Improving the Performance of Coreference Resolvers Across Datasets by Leveraging Target Entities

This repository contains code for self-labeling method which uses Named Entities from target data to automatically add coreference annotations and improve coreference resolver performance.

Install requirements

Setup a Python virtual environment

virtualenv venv-my --python=python3.6 or python3 -m venv venv-my 
source venv-my/bin/activate

Install the requirements:

pip install -r requirements.txt

Running the experiments

1. Start the Stanford CoreNLP server

nohup java -mx16g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 30000 &

2. Run self-labeling process, the parameter should be a json file consistent with the dataset above:

python3 annotate.py manual.json

3. Split the annoated data into training subset

python3 split_data.py path/to/annotated/data

Now the data is ready for training coreference resolution models.

4. Train e2e-coref model on the split data

Please refer to e2e-coref for detailed instructions about training and evaluating the model.

5. Pronoun evaluation

python3 scorer.py key_file prediction_file pro 

Please refer to CoVal for more details about pronoun evaluation.