Relation Aware Joint Entity Detection and Relation Extraction

Demo

For replicating the reported results, you can simply open Run.ipynb in Google Colab or use this link, and follow its instruction to replicate the results.

Use the trained model

1. Download the following files, which are our trained models, and put them in the /checkpoints directory.

• Rethinking: download
• CASREL: download

2- Now, you can evaluate the model using the following bash command:

python ./test.py \
--model_name="Casrel_Rethinking" \
--rel_num=171 \
--path="./checkpoints/model_rethinking" \
--test_prefix="test_triples" \
--dataset="WebNLG" \

Several important options include:

• --model_name: the model that is used, CASREL or Rethinking
• --path: the path of the downloaded checkpoint
• --test_prefix: the prefix of the data file that you want to evaluate the model with (you can use any file name in data/WebNLG without the .json extension)
• --dataset: the dataset that you are using

The expected results is as follows:

correct_num: 1406, predict_num: 1572, gold_num: 1581
f1: 0.8918, precision: 0.8944, recall: 0.8893

Train models from scratch

You can train the model from scratch using the following bash command:

python ./train.py \
--model_name="Casrel_Rethinking" \
--batch_size=6 \
--max_epoch=50 \
--test_epoch=10 \
--max_len=64 \
--rel_num=24 \
--dataset="NYT" \

Acknowledge

In this project, we used the following sources:

• PyTorch implementation of the original paper.
• Datasets and data processors provided by the original CASREL.