/FAST-NER

Code and dataset for IJCNN2019 paper 'Dilated Convolutional Networks Incorporating Soft Entity Type Constraints for Distant Supervised Relation Extraction'

Primary LanguagePythonMIT LicenseMIT

Dilated Convolutional Networks Incorporating Soft Entity Type Constraints for Distant Supervised Relation Extraction

Source code for IJCNN 2019 paper: Dilated Convolutional Networks Incorporating Soft Entity Type Constraints for Distant Supervised Relation Extraction

Overview

The architecture of our model, where $m_i$ indicates the original sentence for an entity pair. $p_j$ indicates the sentence feature encoded by DCNNs. Before calculating attention weight $\alpha_i$ for mention mi, its entity types are dynamically ajusted via the model prediction and current type labels to denoise and acquire soft entity type. Taking soft entity types into account, $\alpha_i$ is specified to the entity types of $m_i$. Finally, weighted summation is applied on all sentence features for producing bag feature $\mathbf{b}$.

Dependencies

  • Pytorch 1.0.1
  • tqdm 4.31.1
  • scikit-learn 0.20.3
  • Compatible with Python 3.X

Dataset and Pretrained models

  • Like the paper "Neural Relation Extraction with Selective Attention over Instances", we use Riedel NYT dataset for evaluation.
  • Our dataset can be downloaded from here, and the extraction code is "e9zu".
  • Our models can be downloaded from here, and the extraction code is "h37z".

Training from scratch

  • Options for preprocessing, traning and testing

    All options can be modified in nre/options.py

  • Preprocessing the dataset:

    python preprocess.py

  • Train the model:

    python train.py -batch_size 160 -model_type PDCNN+TATT -gpu_num 1

    After training, the trained model can be found in ckpt directory.

  • Test the model:

    python test.py -batch_size 160 -model_type PDCNN+TATT -gpu_num 1 -pretrain_model ckpt/model_step_best.pt

  • Plot precision-recall curves

    python plot_tool.py -pr_data_dir pr_data/baselines

Evaluation Results

Precision Recall Curve

Precision@N

P@N(%) 100 200 300 Mean
CNN+AVE 67.3 64.7 58.1 63.4
CNN+ATT 76.2 68.6 59.8 68.2
PCNN+AVE 73.3 66.7 62.8 67.6
PCNN+ATT 76.2 73.1 67.4 72.2
ResCNN-9 79.0 69.0 61.0 69.7
DCNN+ATT 75.2 72.1 67.0 71.5
PDCNN+ATT 77.2 77.6 71.4 75.4
CNN+TATT 79.2 73.1 66.4 72.9
PCNN+TATT 76.2 73.1 68.4 72.6
DCNN+TATT 77.2 70.6 72.1 73.3
PDCNN+TATT 80.2 76.1 70.8 75.7

Supplements

We notice the training set used by Lin et al., 2016 has 522,611 sentences, while some recent works used the training set which has 570,088 sentences. For a fair comparison, we also conduct experiments on the latter. The following are the results:

Precision Recall Curve

Precision@N

P@N(%) 100 200 300 Mean
PDCNN+TATT 83.2 81.1 76.4 80.2

Citation:

Please cite the following paper if you use this code in your work.

For any clarification, please create an issue or contact huweilong@whu.edu.cn.