Skip to content

Commit 354d97a

Browse files
author
tianxin04
committed
update README
1 parent d8afc4d commit 354d97a

File tree

1 file changed

+130
-1
lines changed

1 file changed

+130
-1
lines changed

‎ERNIE/README.md‎

Lines changed: 130 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,3 @@
1-
21
## Ernie: **E**nhanced **R**epresentation from k**N**owledge **I**nt**E**gration
32

43
*Ernie* 通过建模海量数据中的词、实体及实体关系，学习真实世界的语义知识。相较于 *Bert* 学习局部语言共现的语义表示，*Ernie* 直接对语义知识进行建模，增强了模型语义表示能力。
@@ -14,3 +13,133 @@
1413
此外， *Ernie* 引入了百科、新闻、论坛回帖等多源中文语料进行训练。
1514

1615
我们在多个公开的中文数据集合上进行了效果验证，*Ernie* 模型相较 *Bert*， 取得了更好的效果。
16+
17+
<table style="margin-left: 30.0px;">
18+
<tbody style="margin-left: 30.0px;">
19+
<tr style="margin-left: 30.0px;">
20+
<th class="confluenceTh"><strong>数据集</strong>
21+
<br></th>
22+
<th style="text-align: center;margin-left: 30.0px;" colspan="2"><strong>xnli</strong></th>
23+
<th style="text-align: center;margin-left: 30.0px;" colspan="2"><strong>lcqmc</strong></th>
24+
<th style="text-align: center;margin-left: 30.0px;" colspan="2"><strong>msra ner</strong></th>
25+
<th style="text-align: center;margin-left: 30.0px;" colspan="2"><strong>chnsenticorp</strong></th>
26+
<th style="text-align: center;margin-left: 30.0px;" colspan="4"><strong>nlpcc-dbqa</strong></th></tr>
27+
<tr style="margin-left: 30.0px;">
28+
<td rowspan="2">
29+
<p>
30+
<strong>评估</strong></p>
31+
<p>
32+
<strong>指标</strong>
33+
<br></p>
34+
</td>
35+
<td style="margin-left: 30px; text-align: center;" colspan="2">
36+
<strong>acc</strong>
37+
<br></td>
38+
<td style="margin-left: 30px; text-align: center;" colspan="2">
39+
<strong>acc</strong>
40+
<br></td>
41+
<td style="margin-left: 30px; text-align: center;" colspan="2">
42+
<strong>f1-score</strong>
43+
<br></td>
44+
<td style="margin-left: 30px; text-align: center;" colspan="2">
45+
<strong>acc</strong>
46+
<strong></strong>
47+
<br></td>
48+
<td style="margin-left: 30px; text-align: center;" colspan="2">
49+
<strong>mrr</strong>
50+
<br></td>
51+
<td style="margin-left: 30px; text-align: center;" colspan="2">
52+
<strong>f1-score</strong>
53+
<br></td>
54+
</tr>
55+
<tr style="margin-left: 30.0px;">
56+
<td colspan="1" style="text-align: center;" width="">
57+
<strong>dev</strong>
58+
<br></td>
59+
<td colspan="1" style="text-align: center;" width="">
60+
<strong>test</strong>
61+
<br></td>
62+
<td colspan="1" style="text-align: center;" width="">
63+
<strong>dev</strong>
64+
<br></td>
65+
<td colspan="1" style="text-align: center;" width="">
66+
<strong>test</strong>
67+
<br></td>
68+
<td colspan="1" style="text-align: center;" width="">
69+
<strong>dev</strong>
70+
<br></td>
71+
<td colspan="1" style="text-align: center;" width="">
72+
<strong>test</strong>
73+
<br></td>
74+
<td colspan="1" style="text-align: center;" width="">
75+
<strong>dev</strong>
76+
<br></td>
77+
<td colspan="1" style="text-align: center;" width="">
78+
<strong>test</strong>
79+
<br></td>
80+
<td colspan="1" style="text-align: center;" width="">
81+
<strong>dev</strong>
82+
<br></td>
83+
<td colspan="1" style="text-align: center;" width="">
84+
<strong>test</strong>
85+
<br></td>
86+
<td colspan="1" style="text-align: center;" width="">
87+
<strong>dev</strong>
88+
<br></td>
89+
<td colspan="1" style="text-align: center;" width="">
90+
<strong>test</strong>
91+
<br></td>
92+
</tr>
93+
<tr style="margin-left: 30.0px;">
