Detecting Promotional Content in Wikipedia

34
Detecting Promotional Content in Wikipedia Shruti Bhosale Heath Vinicombe Ray Mooney University of Texas at Austin 1

description

Detecting Promotional Content in Wikipedia. Shruti Bhosale Heath Vinicombe Ray Mooney University of Texas at Austin. Outline. Introduction Related Work Our contribution Evaluation Conclusion. Outline. Introduction Related Work Our contribution Evaluation Conclusion. - PowerPoint PPT Presentation

Transcript of Detecting Promotional Content in Wikipedia

Page 1: Detecting Promotional Content  in Wikipedia

1

Detecting Promotional Content in Wikipedia

Shruti BhosaleHeath Vinicombe

Ray MooneyUniversity of Texas at Austin

Page 2: Detecting Promotional Content  in Wikipedia

2

Outline• Introduction• Related Work• Our contribution• Evaluation• Conclusion

Page 3: Detecting Promotional Content  in Wikipedia

3

Outline• Introduction• Related Work• Our contribution• Evaluation• Conclusion

Page 4: Detecting Promotional Content  in Wikipedia

4

Wikipedia’s Core Policies

Can be edited by anyone

Neutral-point-of-view

Verifiability

Page 5: Detecting Promotional Content  in Wikipedia

5

Quality Control in Wikipedia

• Wikipedia editors and administrators• Clean-up Tags

Page 6: Detecting Promotional Content  in Wikipedia

6

Wikipedia Articles with a promotional tone

• Wikipedia Article on Steve Angello– …Since then, he has exploded onto the house

music scene…– …Steve Angello encompasses enough fame as a

stand alone producer. Add astounding remixes for….his unassailable musical sights have truly made for an intense discography…

Page 7: Detecting Promotional Content  in Wikipedia

7

Wikipedia Articles with a promotional tone

Identified manually and tagged with an Cleanup message by Wikipedia editors

Page 8: Detecting Promotional Content  in Wikipedia

8

Outline• Introduction• Related Work• Our contribution• Evaluation• Conclusion

Page 9: Detecting Promotional Content  in Wikipedia

9

Quality Flaw Prediction in Wikipedia(Anderka et al., 2012)

• Classifiers for ten most frequent quality flaws in Wikipedia

• One of the ten flaws is “Advert” => Written like an advertisement

• Majority of promotional Wikipedia articles

“Advert”

Page 10: Detecting Promotional Content  in Wikipedia

10

Features Used in Classification(Anderka et al., 2012)

Content-based features

Structure features

Network features

Edit history features

Page 11: Detecting Promotional Content  in Wikipedia

11

Outline• Introduction• Related Work• Our Approach–Motivation– Dataset Collection– Features– Classification

• Evaluation• Conclusion

Page 12: Detecting Promotional Content  in Wikipedia

12

Style of Writing

• Our hypothesis - Promotional Articles could contain a distinct style of writing.

• Style of writing could be captured using –PCFG models–n-gram models

Page 13: Detecting Promotional Content  in Wikipedia

13

Our Approach

• Training PCFG models, character trigram models and word trigram models for the sets of promotional and non-promotional Wikipedia articles

• Compute features based on these models

Page 14: Detecting Promotional Content  in Wikipedia

14

Dataset Collection

• Positive Examples:13,000 articles from English Wikipedia’s

category, “Category:All articles with a promotional tone” (April 2013)

• Negative Examples:Randomly selected untagged articles (April 2013)

Page 15: Detecting Promotional Content  in Wikipedia

15

Training and Testing

• 70% of the data is used to train language models for each category of articles

• 30% of the data is used to train and test the classifiers for detecting promotional articles

Page 16: Detecting Promotional Content  in Wikipedia

16

Training N-gram Models

• For each categories of articles, we train – Word Trigram language models and – Character Trigram language models

• We also train a unigram word (BOW) model as a baseline for evaluation

Page 17: Detecting Promotional Content  in Wikipedia

17

N-gram Model Features

• Difference in the probabilities assigned to an article by the positive and the negative class character trigram language models

• Difference in the probabilities assigned to an article by the positive and the negative class word trigram language models

Page 18: Detecting Promotional Content  in Wikipedia

18

Training PCFG models(Raghavan et al., 2010; Harpalani et al., 2011)

Promotional Articles

Non-Promotional

Articles

Promotional Category Treebank

Non-promotional

Category Treebank

Promotional PCFG model

Non-Promotional

Articles PCFG model

PCFG PARSING

PCFG MODEL TRAINING

Page 19: Detecting Promotional Content  in Wikipedia

19

PCFG Model Features

• Calculate probabilities assigned to all sentences of an article by each of the two PCFG models

• Compute Mean, Maximum, Minimum and Standard Deviation of all probabilities, per PCFG model.

• Compute the difference in the values for these statistics assigned by the positive and negative class PCFG models

Page 20: Detecting Promotional Content  in Wikipedia

20

Classification

• LogitBoost with Decision Stumps (Friedman et al., 2000)

• 10-fold cross-validation

Page 21: Detecting Promotional Content  in Wikipedia

21

Outline

• Introduction• Related Work• Our contribution• Evaluation• Conclusion

Page 22: Detecting Promotional Content  in Wikipedia

22

EvaluationFeatures Precision Recall F1 AUC

Bag-of-words Baseline 0.82 0.82 0.82 0.89PCFG 0.88 0.87 0.87 0.94Char. Trigram 0.89 0.89 0.89 0.95Word Trigram 0.86 0.86 0.86 0.93PCFG + Char. Trigram + Word Trigram

0.91 0.92 0.91 0.97

Content and Meta Features (Anderka et al.)

