python 3+ tensorflow 1.10+ keras 2.2.4+
我们将所使用的依赖环境已打包为 requirements.txt keras-bert from here https://github.com/CyberZHG/keras-bert
pip install keras-bert
conda install --yes --file requirements.txt请将以下文件放入对应文件夹
1,
./inputs (原始数据存放路径)
应包含以下文件：
train_data.json, dev_data.json, all_50_chemas, test_data_postag.json
2,
./bert (存放预训练模型权重路径)
应包含一下文件：
./bert/chinese_L-12_H-768_A-12/bert_config.json;
./bert/chinese_L-12_H-768_A-12/bert_model.ckpt;
./bert/chinese_L-12_H-768_A-12/vocab.txt
以下是预训练权重下载地址
chinese_L-12_H-768_A-12 ：https://storage.googleapis.com/bert_models/2018_11_03/chinese_L-12_H-768_A-12.zip
run python main.py gpu_num1 gpu_num2 gpu_num3 gpu_num4 gpu_num5
The test datasets predictions will be saved into a file called final_data.json in the outputs, and the models be trained will be saved into ensemble_part_x.weights in the models.
We used multiple GPUs for training and prediction, so we also specified multiple GPUS for training and forecasting in this program.
我们在苏神baseline上的工作:
1, BERT
2，优化了标注方式，针对重叠关系的重新设定了多信息的标注方式
3，简化了下游模型结构，尝试了self-attention和普通点乘attention
4，投票方式简单集成
5，规则数据后处理和预处理。
参考文献：
Global Normalization of Convolutional Neural Networks for Joint Entity and Relation Classification
One for All Neural Joint Modeling of Entities and Even
Table filling multi-task recurrent neural network for joint entity and relation extraction
Joint Extraction of Entities and Relations Based on a Novel Tagging Scheme
End-to-End Neural Relation Extraction with Global Optimization
A榜：0.889， B榜：0.8872 , 最终B榜第五（原本第六，第四名放弃）。
ps: A,B榜有差距差距的原因，是因为我不小心把几个用来集成的模型权重给用debug数据跑的模型覆盖了。。。。重新跑全部EPOCH时间又不够，就以重新跑了几个更少EPOCH的模型替代，导致重新预测效果变差，坑了。。。。