This is a repository with the code to reproduce the experiments described in the paper of the same name, which was accepted to EMNLP 2021. The paper is available here.

1. Preparation
2. Training
3. Evaluation
4. Prediction
5. Citation

The following instruction has been tested with Python 3.7 on an Ubuntu 20.04 machine.

You will need:

• OntoNotes 5.0 corpus (download here, registration needed)
• Python 2.7 to run conll-2012 scripts
• Java runtime to run Stanford Parser
• Python 3.7+ to run the model
• Perl to run conll-2012 evaluation scripts
• CUDA-enabled machine (48 GB to train, 4 GB to evaluate)

1. Extract OntoNotes 5.0 arhive. In case it's in the repo's root directory:

    tar -xzvf ontonotes-release-5.0_LDC2013T19.tgz
   

2. Switch to Python 2.7 environment (where python would run 2.7 version). This is necessary for conll scripts to run correctly. To do it with with conda:

    conda create -y --name py27 python=2.7 && conda activate py27
   

3. Run the conll data preparation scripts (~30min):

    sh get_conll_data.sh ontonotes-release-5.0 data
   

4. Download conll scorers and Stanford Parser:

    sh get_third_party.sh
   

5. Prepare your environment. To do it with conda:

    conda create -y --name wl-coref python=3.7 openjdk perl
    conda activate wl-coref
    python -m pip install -r requirements.txt
   

6. Build the corpus in jsonlines format (~20 min):

    python convert_to_jsonlines.py data/conll-2012/ --out-dir data
    python convert_to_heads.py
   

You're all set!

If you have completed all the steps in the previous section, then just run:

python run.py train roberta

Use -h flag for more parameters and CUDA_VISIBLE_DEVICES environment variable to limit the cuda devices visible to the script. Refer to config.toml to modify existing model configurations or create your own.

Make sure that you have successfully completed all steps of the Preparation section.

1. Download and save the pretrained model to the data directory.

    https://www.dropbox.com/s/vf7zadyksgj40zu/roberta_%28e20_2021.05.02_01.16%29_release.pt?dl=0
   

2. Generate the conll-formatted output:

    python run.py eval roberta --data-split test
   

3. Run the conll-2012 scripts to obtain the metrics:

    python calculate_conll.py roberta test 20
   

To predict coreference relations on an arbitrary text, you will need to prepare the data in the jsonlines format (one json-formatted document per line). The following fields are requred:

    {
            "document_id": "tc_mydoc_001",
            "cased_words": ["Hi", "!", "Bye", "."],
            "sent_id": [0, 0, 1, 1]
    }

You can optionally provide the speaker data:

    {
            "speaker": ["Tom", "Tom", "#2", "#2"]
    }

document_id can be any string that starts with a two-letter genre identifier. The genres recognized are the following:

• bc: broadcast conversation
• bn: broadcast news
• mz: magazine genre (Sinorama magazine)
• nw: newswire genre
• pt: pivot text (The Bible)
• tc: telephone conversation (CallHome corpus)
• wb: web data

You can check a sample input file for reference.

Then run:

    python predict.py roberta input.jsonlines output.jsonlines

This will utilize the latest weights available in the data directory for the chosen configuration. To load other weights, use the --weights argument.

@inproceedings{dobrovolskii-2021-word,
title = "Word-Level Coreference Resolution",
author = "Dobrovolskii, Vladimir",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.605",
pages = "7670--7675"}