0.87 0.87 0.87 0.94

All Features 0.94 0.94 0.94 0.99

Page 23: Detecting Promotional Content  in Wikipedia

23

EvaluationFeatures Precision Recall F1 AUC

Bag-of-words Baseline 0.82 0.82 0.82 0.89PCFG 0.88 0.87 0.87 0.94Char. Trigram 0.89 0.89 0.89 0.95Word Trigram 0.86 0.86 0.86 0.93PCFG + Char. Trigram + Word Trigram

0.91 0.92 0.91 0.97

Content and Meta Features (Anderka et al.)

0.87 0.87 0.87 0.94

All Features 0.94 0.94 0.94 0.99

Page 24: Detecting Promotional Content  in Wikipedia

24

EvaluationFeatures Precision Recall F1 AUC

Bag-of-words 0.82 0.82 0.82 0.89PCFG 0.88 0.87 0.87 0.94Char. Trigram 0.89 0.89 0.89 0.95Word Trigram 0.86 0.86 0.86 0.93PCFG + Char. Trigram + Word Trigram

0.91 0.92 0.91 0.97

Content and Meta Features (Anderka et al.)

0.87 0.87 0.87 0.94

All Features 0.94 0.94 0.94 0.99

Page 25: Detecting Promotional Content  in Wikipedia

25

EvaluationFeatures Precision Recall F1 AUC

Bag-of-words 0.82 0.82 0.82 0.89PCFG 0.88 0.87 0.87 0.94Char. Trigram 0.89 0.89 0.89 0.95Word Trigram 0.86 0.86 0.86 0.93PCFG + Char. Trigram + Word Trigram

0.91 0.92 0.91 0.97

Content and Meta Features (Anderka et al.)

0.87 0.87 0.87 0.94

All Features 0.94 0.94 0.94 0.99

Page 26: Detecting Promotional Content  in Wikipedia

26

EvaluationFeatures Precision Recall F1 AUC

Bag-of-words 0.82 0.82 0.82 0.89PCFG 0.88 0.87 0.87 0.94Char. Trigram 0.89 0.89 0.89 0.95Word Trigram 0.86 0.86 0.86 0.93PCFG + Char. Trigram + Word Trigram

0.91 0.92 0.91 0.97

Content and Meta Features (Anderka et al.)

0.87 0.87 0.87 0.94

All Features 0.94 0.94 0.94 0.99

Page 27: Detecting Promotional Content  in Wikipedia

27

Top 10 Features(Based on Information Gain)

1. LM char trigram 2. LM word trigram 3. PCFG min 4. PCFG max 5. PCFG mean 6. PCFG std. deviation 7. Number of Characters 8. Number of Words 9. Number of Categories 10. Number of Sentences

Page 28: Detecting Promotional Content  in Wikipedia

28

Average Sentiment Score

• Average sentiment of all words in an article using SentiWordNet (Baccianella et al., 2010)

• Intuitively seems like a discriminative feature• 18th most informative feature• Reinforces our hypothesis that surface level

features are insufficient

Page 29: Detecting Promotional Content  in Wikipedia

29

Conclusion

• Features based on n-gram language models and PCFG models work very well in detecting promotional articles in Wikipedia.

• Main advantages – –Depend on the article’s content only

and not on external meta-data –Perform with high accuracy

Page 30: Detecting Promotional Content  in Wikipedia

30

Questions?

Page 31: Detecting Promotional Content  in Wikipedia

31

Content-based Features• Number of characters, words,

sentences• Avg. Word Length • Avg., min., max. Sentence

Lengths, Ratio of max. to min. sentence lengths

• Ratio of long sentences (>48 words) to Short Sentences (<33 words)

• % of Sentences in the passive voice

• Relative Frequencies of POS tags

• % of sentences beginning with selected POS tags

• % of special phrases (e.g. editorializing terms like ‘without a doubt’, ‘of course’ )

• % of easy words, difficult words, long words and stop words

• Overall Sentiment Score based on SentiWordNet*

* Baccianella et al., 2010

Page 32: Detecting Promotional Content  in Wikipedia

32

Structure Features

• Number of Sections • Number of Images • Number of Categories • Number of Wikipedia Templates used • Number of References, Number of References

per sentence and Number of references per section

Page 33: Detecting Promotional Content  in Wikipedia

33

Wikipedia Network Features

• Number of Internal Wikilinks (to other Wikipedia pages)

• Number of External Links (to other websites) • Number of Backlinks (i.e. Number of wikilinks

from other Wikipedia articles to an article) • Number of Language Links (i.e. Number of

links to the same article in other languages)

Page 34: Detecting Promotional Content  in Wikipedia

34

Edit History Features

• Age of the article • Days since last revision of the article • Number of edits to the article • Number of unique editors • Number of edits made by registered users and by

anonymous IP addresses • Number of edits per editor • Percentage of edits by top 5% of the top

contributors to the article