ChineseNLP

中文关系提取 (Relation Extraction)

背景

给定两个实体 (entity)，识别它们的关系并对关系进行分类。

示例

输入:

[李晓华]和她的丈夫[王大牛]前日一起去[英国]旅行了。

输出:

(entity1: 李晓华, entity2: 王大牛, relation: 夫妻) 

标准评价指标

精确率 (Precision), 召回率 (Recall), F1

输入:

[李晓华]和她的丈夫[王大牛]前日一起去[英国]旅行了。

参考答案 (Reference):

(entity1: 李晓华, entity2: 王大牛, relation: 夫妻) 
(entity1: 李晓华, entity2: 英国, relation: Other) 
(entity1: 王大牛, entity2: 李晓华, relation: 夫妻) 
(entity1: 王大牛, entity2: 英国, relation: Other) 
(entity1: 英国, entity2: 李晓华, relation: Other) 
(entity1: 英国, entity2: 王大牛, relation: Other)

系统输出结果:

(entity1: 李晓华, entity2: 王大牛, relation: 夫妻) 
(entity1: 李晓华, entity2: 英国, relation: 夫妻) 
(entity1: 王大牛, entity2: 李晓华, relation: 夫妻) 
(entity1: 王大牛, entity2: 英国, relation: Other) 
(entity1: 英国, entity2: 李晓华, relation: Other) 
(entity1: 英国, entity2: 王大牛, relation: Other) 

评价指标:

Precision = 2 / 3 = 0.66
Recall = 2 / 2 = 1.0

ACE 2005 Relation Extraction.

• 数据链接
• 综述论文
• 不同的论文以不同的方式分割训练集和测试集。

ACE 2005包含6种关系类型和18种子关系类型 (subtypes), 具体类型如下所示:

Type	Subtypes
ART (artifact)	User-Owner-Inventor-Manufacturer
GEN-AFF (Gen-affiliation)	Citizen-Resident-Religion-Ethnicity, Org-Location
* METONYMY	none
ORG-AFF (Org-affiliation)	Employment, Founder, Ownership, Student-Alum, Sports-Affiliation, Investor-Shareholder, Membership
P ART-WHOLE (part-whole)	Artifact, Geographical, Subsidiary
* PER-SOC (person-social)	Business, Family, Lasting-Personal
* PHYS (physical)	Located, Near

结果

	F1 (6 relation types)	F1 (18 relation types)	Train/Test split
Zhang et al. (2018)	87.87	83.40	80% / 20%
Li et. al. (2019)	-	78.17	75% / 25%
Chen et al. 2014	90.35	75.44	-

相关资源

ACE 2005 Chinese Corpus	字符数量	文件数量
Newswire	121797	238
Broadcast news	120513	298
Web blogs	65681	97
Total	307991	633

Chinese-Literature-NER-RE-Dataset .

• Data
• 论文
• 清晰的train/dev/test数据集划分
• 包括了9种relation(Located, Part-Whole, Family, General-Special, Social, Ownership, Use, Create, Near)

结果

	F1
Li et. al. (2019)	65.61
Zhang et. al. (2020)	63.13
Xu et. al. (2020)	57.43

---

建议? 修改? 请发邮件到 chinesenlp.xyz@gmail.com