CatNER

CatNER基于NCRF++和cove,打算探究中间向量对NER的影响。

1 安装环境

Ubuntu：14.04（测试系统）
Python: 2.7
Pytorch：0.3
pip install -r requirements.txt

2 使用

>>> ./run[_cove]_main.sh

注意: 运行前需要更改权限

>>> chmod 755 run[_cove]_main.sh

3 实验

3.1 charlstm+lstm+crf

pretrain word:400000, prefect match:11415, case_match:11656, oov:2233, oov%:0.0882434301521
     Hyper       iteration: 100
     Hyper      batch size: 10
     Hyper   average batch: False
     Hyper              lr: 0.015
     Hyper        lr_decay: 0.05
     Hyper         HP_clip: None
     Hyper        momentum: 0
     Hyper      hidden_dim: 200
     Hyper         dropout: 0.5
     Hyper      lstm_layer: 1
     Hyper          bilstm: True
     Hyper             GPU: True
     Hyper        use_char: True
             Char_features: LSTM
best F1: 0.913147

3.2 训练100个epoch:

模型	超参数	F1
charlstm+lstm+crf	50_200	0.913147
average_batch_loss	50_200	0.897811
cnnlstm+lstm+crf	50_200	0.912601
clip+charlstm+lstm+crf	50.0_50_200	0.912119
batch+charlstm+lstm+crf	16_50_200	0.911215

3.3 cove+charlstm+lstm+crf

     Char embedding size: 30
     Norm   word     emb: False
     Norm   char     emb: False
     Train instance number: 14041
     Dev   instance number: 3250
     Test  instance number: 3453
     Raw   instance number: 0
     Hyper       iteration: 100
     Hyper      batch size: 10
     Hyper   average batch: False
     Hyper              lr: 0.015
     Hyper        lr_decay: 0.05
     Hyper         HP_clip: None
     Hyper        momentum: 0
     Hyper      hidden_dim: 1000
     Hyper         dropout: 0.5
     Hyper      lstm_layer: 1
     Hyper          bilstm: True
     Hyper             GPU: True
     Hyper        use_char: True
             Char_features: LSTM

3.4 训练100个epoch:

模型	超参数	F1
char+cove+word2vec	30_300_300-50_1000	0.905241
char+cove+noword2vec	30_300-50_650	0.877579
char+nocove+word2vec	30_300-50_350	0.910298
char+cove+word2vec	30_300_300-50_350	0.906286