Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 

Rewarding Coreference Resolvers for Being Consistent with World Knowledge

In proceedings of EMNLP 2019.

Datasets

For convinience, create a symlink: cd e2e-coref && ln -s ../wiki ./wiki

For pre-training the coreference resolution system, OntoNotes 5.0 is required. [Download] [Create splits]

Data for training the reward models and fine-tuning the coreference resolver (place in <PROJECT_HOME>/data):

Note: If you want to make these files from scratch, follow the instructions in the triples folder.

Pre-trained models

  • Best performing reward model (RE-Distill) [Download]
  • Best performing coreference resolver (Coref-Distill) [Download]

Evaluation

Unzip Coref-Distill into e2e-coref/logs folder and run GPU=x python evaluate.py final

Training

Reward models

  • Download pytorch big-graph embeddings (~40G, place in <PROJECT_HOME>/embeddings) [Download]
  • Run wiki/embs.py to create an index of the embeddings (you need to do this only once)
  • Run reward module training with cd wiki/reward && python train.py <dataset-name>

Coreference resolver

Pre-training
  • Follow e2e-coref/README.md to setup environment, create ELMO embeddings, etc.
  • Run coreference pre-training with cd e2e-coref && GPU=x python train.py <experiment>
Fine-tuning
  • Start the sling server with python wiki/reward/sling_server.py
  • Change SLING_IP in wiki/reward/reward.py to the IP of the sling server
  • Run coreference fine-tuning with cd e2e-coref && GPU=x python finetune.py <experiment> (see e2e-coref/experiments.conf for the different configurations)

Misc

  • wiki/reward/combine_models.py can be used to distill the various reward models
  • e2e-coref/save_weights.py can be used to save the weights of the fine-tuned coreference models so that they can be combined by setting the distill flag in the configuration file

Citation

@inproceedings{aralikatte-etal-2019-rewarding,
    title = "Rewarding Coreference Resolvers for Being Consistent with World Knowledge",
    author = "Aralikatte, Rahul  and
      Lent, Heather  and
      Gonzalez, Ana Valeria  and
      Herschcovich, Daniel  and
      Qiu, Chen  and
      Sandholm, Anders  and
      Ringaard, Michael  and
      S{\o}gaard, Anders",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
    month = nov,
    year = "2019",
    address = "Hong Kong, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D19-1118",
    doi = "10.18653/v1/D19-1118",
    pages = "1229--1235"
}

About

Rewarding Coreference Resolvers for Being Consistent with World Knowledge

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published