-
Notifications
You must be signed in to change notification settings - Fork 5
命名实体识别实验结果
memeda edited this page Aug 30, 2016
·
1 revision
###参数设置:
参数 | 值 |
---|---|
词Embedding(随机初始化)维度 | 50 |
POSTAG Embedding(随机初始化)维度 | 5 |
NER Embedding(随机初始化)维度 | 5 |
LSTM X 维度 | 50 |
LSTM H 维度 | 100 |
TAG层隐层(MERGE层)维度 | 26 |
词典大小 | 47,216 |
输出层大小(输出标签数) | 13 |
-
进行连续数字转换
-
TAG层使用前一个TAG信息
-
以一定概率随机替换低频词为UNK
###实验结果*
语料 | LTP F1 | P | R | F1 |
---|---|---|---|---|
pku-holdout | 91.38 | 94.23% | 91.85% | 93.02% ↑ (+1.64) |
pku-test | 93.78 | 93.90% | 92.80% | 93.34% ↓ (-0.54) |
- 迭代轮次15轮
全部数据
PKU-HOLDOUT
accuracy: 99.02%; precision: 94.23%; recall: 91.85%; FB1: 93.02
Nh: precision: 95.37%; recall: 92.36%; FB1: 93.84 1318
Ni: precision: 92.14%; recall: 87.32%; FB1: 89.66 725
Ns: precision: 94.24%; recall: 93.14%; FB1: 93.69 2118
PKU-TEST
accuracy: 99.21%; precision: 93.90%; recall: 92.80%; FB1: 93.34
Nh: precision: 96.97%; recall: 95.06%; FB1: 96.00 3172
Ni: precision: 90.13%; recall: 86.37%; FB1: 88.21 2524
Ns: precision: 93.85%; recall: 94.55%; FB1: 94.20 5511
基于神经网络的序列标注任务 - WIKI (wiki语法见gollum)