runxinxu / git Goto Github PK

Source code for ACL-IJCNLP 2021 Long paper: Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker.

Python 99.49% Shell 0.51%

document-level event-extraction nlp information-extraction

git's Introduction

Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker

Source code for ACL-IJCNLP 2021 Long paper: Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker.

Our code is based on Doc2EDAG.

0. Introduction

Document-level event extraction aims to extract events within a document. Different from sentence-level event extraction, the arguments of an event record may scatter across sentences, which requires a comprehensive understanding of the cross-sentence context. Besides, a document may express several correlated events simultaneously, and recognizing the interdependency among them is fundamental to successful extraction. To tackle the aforementioned two challenges, We propose a novel heterogeneous Graph-based Interaction Model with a Tracker (GIT). A graph-based interaction network is introduced to capture the global context for the scattered event arguments across sentences with different heterogeneous edges. We also decode event records with a Tracker module, which tracks the extracted event records, so that the interdependency among events is taken into consideration. Our approach delivers better results over the state-of-the-art methods, especially in cross-sentence events and multiple events scenarios.

Architecture
Overall Results

1. Package Description

GIT/
├─ dee/
    ├── __init__.py
    ├── base_task.py
    ├── dee_task.py
    ├── ner_task.py
    ├── dee_helper.py: data features constrcution and evaluation utils
    ├── dee_metric.py: data evaluation utils
    ├── config.py: process command arguments
    ├── dee_model.py: GIT model
    ├── ner_model.py
    ├── transformer.py: transformer module
    ├── utils.py: utils
├─ run_dee_task.py: the main entry
├─ train_multi.sh
├─ run_train.sh: script for training (including evaluation)
├─ run_eval.sh: script for evaluation
├─ Exps/: experiment outputs
├─ Data.zip
├─ Data: unzip Data.zip
├─ LICENSE
├─ README.md

2. Environments

python (3.6.9)
cuda (11.1)
Ubuntu-18.0.4 (5.4.0-73-generic)

3. Dependencies

numpy (1.19.5)
torch (1.8.1+cu111)
pytorch-pretrained-bert (0.4.0)
dgl-cu111 (0.6.1)
tensorboardX (2.2)

PS: The environments and dependencies listed here is different from what we use in our paper, so the results may be a bit different.

4. Preparation

Unzip Data.zip and you can get an Data folder, where the training/dev/test data locate.

5. Training

>> bash run_train.sh

6. Evaluation

>> bash run_eval.sh

(The evaluation is also conducted after the training)

7. License

This project is licensed under the MIT License - see the LICENSE file for details.

8. Citation

If you use this work or code, please kindly cite the following paper:

@inproceedings{xu-etal-2021-git,
    title = "Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker",
    author = "Runxin Xu  and
      Tianyu Liu  and
      Lei Li and
      Baobao Chang",
    booktitle = "The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021)",
    year = "2021",
    publisher = "Association for Computational Linguistics",
}

git's People

Contributors

Stargazers

Watchers

git's Issues

询问下训练的设备和时间

Released code fail to achieve the performance in paper. 代码无法复现出论文中的指标。

We find the released code fail to achieve the performance in paper. We observe that the max_sent_num in your code is 32 while it's 64 in previous Doc2EDAG. We wonder if it's the reason?
我们发现开源代码无法复现出论文的指标，我们观察到GIT最大句数的设置是32而Doc2EDAG是64。不知道是否是这个原因呢？

您好，代码好像有处小错误

dee_model.py 构建图时，第三种边，第720行到721行，len函数是不是应该不需要呢

您好，我想问一下gcn输出的计算

您好，我正在阅读您的代码和文章，有一些看不懂，希望您帮我解惑

请问这个公式怎么理解呢，
还有

是把L层拼接起来计算这个节点的最后一层embedding嘛

a query about role embedding

Hi, when I read the paper, I notice the words "role embedding" as follows. I read your codes carefully, but I still can't understand how to get it. Is it embedded by role_idx mapping? Moreover, I'd appreciate it if you could tell me where it is implemented in the code. Thanks!

