This repository contains the code for replicating results from

• Higher-order Coreference Resolution with Coarse-to-fine Inference
• Kenton Lee, Luheng He, and Luke Zettlemoyer
• In NAACL 2018

• Install python (either 2 or 3) requirements: pip install -r requirements.txt
• Download pretrained word embeddings and build custom kernels by running setup_all.sh.
  • There are 3 platform-dependent ways to build custom TensorFlow kernels. Please comment/uncomment the appropriate lines in the script.
• Run one of the following:
  • To use the pretrained model only, run setup_pretrained.sh
  • To train your own models, run setup_training.sh
    • This assumes access to OntoNotes 5.0. Please edit the ontonotes_path variable.

• Experiment configurations are found in experiments.conf
• Choose an experiment that you would like to run, e.g. best
• Training: python train.py <experiment>
• Results are stored in the logs directory and can be viewed via TensorBoard.
• Evaluation: python evaluate.py <experiment>

• Command-line demo: python demo.py final
• To run the demo with other experiments, replace final with your configuration name.

• Create a file where each line is in the following json format (make sure to strip the newlines so each line is well-formed json):

{
  "clusters": [],
  "doc_key": "nw",
  "sentences": [["This", "is", "the", "first", "sentence", "."], ["This", "is", "the", "second", "."]],
  "speakers": [["spk1", "spk1", "spk1", "spk1", "spk1", "spk1"], ["spk2", "spk2", "spk2", "spk2", "spk2"]]
}

• clusters should be left empty and is only used for evaluation purposes.
• doc_key indicates the genre, which can be one of the following: "bc", "bn", "mz", "nw", "pt", "tc", "wb"
• speakers indicates the speaker of each word. These can be all empty strings if there is only one known speaker.
• Run python predict.py <experiment> <input_file> <output_file>, which outputs the input jsonlines with predicted clusters.

• It does not use GPUs by default. Instead, it looks for the GPU environment variable, which the code treats as shorthand for CUDA_VISIBLE_DEVICES.
• The training runs indefinitely and needs to be terminated manually. The model generally converges at about 400k steps.