lemonhu
lemonhu
@sakuranew 谢谢关注与反馈。 我最近的实验，也表明在计算损失时引入mask会严重降低NER的效果，有些神奇（新版的pytorch-transformer引入了mask了loss）。 这个也作为该问题的一个实验性解答吧。
谢谢关注。 建议参考文件[evaluate.py](https://github.com/lemonhu/NER-BERT-pytorch/blob/master/evaluate.py)，即可实现模型的测试评估。
利用训练好的模型提供离线/在线服务，结合[data_loader.py](https://github.com/lemonhu/NER-BERT-pytorch/blob/master/data_loader.py)和[evaluate.py](https://github.com/lemonhu/NER-BERT-pytorch/blob/master/evaluate.py)，进行适当改写即可了。
Thank you for your attention. The msra dataset in this repo, I obtained from the [repo](https://github.com/dox1994/nlp_datasets), can not guarantee the authority of the data. Test data and statistical results should...
The reason is found. Due to the limitation of the maximum length of the sequence, sequences longer than this will be truncated, and the entities in the original sentence will...
谢谢关注。 根据你所给的结果，不太好直接给出结论，一个可能的原因是训练时的标签顺序编码`tag2idx`和测试时不一样。
The custom data can be trained in the format of the standard dataset (the BIO format), hopefully can help you.