TESTED ON CUDA 10.2
Suppose you are in /home/myuser/
git clone git@github.com:alexeyev/openie6.git
cd openie6
zenodo_get 4055395
tar -xvf openie6_models.tar.gz
docker build -t openie6 .
nvidia-docker run --gpus all -v /home:/home -it -p 8084:8084 openie6:latest bash
When 'inside':
cd /home/myuser/openie6/
echo "I love Luciano Pavarotti and Josep Maria Carreras Coll ." > sentences.txt
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 PYTHONPATH=imojie:imojie/allennlp:imojie/pytorch_transformers:$PYTHONPATH python run.py --mode splitpredict --inp sentences.txt --out predictions.txt --rescoring --task oie --gpus 1 --oie_model models/oie_model/epoch=14_eval_acc=0.551_v0.ckpt --conj_model models/conj_model/epoch=28_eval_acc=0.854.ckpt --rescore_model models/rescore_model --num_extractions 5
Result:
$ cat predictions.txt
I love Luciano Pavarotti and Josep Maria Carreras Coll .
0.52: (I; love; Luciano Pavarotti)
0.40: (I; love; Josep Maria Carreras Coll)
DO NOT install allennlp or pytorch_transformers! this code uses a modified version of those.
We have also managed to put together a Colab Notebook for running the conversion to 'simple' sentences. Works both in GPU and CPU environments.
I love Luciano Pavarotti and Jose Carreras .
I love Luciano Pavarotti .
I love Jose Carreras .
Have fun.
This repository contains the code for the paper:
OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction
Keshav Kolluru*, Vaibhav Adlakha*, Samarth Aggarwal, Mausam and Soumen Chakrabarti
EMNLP 2020
* denotes equal contribution
conda create -n openie6 python=3.6
conda activate openie6
pip install -r requirements.txt
python -m nltk.downloader stopwords
python -m nltk.downloader punkt
All results have been obtained on V100 GPU with CUDA 10.0 NOTE: HuggingFace transformers==2.6.0 is necessary. The latest version has a breaking change in the way tokenizer is used in the code. It will not raise an error but will give wrong results!
Download Data (50 MB)
zenodo_get 4054476
tar -xvf openie6_data.tar.gz
Download Models (6.6 GB)
zenodo_get 4055395
tar -xvf openie6_models.tar.gz
New command:
python run.py --mode splitpredict --inp sentences.txt --out predictions.txt --rescoring --task oie --gpus 1 --oie_model models/oie_model/epoch=14_eval_acc=0.551_v0.ckpt --conj_model models/conj_model/epoch=28_eval_acc=0.854.ckpt --rescore_model models/rescore_model --num_extractions 5
Expected models:
models/conj_model: Performs coordination analysis
models/oie_model: Performs OpenIE extraction
models/rescore_model: Performs the final rescoring
--inp sentences.txt - File with one sentence in each line --out predictions.txt - File containing the generated extractions
gpus - 0 for no GPU, 1 for single GPU
Additional flags -
--type labels // outputs word-level aligned labels to the file path `out`+'.labels'
--type sentences // outputs decomposed sentences to the file path `out`+'.sentences'
Additional Notes:
- The model is trained with tokenized sentences and hence requires tokenized sentences during prediction as well. The code currently uses nltk tokenization for this purpose. This will lead to the final sentences being different from the input sentences, as they will be the tokenized version. If this is not desirable you can comment the nltk tokenization in data.py and make sure that your sentences are tokenized beforehand.
- Due to an artifact of training data in conjunction model, it requires the sentence to end with full stop to function correctly.
Training:
python run.py --save models/warmup_oie_model --mode train_test --model_str bert-base-cased --task oie --epochs 30 --gpus 1 --batch_size 24 --optimizer adamW --lr 2e-05 --iterative_layers 2
Testing:
python run.py --save models/warmup_oie_model --mode test --batch_size 24 --model_str bert-base-cased --task oie --gpus 1
Carb F1: 52.4, Carb AUC: 33.8
Predicting
python run.py --save models/warmup_oie_model --mode predict --model_str bert-base-cased --task oie --gpus 1 --inp sentences.txt --out predictions.txt
Time (Approx): 142 extractions/second
Training
python run.py --save models/oie_model --mode resume --model_str bert-base-cased --task oie --epochs 16 --gpus 1 --batch_size 16 --optimizer adam --lr 5e-06 --iterative_layers 2 --checkpoint models/warmup_oie_model/epoch=13_eval_acc=0.544.ckpt --constraints posm_hvc_hvr_hve --save_k 3 --accumulate_grad_batches 2 --gradient_clip_val 1 --multi_opt --lr 2e-5 --wreg 1 --cweights 3_3_3_3 --val_check_interval 0.1
Testing
python run.py --save models/oie_model --mode test --batch_size 16 --model_str bert-base-cased --task oie --gpus 1
Carb F1: 54.0, Carb AUC: 35.7
Predicting
python run.py --save models/oie_model --mode predict --model_str bert-base-cased --task oie --gpus 1 --inp sentences.txt --out predictions.txt
Time (Approx): 142 extractions/second
NOTE: Due to a bug in the code, link, we end up using a loss function based only on the constrained loss term and not the original Cross Entropy (CE) loss. It still seems to work well as the warmup model is already trained with the CE loss and the constrained training is initialized from the warmup model.
python run.py --save models/conj_model --mode train_test --model_str bert-large-cased --task conj --epochs 40 --gpus 1 --batch_size 32 --optimizer adamW --lr 2e-05 --iterative_layers 2
F1: 87.8
Running
python run.py --mode splitpredict --inp carb/data/carb_sentences.txt --out models/results/final --rescoring --task oie --gpus 1 --oie_model models/oie_model/epoch=14_eval_acc=0.551_v0.ckpt --conj_model models/conj_model/epoch=28_eval_acc=0.854.ckpt --rescore_model models/rescore_model --num_extractions 5
python utils/oie_to_allennlp.py --inp models/results/final --out models/results/final.carb
python carb/carb.py --allennlp models/results/final.carb --gold carb/data/gold/test.tsv --out /dev/null
Carb F1: 52.7, Carb AUC: 33.7 Time (Approx): 31 extractions/second
Evaluate using other metrics (Carb(s,s), Wire57 and OIE-16)
bash carb/evaluate_all.sh models/results/final.carb carb/data/gold/test.tsv
Carb(s,s): F1: 46.4, AUC: 26.8 Carb(s,m) ==> Carb: F1: 52.7, AUC: 33.7 OIE16: F1: 65.6, AUC: 48.4 Wire57: F1: 40.0
If you use this code in your research, please cite:
@inproceedings{kolluru&al20,
title = "{O}pen{IE}6: {I}terative {G}rid {L}abeling and {C}oordination {A}nalysis for {O}pen {I}nformation {E}xtraction",\
author = "Kolluru, Keshav and
Adlakha, Vaibhav and
Aggarwal, Samarth and
Mausam, and
Chakrabarti, Soumen",
booktitle = "The 58th Annual Meeting of the Association for Computational Linguistics (ACL)",
month = July,
year = "2020",
address = {Seattle, U.S.A}
}
Note that the license is the full GPL, which allows many free uses, but not its use in proprietary software which is distributed to others. For distributors of proprietary software, you can contact us for commercial licensing.
In case of any issues, please send a mail to keshav.kolluru (at) gmail (dot) com