Why does NER module in GIT use the encoder-decoder transformer rather than the encoder part only(like BERT used in Doc2EDAG)

Thanks for the excellent paper. But I have a question for the experiment setting:
As a sequence labeling task, we always use transformer encoder(like BERT) to solve the NER problem, for example, the author in Doc2EDAG use BERT as the first step backbone. However, in the paper, it is said that the vanilla transformer(encoder-decoder structure) is used in NER module, which confuse me a lot. I am wondering wher the decoder part of transformer is used for? Thanks.

A problem occurred at runtime

subprocess.CalledProcessError: Command '['/opt/anaconda3/envs/Doc2edag/bin/python', '-u', 'run_dee_task.py', '--local_rank=1', '--resume_latest_cpt', 'True', '--save_cpt_flag', 'True', '--data_dir', './Data', '--exp_dir', './Exps', '--task_name', 'try', '--num_train_epochs', '50', '--train_batch_size', '64', '--gradient_accumulation_steps', '8', '--eval_batch_size', '2', '--cpt_file_name', 'GIT']' returned non-zero exit status 1.

May I ask why this problem is? In addition, I have successfully reproduced Doc2EDAG, but the extraction results of event parameters are all numbers and cannot be visualized. How do you solve this problem

结果文件如何解析成数据集的样式

您好，在尝试用您的模型跑出结果后，得到如：./Exps/try/Output/dee_eval.test.pred_span.GIT.5.pkl的结果文件，其中一条数据是：
'''
(0, [0, 0, 1, 1, 0], [None, None, [[None, None, None, None, None, None]], [[(3330, 1290), None, None, None, None, None]], None], DocSpanInfo(span_token_tup_list=[(121, 121, 121, 127, 121, 126), (3943, 3862, 5500, 819), (4507,), (3330, 1290)], span_dranges_list=[[(0, 5, 11)], [(0, 16, 20)], [(9, 0, 1)], [(12, 10, 12), (13, 8, 10)]], span_mention_range_list=[(0, 1), (1, 2), (2, 3), (3, 5)], mention_drange_list=[(0, 5, 11), (0, 16, 20), (9, 0, 1), (12, 10, 12), (13, 8, 10)], mention_type_list=[1, 3, 5, 7, 7], event_dag_info=[None, None, None, None, [{(): {None}}, {(None,): {None}}, {(None, None): {None}}, {(None, None, None): {None}}, {(None, None, None, None): {None}}, {(None, None, None, None, None): {None}}, {(None, None, None, None, None, None): {None}}, {(None, None, None, None, None, None, None): {None}}, {(None, None, None, None, None, None, None, None): {None}}]], missed_sent_idx_list=[1, 4, 7, 8, 9, 10, 12, 13, 14, 16, 19]), [None, None, [(None, None, None, None, None, None)], [(3, None, None, None, None, None)], None])
'''
请问如何对上述数据进行解析，得到数据集中
'''
"recguid_eventname_eventdict_list": [
[
0,
"EquityPledge",
{
"Pledger": "李华青",
"PledgedShares": "1188600股",
"Pledgee": "海通证券股份有限公司",
"TotalHoldingShares": "22619999股",
"TotalHoldingRatio": "6.41%",
"TotalPledgedShares": "18200000股",
"StartDate": "2018年9月6日",
"EndDate": null,
"ReleasedDate": null
}
],
[
1,
"EquityPledge",
{
"Pledger": "李华青",
"PledgedShares": "12151000股",
"Pledgee": "海通证券股份有限公司",
"TotalHoldingShares": null,
"TotalHoldingRatio": "6.41%",
"TotalPledgedShares": "12151000股",
"StartDate": "2017年12月7日",
"EndDate": null,
"ReleasedDate": null
}
]
]
'''
的结果？
是否有结果解析的相关代码方便提供？
谢谢！

运行不了

请问怎么使用单GPU运行