中文NER的那些事儿

The code is not rigorously tested, if you find a bug, welcome PR ^_^ ~

版本和环境配置详见requirement.txt, 数据和预训练模型的下载链接在对应folder的README中～

支持模型

1. 字符输入单任务: bilstm_crf，bert_ce，bert_crf，bert_bilstm_crf，bert_cnn_crf，bert_bilstm_crf_bigram

2. 词汇增强: bilstm_crf_softword，bilstm_crf_ex_softword，bilstm_crf_softlexicon, bilstm_crf_bichar

3. 多任务

• bert_bilstm_crf_mtl: 共享Bert的多任务联合学习
• bert_bilstm_crf_adv: 对抗迁移联合学习

4. Transformer结构：默认用bichar输入 transformer_crf_bichar, transformer_tener_crf_bichar

5. 数据增强：data/people_daily_augment，支持实体替换，Bert MASK替换，句子shuffle，同义词替换

6. MRC框架 + BIO Tagging Schema

训练&评估

1. pretrain_model中下载对应预训练模型到对应Folder，具体详见Folder中README.md
2. data中运行对应数据集preprocess.py得到tfrecord和data_params，训练会根据model_name选择以下tokenizer生成的tfrecord
   • Bert类模型用了wordPiece tokenizer，依赖以上预训练Bert模型的vocab文件
   • 非Bert类模型，包括词汇增强模型用了Giga和ctb50的预训练词向量
3. 运行单任务NER模型

[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object]

4. 运行多任务NER模型：按输入数据集类型可以是NER+NER的迁移/联合任务，也可以是NER+CWS的分词增强的NER任务。当前都是Joint Train暂不支持Alternative Train

[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object]

5. 评估：以上模型训练完会dump测试集的预测结果到data，repo里已经上传了现成的预测结果可用

[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object],[object Object]

<p float="left"> <img src="https://files.mdnice.com/user/8955/a112ebb1-eb85-45d8-8ada-16ce5906b5d9.png" width="70%" /> &nbsp; &nbsp; &nbsp; &nbsp; <img src="https://files.mdnice.com/user/8955/c13cf469-76d6-47b2-a99a-b19083cfae4b.png" width="70%" /> </p>

推理

1. 下载docker https://docs.docker.com/get-docker/

2. 下载tf docker image

[object Object],[object Object]

3. warmup (optional), serving_model中提供的3个模型已经运行过warmup

[object Object],[object Object],[object Object]

4. run server: server会从inferece.py中读取推理用的model_name

[object Object],[object Object]

5. run client: 输入文本返回预测

[object Object],[object Object],[object Object]

下图为无warmp的infer latency 下图为加入warmup后的infer latency

博客

中文NER的那些事儿1. Bert-Bilstm-CRF基线模型详解&代码实现

中文NER的那些事儿2. 多任务，对抗迁移学习详解&代码实现

中文NER的那些事儿3. SoftLexicon等词汇增强详解&代码实现

tensorflow踩坑合集2. TF Serving & gRPC 踩坑

中文NER的那些事儿4. 数据增强在NER的尝试

中文NER的那些事儿5. Transformer相对位置编码&TENER代码实现

中文NER的那些事儿6. NER新范式！你问我答之MRC