATA_DIR

Credit: This script was taken from the e2e-coref repo.

Evaluation

Download our trained model:

export MODEL_DIR=<model_dir>
gdown --id 1uPzu-wAnMoO84tK_urRLxO7zN6eRQ2fy --output temp_model.zip
unzip temp_model.zip -d $MODEL_DIR
rm -rf temp_model.zip

and run:

export OUTPUT_DIR=<output_dir>
export CACHE_DIR=<cache_dir>
export MODEL_DIR=<model_dir>
export DATA_DIR=<data_dir>
export SPLIT_FOR_EVAL=<dev or test>
python run_config.py \
            --model_type t5-base \
            --split_for_eval test \
            --epochs 1 \
            --model_name_or_path $MODEL_DIR \
            --sent_num 10 \
            --step_num 10

Training

Train a coreference model using the run_config.py configuration:

export OUTPUT_DIR=<output_dir>
export CACHE_DIR=<cache_dir>
export MODEL_DIR=<model_dir>
export DATA_DIR=<data_dir>
export SPLIT_FOR_EVAL=<dev or test>
python run_config.py \
            --model_type t5-base \
            --split_for_eval test \
            --epochs <num of epochs> \
            --model_name_or_path $MODEL_DIR \
            --sent_num 10 \
            --step_num 10 \
            --do_train

For changes of more parameters run:

python run_config.py \
           run_config.py -h

and run accordingly

To evaluate your trained model on test go here.

References

[1] Coreference Resolution without Span Representations, 2021, Kirstain et al.