topic_classification.md

Chinese Text Classification

Background

Text classification assigns tags or categories to text according to its topical content, typically training on labeled documents. Topics are sometimes broad and akin to genre (news, sports, arts) but sometimes as fine-grained as hashtags.

Example input/output

Input:

[国足]有信心了 中国国奥队取得热身赛三连胜

Output:

Sports

Standard Metrics

• Accuracy: the percentage of correctly classified samples.

THUCNews.

Sina News RSS subscription channel data from 2005 to 2011, which contains 74 million news documents (2.19 GB), 14 topics, all in UTF-8 plain text format.

• Website

Source	# Classes	Size(sentences)
THUCNews	14	740,000

Metrics

• Accuracy

Results

	Accuracy
J. Chen, C. Cao, X. Jiang	98.7%
Y. Song	97.56%
W. Liu, P. Zhou, et al	96.71%
S. Xin	96.04%
Sun, Baohua, et al	94.85%

SogouCS.

Sohu News from June to July 2012 in 18 channels.

• Website

Source	# Classes	Size(sentences)
Sougou news dataset	5	86,597

Metrics

• Accuracy

Results

	Error rate
Chung, Tonglee, et al	3.37%

Resources

Dataset	Classes	Train(samples size)
Sougou news dataset	5	490,717

Fudan corpus.

contains 9804 documents of long sentences and paragraphs in 20 categories.

Source	# Classes	Size(sentences)
Fudan corpus	5	1836

Metrics

• Accuracy

Results

	Accuracy
Sun, Baohua, et al	97.8%
Meng et al, 2019	96.3%

Resources

Source	# Classes	Size(sentences)
Fudan corpus	5	4284

Ifeng.

First paragraphs of Chinese news articles from 2006-2016 were evenly split into 5 news channels.

• Github link

Source	# Classes	Size(sentences)
Ifeng	5	50,000

Metrics

• Accuracy

Results

	Accuracy
Meng et al, 2019	85.8%
Sun, Baohua, et al	84.4%
Zhang and Lecun 2017	83.7%

Resources

Dataset	Classes	Train(samples size)
Ifeng	5	800,000

Chinanews.

Chinese news articles from 2008- 2016 were evenly split into 7 news channels, removing duplicates.

• Github link

Source	# Classes	Size(sentences)
Chinanews	7	112,000

Metrics

• Accuracy

Results

	Accuracy
Sun, Baohua, et al	92.0%
Meng et al, 2019	91.9%
Zhang and Lecun 2017	90.9%

Resources

Dataset	Classes	Train(samples size)
China news	7	1,400,000

---

Suggestions? Changes? Please send email to chinesenlp.xyz@gmail.com