receiling / unire Goto Github PK

当我训练SciERC时，发生如下错误，有人遇到过吗？
[2021-12-30 12:30:40,742 - entity_relation_joint_decoder.py - line:274 - INFO]: Load bert tokenizer successfully.
Traceback (most recent call last):
File "C:\Users\Administrator\Desktop\UniRE-master\entity_relation_joint_decoder.py", line 330, in
main()
File "C:\Users\Administrator\Desktop\UniRE-master\entity_relation_joint_decoder.py", line 295, in main
ace_dataset.build_dataset(vocab=vocab,
File "C:\Users\Administrator\Desktop\UniRE-master\inputs\datasets\dataset.py", line 78, in build_dataset
instance_settting['instance'].count_vocab_items(counter,
File "C:\Users\Administrator\Desktop\UniRE-master\inputs\instance.py", line 61, in count_vocab_items
field.count_vocab_items(counter, sentences)
File "C:\Users\Administrator\Desktop\UniRE-master\inputs\fields\token_field.py", line 39, in count_vocab_items
for sentence in sentences:
File "C:\Users\Administrator\Desktop\UniRE-master\inputs\dataset_readers\ace_reader_for_joint_decoding.py", line 37, in iter
state, results = self.get_tokens(line)
File "C:\Users\Administrator\Desktop\UniRE-master\inputs\dataset_readers\ace_reader_for_joint_decoding.py", line 90, in get_tokens
logger.error("article id: {} sentence id: {} doesn't contain 'sentText'.".format(line['articleId'], line['sentId']))
KeyError: 'articleId'

你好，Pre-trained Models 里面的百度网盘链接里面文件没了，可以重新提供一下文件吗？

感谢作者分享的代码，我刚接触实体关系联合抽取任务。

为什么作者在文中提到了overlapping triples，可是不用NYT和WebNLG数据集，而要使用ACE呢？

我也注意到，实体关系联合抽取的文章，一部分用ACE，一部分用ＮＹＴ，这之间又有什么区别呢？

感谢楼主！

尊敬的作者，您好！

看了您的文章很受启发，但有些疑惑在于，您是怎么做ACE2004数据集上的5-fold交叉验证的，您是在这五个数据集上跑了五次然后取平均值吗？

其次，如果可以的话，希望获取ACE的数据集。([email protected])

祝您工作顺利，平安喜乐！

你好，我想请教一下关于数据集中两个实体之间关系为None的是如何在评估里面处理的？

Thanks for your amazing work!
I do the data-preprocess and train as the guide in the Readme.md
I see the code set the seed for random、torch and np.
I train the model with the SciERC dataset and early stop at epoch 229 , get the result:



I train it with a RTX3090 GPU without any modify the code (same as the Readme.md)
I see, in the paper, the result is:

the F1 for the ent is 68.4 and i get 66.03
How to reproduce the result in the paper?
Sincerely looking forward to your reply!

作者你好，感谢你们开源了代码，代码非常整洁易读。我注意到在论文中你们也提到UniRE无法处理overlap entity的情况，在阅读代码时，我发现这部分包含overlap的entity的句子似乎被直接扔掉了，我是根据

请问作者能分享一下数据集吗，readme没有提供数据集下载方式

ACE05数据拿来后要咋处理？处理成啥格式？可以提供一份处理后的样例看看嘛？

When I test on scierc dataset, the following errors are reported: error message:

Traceback (most recent call last):
  File "entity_relation_joint_decoder.py", line 330, in <module>
    main()
  File "entity_relation_joint_decoder.py", line 316, in main
    model.load_state_dict(state_dict,False)
  File "/home/jsj201-9/anaconda3/envs/unire/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1407, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for EntRelJointDecoder:
        size mismatch for embedding_model.bert_encoder.bert_model.embeddings.word_embeddings.weight: copying a param with shapetorch.Size([31090, 768]) from checkpoint, the shape in current model is torch.Size([30522, 768]).

Where is the file 'train.json'

使用如下提供的流程和代码，以及ACE2004数据集进行处理

https://github.com/LorrinWWW/two-are-better-than-one/tree/master/datasets

① 执行python ace2json.py后生成了如下文件

② 执行python unify.py后在unified目录下生成了如下文件：

请问针对您的脚本命令./ace2004.sh ace2004_folder，以及split.py，我该使用哪一个路径作为ace2004_folder这一参数？谢谢！

修改shell脚本中的

--device 0

为

--device 0, 1, 2 ,3

会报错，这里是否需要我自己进行实现，还是说如何可以做到？ 麻烦您了，谢谢！

不知道是什么情况呀？