中文文档

pip install bert-multitask-learning

This a project that uses BERT to do multi-task learning with multiple GPU support.

In the original BERT code, neither multi-task learning or multiple GPU training is possible. Plus, the original purpose of this project is NER which dose not have a working script in the original BERT code.

To sum up, compared to the original bert repo, this repo has the following features:

1. Multi-task learning(major reason of re-writing the majority of code).
2. Multiple GPU training
3. Support sequence labeling (for example, NER) and Encoder-Decoder Seq2Seq(with transformer decoder).

• Classification(cls)
• Sequence Labeling(seq_tag)
• Seq2seq Labeling(seq2seq_tag)
• Seq2seq Text Generation(seq2seq_text)
• (TODO)Multi-Label Classification(multi_cls)

There are two types of chaining operations can be used to chain problems.

• &. If two problems have the same inputs, they can be chained using &. Problems chained by & will be trained at the same time.
• |. If two problems don't have the same inputs, they need to be chained using |. Problems chained by | will be sampled to train at every instance.

For example, cws|NER|weibo_ner&weibo_cws, one problem will be sampled at each turn, say weibo_ner&weibo_cws, then weibo_ner and weibo_cws will trained for this turn together. Therefore, in a particular batch, some tasks might not be sampled, and their loss could be 0 in this batch.

Please see the examples in notebooks for more details about training, evaluation and export models.

pip install bert-multitask-learning

这是利用BERT进行多任务学习并且支持多GPU训练的项目.

在原始的BERT代码中, 是没有办法直接用多GPU进行多任务学习的. 另外, BERT并没有给出序列标注和Seq2seq的训练代码.

因此, 和原来的BERT相比, 这个项目具有以下特点:

1. 多任务学习
2. 多GPU训练
3. 序列标注以及Encoder-decoder seq2seq的支持(用transformer decoder)

• 单标签分类(cls)
• 序列标注(seq_tag)
• 序列到序列标签标注(seq2seq_tag)
• 序列到序列文本生成(seq2seq_text)
• (TODO)多标签分类(multi_cls)

• 中文命名实体识别
• 中文分词
• 中文词性标注

可以用两种方法来将多个任务连接起来.

• &. 如果两个任务有相同的输入, 不同标签的话, 那么他们可以用&来连接. 被&连接起来的任务会被同时训练.
• |. 如果两个任务为不同的输入, 那么他们必须用|来连接. 被|连接起来的任务会被随机抽取来训练.

例如, 我们定义任务cws|NER|weibo_ner&weibo_cws, 那么在生成每一条数据时, 一个任务块会被随机抽取出来, 例如在这一次抽样中, weibo_ner&weibo_cws被选中. 那么这次weibo_ner和weibo_cws会被同时训练. 因此, 在一个batch中, 有可能某些任务没有被抽中, loss为0.

训练, eval和导出模型请见notebooks