cpf-nlpr/at4chinesener

Adversarial Transfer Learning for Chinese Named Entity Recognition with Self-Attention Mechanism

at4chinesener's Issues

paper

Can I ask where to find this paper?

how can I get SighanNER dataset？

有谁能分享一下这个数据集吗

could not find the SigHan dataset

Hi, dear PengFei Cao, I have try my best to get the SigHan Dataset, But still not find. I have send a email to you for help. So if you can, can you share the dataset with me?

Is it necessary to add self-attention?

In my NER experiments, it seems that self-attention did not improve the model's effectiveness.
I just had a problem. I want to confirm it.

Weibo_model代码疑惑

在Weibo_model.py中，181行和182行中，计算output和W_ner时，请问4这个数字是指什么，有些疑惑？是否可以替换为其他数字呢

how adv_loss improve the accuracy of NER

好像adv_loss是在训练一个数据集来自那个任务的分类器，那么这个分类器是如何提高命名实体的精度的呢，菜鸟求解答

找不到文件

data目录下的msr_training.utf8怎么没有?

预训练中的embed.txt是怎么得来的呢？

预训练中的embed.txt是怎么得来的呢？是通过word2vec得来的，还是随机生成的？

最大长度80

按照您设置的最大长度80，我对三个数据集进行统计，
训练集，验证集，测试集，句子总数为：135 270 270
大于80的句子数占各个数据集的比例为：0.22，0.20,0.22，跟您前面说过的小于80的占95%以上不符，也就是你说的这个比例是小于0.05，这个差距有点大，谢谢。

how can you sent this paper to me?

[email protected]
thanks

pytorch version

I release a pytorch version about this model in https://github.com/tyistyler/pytorch-AT4ChineseNER, maybe it has some questions and we can talk about it~

weibo数据达不到论文效果

我使用这个代码发现在测试集上最好的F1：51.51,达不到53.08。

InvalidArgumentError with Other MSRA Dataset

Hi there,

As SIGHANNER dataset isn't available on your given link, if other MSRA datasets are used it shows following errors.

GPU:

InvalidArgumentError (see above for traceback): indices[0,1] = 59 is not in [0, 49)
	 [[node ner_model/ner_crf/GatherV2_1 (defined at <ipython-input-2-f9f661903dfb>:176) ]]

CPU:

InvalidArgumentError (see above for traceback): indices[0,0] = 59 is not in [0, 49)
         [[Node: ner_model/ner_crf/Gather_1 = Gather[Tindices=DT_INT32, Tparams=DT_FLOAT, 
validate_indices=true, _device="/job:localhost/replica:0/task:0/cpu:0"]
(ner_model/ner_crf/Reshape_3, ner_model/ner_crf/add_3)]]

代码的模型好像跟论文中的模型不太一样

代码里的ner和cws训练时都将input给到了‘common_bilstm'和’cws_private_bilstm'和‘ner_private_bilstm'，论文里不是分开的吗？

加载模型进行测试

用了作者训练的模型，后来又用自己训练的模型测试，但测试结果都很差，请问test_weibo.py是否完整

something about dev data

In your published paper, you claim that "Since the SighanNER does not have development set, we sample 10% data of training set as development set". However, I can neither find development set of SighanNER ,or development set of MSRANER that used for selecting best model in your code.

where can I get the paper?

Excuse me .where can I get the paper? Or can you send the paper to my email?[email protected]
thx