This code is for ECAI 2020 paper "Joint Extraction of Entities and Relations Based on a Novel Decomposition Strategy", in which we first decompose the joint extraction task into two interrelated subtasks, namely HE extraction and TER extraction. The former subtask is to distinguish all head-entities that may be involved with target relations, and the latter is to identify corresponding tail-entities and relations for each extracted head-entity. Next, these two subtasks are further deconstructed into several sequence labeling problems based on our proposed span-based tagging scheme, which are conveniently solved by a hierarchical boundary tagger and a multi-span decoding algorithm. Experimental results show that our method outperforms previous work by 5.2%, 5.9% and 21.5% (F1 score), achieving a new state-of-the-art on three public datasets as follows:
You can find the paper here
See below for an overview of the model architecture:
Our model was trained on GPU GeForce 2080 RTX Ti.
-
Python 3 (tested on 3.6.8)
-
PyTorch (tested on 1.0.1)
-
CUDA (tested on 9.0.176)
-
unzip, wget (for downloading only)
The pre-processed JSON files of NYT-multi and WebNLG have been put under the directory dataset/NYT-multi/data and dataset/WebNLG/data, respectively.
First, download and unzip GloVe vectors:
chmod +x download.sh; ./download.sh
Then prepare vocabulary, initial word vectors and tagging scheme of NYT-multi with:
python prepare_data.py --data_dir dataset/NYT-multi/data --vocab_dir dataset/NYT-multi/vocab
Similar operation can also be applied to the WebNLG dataset.
To train the model on the NYT-multi dataset, run:
python train.py --data_dir dataset/NYT-multi/data --vocab_dir dataset/NYT-multi/vocab --id multi
Model checkpoints, logs and the output file of dev setwill be saved to ./saved_models/multi.
For details on the use of other parameters, please refer to train.py.
After training, our pretrained model is saved under the dir saved_models/multi. To run evaluation on the test set, run:
python eval.py saved_models/multi --data_dir dataset/NYT-multi/data --dataset test
This will use the best_model.pt file by default. Use --model checkpoint_epoch_10.pt to specify a model checkpoint file.
Codes are adapted from the repo of Position-aware Attention and Supervised Data Improve Slot Filling. We obtained the original datasets of NYT-multi and WebNLG from Extracting Relational Facts by an End-to-End Neural Model with Copy Mechanism, and pre-processed them with StanfordCoreNLP.
@inproceedings{yu2020jointer,
author = {Yu, Bowen and Zhang, Zhenyu and Shu, Xiaobo and Wang, Yubin and Liu, Tingwen and Wang, Bin and Li, Sujian},
booktitle = {Proc. of ECAI},
title = {Joint Extraction of Entities and Relations Based on a Novel Decomposition Strategy},
year = {2020}
}