Home

Awesome

Structured Minimally Supervised Learning for Neural Relation Extraction

This repo contains the pytorch code for paper Structured Minimally Supervised Learning for Neural Relation Extraction.

@inproceedings{bai-ritter-2019-structured,
    title = "{S}tructured {M}inimally {S}upervised {L}earning for {N}eural {R}elation {E}xtraction",
    author = "Bai, Fan  and
      Ritter, Alan",
    booktitle = "Proceedings of the 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)",
    month = jun,
    year = "2019",
    address = "Minneapolis, Minnesota",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/N19-1310",
    doi = "10.18653/v1/N19-1310",
    pages = "3057--3069",
}

Requirements

Dataset

Two datasets NYTFB-68K/NYTFB-280K can be found here.

NYTFB-68K: Riedel et. al. HeldOut dataset.

NYTFB-280K: Lin et. al. dataset removing overlapping entity pairs from training data(only entity names are shared not sentences).

Sentential DEV/TEST data: Manually annotated data created by Hoffmann et. al.

Please checkout the Appendix B of our paper for detailed introduction and comparison about these two datasets.

Training

Train a PCNN-NMAR model with a specific configuration:

python train.py --data_dir data/NYTFB-68K --lr 0.001 --penal_scalar 1000 --num_epoch 15 --save_dir saved_models/

With the above command, the model's checkpoint with best sentential AUC performance will be saved to ./saved_models/ as NYTFB-68K_lr0.001_penal1000_best_model.tar. You can save the checkpoint of every epoch by setting --save_each_epoch True, and perform heldou evaluation with --heldout_eval True.

Note: In the newest arXiv version of the paper, "Bag-Size Adaptive Learning Rate" has been rephrased as "Bag-Size Weighting Function" (a better presentation of the algorithm). Since it doesn't affect how the algorithm works, we don't modify the code.

Evaluation

All checkpoints used in our paper are stored under ./checkpoints_in_paper/

Sentential evaluation.

python eval.py --data_dir data/NYTFB-68K --model_dir checkpoints_in_paper/ --model_name NYTFB-68K_sentential.tar --sentential_eval True --sen_file sentential_DEV.txt

Heldout evaluation.

python eval.py --data_dir data/NYTFB-68K --model_dir checkpoints_in_paper/ --model_name NYTFB-68K_heldout.tar --heldout_eval True

You can also print out the configuration of the model by setting --print_config True

Reproduce

Since PCNN-NMAR is sensitive to the initialization, if you want to train the model from scratch to reproduce the result in our paper, you can run the script tune.sh with data directory and available GPU ids:

sh tune.sh "data/NYTFB-68K" "0 1 2 3"

This script will tune hyperparameters KB disagreement penalty scalars among {100, 200, ..., 2000} and learning rates among {0.001, 0.01}.

To select the model with best sentential result on DEV set among all saved models:

python eval.py --data_dir data/NYTFB-68K --model_dir saved_models/ --sentential_eval True --sen_file sentential_DEV.txt --tune True

Reference

[Riedel et al., 2010] Sebastian Riedel and Limin Yao and Andrew McCallum. Modeling Relations and Their Mentions without Labeled Text. In Proceedings of ECML.

[Hoffmann et al., 2011] Hoffmann, Raphael and Zhang, Congle and Ling, Xiao and Zettlemoyer, Luke and Weld, Daniel S. Knowledge-Based Weak Supervision for Information Extraction of Overlapping Relations. In Proceedings of ACL.

[Lin et al., 2016] Lin, Yankai and Shen, Shiqi and Liu, Zhiyuan and Luan, Huanbo and Sun, Maosong. Neural Relation Extraction with Selective Attention over Instances. In Proceedings of ACL. C++ code/data

[Bai et al., 2019] Bai, Fan and Ritter, Alan. Structured Minimally Supervised Learning for Neural Relation Extraction. In Proceedings of NAACL-HLT.