Codes and datasets for our paper "Hierarchical Relation Extraction with Coarse-to-Fine Grained Attention"
If you use the code, please cite the following paper:
@inproceedings{han2018hierarchicalRE,
title={Hierarchical Relation Extraction with Coarse-to-Fine Grained Attention},
author={Han, Xu and Yu, Pengfei and Liu, Zhiyuan and Sun, Maosong and Li, Peng},
booktitle={Proceedings of EMNLP},
year={2018}
}
The model is implemented using tensorflow. The versions of packages used are shown below.
- tensorflow = 1.4.1
- numpy = 1.13.3
- scipy = 0.19.1
First unzip the ./raw_data/data.zip and put all the files under ./raw_data. Once the original raw text corpus data is in ./raw_data, run
python scripts/initial.py
For CNN hierarchical model,
PYTHONPATH=. python scripts/train.py --model cnn_hier
For PCNN,
PYTHONPATH=. python scripts/train.py --model pcnn_hier
Run various evaluation by specifying --mode in commandline, see the paper for detailed description for these evaluation methods.
PYTHONPATH=. python scripts/evaluate.py --mode [test method: pr, pone, ptwo, pall] --test-single --test_start_ckpt [ckpt number to be tested] --model [cnn_hier or pcnn_hier]
The logits are saved at ./outputs/logits/. To see the PR curve, run the following command which directly show() the curve, and you can adjust the codes in ./scripts/show_pr.py for saving the image as pdf file or etc. :
python scripts/show_pr.py [path/to/generated .npy logits file from evaluation]
The pretrained models is already saved at ./outputs/ckpt/. To directly evaluate on them, run the following command:
PYTHONPATH=. python scripts/evaluate.py --mode [test method: hit_k_100, hit_k_200, pr, pone, ptwo, pall] --test_single --test_start_ckpt 0 --model [cnn_hier or pcnn_hier]
And PR curves can be generated same way as above.
As this toolkit is reconstructed based on the original code and the checkpoints are retrained on this toolkit, the results of the released checkpoints are comparable with the reported ones.
- pr
| Model | Micro | Macro |
|---|---|---|
| CNN+HATT | 41.8 | 16.5 |
| PCNN+HATT | 42.0 | 17.1 |
- Hits@N(<100)
| micro | 10 | 15 | 20 |
|---|---|---|---|
| CNN+HATT | 5.6 | 33.3 | 50.0 |
| PCNN+HATT | 33.3 | 50.0 | 61.1 |
| macro | 10 | 15 | 20 |
|---|---|---|---|
| CNN+HATT | 5.6 | 29.6 | 57.4 |
| PCNN+HATT | 29.6 | 50.0 | 61.1 |
- Hits@N(<200)
| micro | 10 | 15 | 20 |
|---|---|---|---|
| CNN+HATT | 41.4 | 58.6 | 69.0 |
| PCNN+HATT | 55.2 | 69.0 | 75.9 |
| macro | 10 | 15 | 20 |
|---|---|---|---|
| CNN+HATT | 22.7 | 42.4 | 65.2 |
| PCNN+HATT | 41.4 | 59.1 | 68.2 |
- pone
| P@N | 100 | 200 | 300 | Mean |
|---|---|---|---|---|
| CNN+HATT | 84.0 | 76.5 | 70.7 | 77.1 |
| PCNN+HATT | 82.0 | 75.0 | 71.0 | 76.0 |
- ptwo
| P@N | 100 | 200 | 300 | Mean |
|---|---|---|---|---|
| CNN+HATT | 83.0 | 79.0 | 72.3 | 78.1 |
| PCNN+HATT | 82.0 | 77.0 | 75.3 | 78.1 |
- pall
| P@N | 100 | 200 | 300 | Mean |
|---|---|---|---|---|
| CNN+HATT | 82.0 | 80.0 | 74.7 | 78.9 |
| PCNN+HATT | 81.0 | 80.5 | 76.0 | 79.2 |
Some of other baselines can be found in other baselines.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent