End-to-end Neural Coreference Resolution

Introduction

This repository contains the code for replicating results from

• End-to-end Neural Coreference Resolution
• Kenton Lee, Luheng He, Mike Lewis and Luke Zettlemoyer
• In Proceedings of the Conference on Empirical Methods in Natural Language Process (EMNLP), 2017

A demo of the code can be found here: http://www.kentonl.com/e2e-coref.

Requirements

• Python 2.7
  • TensorFlow 1.0.0
  • pyhocon (for parsing the configurations)
  • NLTK (for sentence splitting and tokenization in the demo)

Setting Up

• 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.

Training Instructions

• Experiment configurations are found in experiments.conf
• Choose an experiment that you would like to run, e.g. best
• For a single-machine experiment, run the following two commands:
  • python singleton.py <experiment>
  • python evaluator.py <experiment>
• For a distributed multi-gpu experiment, edit the cluster property of the configuration and run the following commands:
  • python parameter_server.py <experiment>
  • python worker.py <experiment> (for every worker in your cluster)
  • python evaluator.py <experiment> (on the same machine as your first worker)
• Results are stored in the logs directory and can be viewed via TensorBoard.
• For final evaluation of the checkpoint with the maximum dev F1:
  • Run python test_single.py <experiment> for the single-model evaluation.
  • Run python test_ensemble.py <experiment1> <experiment2> <experiment3>... for the ensemble-model evaluation.

Demo Instructions

• For the command-line demo with the pretrained model:
  • Run python demo.py final
• For the web demo with the pretrained model:
  • Run python demo.py final 8080
  • Edit the URL at the end of docs/main.js to point to the demo location, e.g. localhost:8080
  • Open docs/index.html in a web browser.
• To run the demo with other experiments, replace final with your configuration name.

Batched Prediction Instructions

• 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.
• Change the value of eval_path in the configuration file to the path to this new file.
• Run python decoder.py <experiment> <output_file>, which outputs the original file extended with annotations of the predicted clusters, the top spans, and the head attention scores.
• To visualize the predictions, place the output file in the viz directory and run run.sh. This will run a web server hosting the files in the viz directory. If run locally, it can be reached at http://localhost:8080?path=<output_file>

Other Quirks

• 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 evaluator should not be run on GPUs, since evaluating full documents does not fit within GPU memory constraints.
• The training runs indefinitely and needs to be terminated manually. The model generally converges at about 400k steps and within 48 hours.
• On some machines, the custom kernels seem to have compatibility issues with virtualenv. If you are using virtualenv and observe segmentation faults, trying running the experiments without virtualenv.