识别3种命名实体，人名、地名、以及有关部门名
对比随机生成嵌入层词向量和使用预训练的词向量效果差别
命名实体使用如下label{O, B-PER, I-PER, B-LOC, I-LOC, B-ORG, I-ORG}，其中B为一个词的开头，I为非开头，O对应其他
所使用数据可到此处下载 https://pan.baidu.com/s/1RYdjzqkLcQkPCPiu9R0tyg 提取码: kw44
预训练词向量可到Chinese Word Vectors 中文词向量下载，本文使用的是人民日报的word 300d
embedding = tf.Variable(self.embedding,dtype = tf.float32)
self.word_embedding = tf.nn.embedding_lookup(embedding,self.x_data,name='word_embedding')
cell_fw = LSTMCell(300)
cell_bw = LSTMCell(300)
(output_fw_seq, output_bw_seq),_ = tf.nn.bidirectional_dynamic_rnn(cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=self.word_embedding,
sequence_length=self.seq_len,
dtype = tf.float32)
预训练的测向量维度为300，这里cell_num统一为300
fc = tf.layers.dense(out_put, self.num_tags)
fc = tf.contrib.layers.dropout(fc,self.dropout)
self.logits = tf.reshape(fc,[self.batch_size, max_len, self.num_tags])
log_likelihood, self.transition = crf_log_likelihood(inputs = self.logits,
tag_indices = self.y_data,
sequence_lengths = self.seq_len)
self.loss = -tf.reduce_mean(log_likelihood)
processed 171959 tokens with 6174 phrases; found: 5923 phrases; correct: 4996.
accuracy: 97.82%; precision: 84.35%; recall: 80.92%; FB1: 82.60
LOC: precision: 89.69%; recall: 84.73%; FB1: 87.14 2716
ORG: precision: 78.47%; recall: 79.41%; FB1: 78.94 1347
PER: precision: 80.81%; recall: 76.37%; FB1: 78.53 1860
从测试集来看，是有一定效果的，但只是对人名的识别效果较好
processed 171959 tokens with 6174 phrases; found: 6051 phrases; correct: 5297.
accuracy: 98.33%; precision: 87.54%; recall: 85.80%; FB1: 86.66
LOC: precision: 90.58%; recall: 88.00%; FB1: 89.27 2793
ORG: precision: 82.71%; recall: 82.64%; FB1: 82.68 1330
PER: precision: 86.46%; recall: 84.71%; FB1: 85.57 1928
使用预训练词向量后，对三种命名实体的识别都有一定的提升，后续如果希望继续提升效果，可以考虑使用其他的预训练词向量。
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent