This repo contains the implementation of our paper.
The original code was cloned from from this repo, which is a PyTorch implementation for paper Position-aware Attention and Supervised Data Improve Slot Filling.
The TACRED dataset: Details on the TAC Relation Extraction Dataset can be found on this dataset website.
- Python 3 (tested on 3.6.2)
- PyTorch (tested on 1.0.0)
- tqdm, bert_as_service, maybe a couple others
- unzip, wget (for downloading only)
First, download and unzip GloVe vectors from the Stanford website, with:
chmod +x download.sh; ./download.sh
Then tokenize data to run with BERT or SciBERT with:
python data/data_tok.py
Train using the commands in cmdcheat.txt. You need two terminal windows open, or two separate tmux sessions. Run the corresponding bert-as-service command and then run the python train.py command, both with the appropriate flags listed in cmdcheat.txt.
Model checkpoints and logs will be saved to ./saved_models/00.
Run evaluation on the test set with:
python eval.py saved_models/00 --dataset test
This will use the best_model.pt by default. Use --model checkpoint_epoch_10.pt to specify a model checkpoint file. Add --out saved_models/out/test1.pkl to write model probability output to files (for ensemble, etc.).
You will need bert-as-service running for the test phase as well.
Please see the example script ensemble.sh.
All work contained in this package is licensed under the Apache License, Version 2.0. See the included LICENSE file.