94+
<td style="margin-left: 30.0px;">
95+
<strong>Bert
96+
<br></strong></td>
97+
<td style="margin-left: 30px; text-align: center;">78.1</td>
98+
<td style="margin-left: 30px; text-align: center;">77.2</td>
99+
<td style="margin-left: 30px; text-align: center;">88.8</td>
100+
<td style="margin-left: 30px; text-align: center;">87.0</td>
101+
<td style="margin-left: 30px; text-align: center;">94.0
102+
<br></td>
103+
<td style="margin-left: 30px; text-align: center;">
104+
<span>92.6</span></td>
105+
<td style="margin-left: 30px; text-align: center;">94.6</td>
106+
<td style="margin-left: 30px; text-align: center;">94.3</td>
107+
<td style="margin-left: 30px; text-align: center;" colspan="1">94.7</td>
108+
<td style="margin-left: 30px; text-align: center;" colspan="1">94.6</td>
109+
<td style="margin-left: 30px; text-align: center;" colspan="1">80.7</td>
110+
<td style="margin-left: 30px; text-align: center;" colspan="1">80.8</td></tr>
111+
<tr style="margin-left: 30.0px;">
112+
<td style="margin-left: 30.0px;">
113+
<strong>Ernie
114+
<br></strong></td>
115+
<td style="margin-left: 30px; text-align: center;">79.9 <span style="color: red;">(<strong>+1.8</strong>)</span></td>
116+
<td style="margin-left: 30px; text-align: center;">78.4 <span style="color: red;">(<strong>+1.2</strong>)</span></td>
117+
<td style="margin-left: 30px; text-align: center;">89.7 <span style="color: red;">(<strong>+0.9</strong>)</span></td>
118+
<td style="margin-left: 30px; text-align: center;">87.4 <span style="color: red;">(<strong>+0.4</strong>)</span></td>
119+
<td style="margin-left: 30px; text-align: center;">95.0 <span style="color: red;">(<strong>+1.0</strong>)</span></td>
120+
<td style="margin-left: 30px; text-align: center;">93.8 <span style="color: red;">(<strong>+1.2</strong>)</span></td>
121+
<td style="margin-left: 30px; text-align: center;">95.2 <span style="color: red;">(<strong>+0.6</strong>)</span></td>
122+
<td style="margin-left: 30px; text-align: center;">95.4 <span style="color: red;">(<strong>+1.1</strong>)</span></td>
123+
<td style="margin-left: 30px; text-align: center;" colspan="1">95.0 <span style="color: red;">(<strong>+0.3</strong>)</span></td>
124+
<td style="margin-left: 30px; text-align: center;" colspan="1">95.1 <span style="color: red;">(<strong>+0.5</strong>)</span></td>
125+
<td style="margin-left: 30px; text-align: center;" colspan="1">82.3 <span style="color: red;">(<strong>+1.6</strong>)</span></td>
126+
<td style="margin-left: 30px; text-align: center;" colspan="1">82.7 <span style="color: red;">(<strong>+1.9</strong>)</span></td></tr>
127+
</tbody>
128+
</table>
129+
130+
#### 数据集介绍
131+
132+
- **自然语言推断任务** XNLI
133+
XNLI 由 Facebook 和纽约大学的研究者联合构建，旨在评测模型多语言的句子理解能力。目标是判断两个句子的关系（矛盾、中立、蕴含）。[链接](https://github.com/facebookresearch/XNLI)
134+
135+
- **语义匹配任务** LCQMC
136+
LCQMC 是哈尔滨工业大学在自然语言处理国际顶会 COLING2018 构建的问答匹配数据集其目，标是判断两个问题的语义是否相同。[链接](http://aclweb.org/anthology/C18-1166)
137+
138+
- **命名实体识别任务** MSRA-NER
139+
MSRA-NER 数据集由微软亚研院发布，其目标是命名实体识别，是指识别文本中具有特定意义的实体，主要包括人名、地名、机构名等。[链接](http://sighan.cs.uchicago.edu/bakeoff2005/)
140+
141+
- **情感分析任务** ChnSentiCorp
142+
ChnSentiCorp 是中文情感分析数据集，其目标是判断一段话的情感态度。
143+
144+
- **检索式问答任务** nlpcc-dbqa
145+
nlpcc-dbqa是由国际自然语言处理和中文计算会议NLPCC于2016年举办的评测任务，其目标是选择能够回答问题的答案。[链接](http://tcci.ccf.org.cn/conference/2016/dldoc/evagline2.pdf)

0 commit comments

Comments
 (0)