Published in the conference STAR-SEM 2018. Link to paper is to be found here.

Run the startup script (bash startup.sh) to get started. It will pull and download all necessary repositories and datasets, including:

• The coreference scorer
• The ECB+ dataset

There are several Python package dependencies, including Theano, which is what we used for our experiments.

If you seek to implement our model, I would recommend a reimplementation in PyTorch or a more well-maintained deep learning library.

The predictions made by each model have been saved in results/, along with the gold standard coreference chains. After switching to the scripts directory (cd scripts/) you can do the following to replicate the results presented in the paper. For the within and cross-doc results:

bash get_scores.sh MODEL_NAME.response_conll

For just within-doc results:

bash get_scores.sh ecb_plus_events_test_mention_based_WITHINDOC_.key_conll  MODEL_NAME__within.response_conll

I do not currently have the time to document the Python code, but on request I can offer assistance over email. All of the code is found in python/. I would recommend reimplementation of the model if you seek to develop upon CORE. If you are interested primarily in the loss function and matrix derivation of CORE, check the file python/neural_cluster_model.py and the definition of the loss in the prepare_model function. Note that several of the files and functions are deprecated and were only used for preliminary experimentation.

Contact Kian Kenyon-Dean at kian.kenyon-dean@mail.mcgill.ca (or, on github) for questions about this repository.