han liu
commited on
Commit
•
6d1608c
1
Parent(s):
29be213
add README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,71 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
# Erlangshen-BERT-120M-IE-Chinese
|
5 |
+
|
6 |
+
* Github: [GTS-Engine](https://github.com/IDEA-CCNL/GTS-Engine)
|
7 |
+
* Documentation: [GTS-Engine](https://gts-engine-doc.readthedocs.io/en/latest/docs/quick_start.html)
|
8 |
+
|
9 |
+
## 简介 Brief Introduction
|
10 |
+
|
11 |
+
本模型基于大规模信息抽取数据进行预训练,可支持few-shot、zero-shot场景下的实体识别、关系三元组抽取任务。
|
12 |
+
|
13 |
+
This model is pre-trained on large-scale information extraction data, to better support Named Entity Recognition (NER) and Relation Extraction (RE) tasks in few-shot/zero-shot scenarios.
|
14 |
+
|
15 |
+
## 模型分类 Model Taxonomy
|
16 |
+
|
17 |
+
| 需求 Demand | 任务 Task | 系列 Series | 模型 Model | 参数 Parameter | 额外 Extra |
|
18 |
+
| ---------- | ---------- | -------------- | -------- | ------------ | -------- |
|
19 |
+
| 通用 General | 信息抽取 Information Extraction | 二郎神 Erlangshen | BagualuIEModel | 120M | Chinese |
|
20 |
+
|
21 |
+
## 下游效果 Performance
|
22 |
+
|
23 |
+
Erlangshen-BERT-120M-IE-Chinese在多个信息抽取任务下进行测试。
|
24 |
+
|
25 |
+
其中,zh_weibo/MSRA/OntoNote4/Resume为NER任务,其中MSRA在原始数据下进行测试;SanWen/FinRE作为实体关系联合抽取任务进行测试,非单一关系分类任务。
|
26 |
+
|
27 |
+
部分参数设置如下:
|
28 |
+
```
|
29 |
+
batch_size=16
|
30 |
+
precision=16
|
31 |
+
max_epoch=50
|
32 |
+
lr=2e-5
|
33 |
+
weight_decay=0.1
|
34 |
+
warmup=0.06
|
35 |
+
max_length=512
|
36 |
+
```
|
37 |
+
|
38 |
+
我们分别在随机种子123/456/789下进行测试,并以[MacBERT-base, Chinese](https://github.com/ymcui/MacBERT)作为预训练模型保持相同参数进行训练作为对比baseline,得到效果计算平均,效果如下:
|
39 |
+
|
40 |
+
| Dataset | Training epochs | Test precision | Test recall | Test f1 | Baseline f1 |
|
41 |
+
| --------- | --------------- | -------------- | ----------- | ------- | ----------- |
|
42 |
+
| zh_weibo | 10.3 | 0.7282 | 0.6447 | 0.6839 | 0.6778 |
|
43 |
+
| MSRA | 5 | 0.9374 | 0.9299 | 0.9336 | 0.8483 |
|
44 |
+
| OntoNote4 | 9 | 0.8640 | 0.8634 | 0.8636 | 0.7996 |
|
45 |
+
| Resume | 15 | 0.9568 | 0.9658 | 0.9613 | 0.9479 |
|
46 |
+
| SanWen | 6.7 | 0.3655 | 0.2072 | 0.2639 | 0.2655 |
|
47 |
+
| FinRE | 7 | 0.5190 | 0.4274 | 0.4685 | 0.4559 |
|
48 |
+
|
49 |
+
## 使用 Usage
|
50 |
+
|
51 |
+
GTS引擎(GTS-Engine)是一款开箱即用且性能强大的自然语言理解引擎,能够仅用小样本就能自动化生产NLP模型。GTS Engine包含两个训练引擎:乾坤鼎和八卦炉。
|
52 |
+
|
53 |
+
本模型为可在GTS-Engine八卦炉引擎信息抽取任务中,作为预训练模型进行finetune。
|
54 |
+
|
55 |
+
GTS-Engine文档参考:[GTS-Engine](https://gts-engine-doc.readthedocs.io/en/latest/docs/about.html)
|
56 |
+
|
57 |
+
## 引用
|
58 |
+
|
59 |
+
如果您在您的工作中使用了我们的模型,可以引用我们的[网站](https://github.com/IDEA-CCNL/GTS-Engine):
|
60 |
+
|
61 |
+
You can also cite our website:
|
62 |
+
|
63 |
+
```
|
64 |
+
@misc{GTS-Engine,
|
65 |
+
title={GTS-Engine},
|
66 |
+
author={IDEA-CCNL},
|
67 |
+
year={2022},
|
68 |
+
howpublished={\url{https://github.com/IDEA-CCNL/GTS-Engine}},
|
69 |
+
}
|
70 |
+
```
|
71 |
+
|