Pytorch implementation of :https://github.com/rainarch/DSNER
python 3.6
pytorch
First: python script/preprocess-EC
main module: python dsner.py
args parameters should be define based on the paper
Result: Dev and Test respectively
| Model | Setup | epoch | P | R | F1 | P | R | F1 |
|---|---|---|---|---|---|---|---|---|
| LSTM+CRF | H | 562 | 66.30 % | 64.21 % | 65.24 % | 63.64 % | 62.53 % | 63.08 % |
| LSTM+CRF | H+A | 769 | 67.80 % | 58.53 % | 62.82 % | 63.65 % | 53.59 % | 58.19 % |
| LSTm+CRF+SL | H+A | 538 | 68.21 % | 61.89 % | 64.90 % | 66.90 % | 61.66 % | 64.17 % |
| LSTm+CRF+PA | H+A | 652 | 62.21 % | 70.11 % | 65.93 % | 61.12 % | 67.63 % | 64.21 % |
| LSTm+CRF+PA+SL | H+A | 370 | 69.12 % | 71.16 % | 70.12 % | 63.46 % | 63.94 % | 63.70 % |
Another paper which is used RL as a post-Processing for denoising with a different reward formula:
Reinforcement-based denoising of distantly supervised NER with partial annotation : https://www.aclweb.org/anthology/D19-6125/