=Paper=
{{Paper
|id=Vol-1177/CLEF2011wn-PAN-Aksit2011
|storemode=property
|title=An Empirical Research: “Wikipedia Vandalism Detection using VandalSense 2.0” Notebook for PAN at CLEF 2011
|pdfUrl=https://ceur-ws.org/Vol-1177/CLEF2011wn-PAN-Aksit2011.pdf
|volume=Vol-1177
}}
==An Empirical Research: “Wikipedia Vandalism Detection using VandalSense 2.0” Notebook for PAN at CLEF 2011==
https://ceur-ws.org/Vol-1177/CLEF2011wn-PAN-Aksit2011.pdf
An Empirical Research: “Wikipedia
Vandalism Detection using VandalSense 2.0”
Notebook for PAN at CLEF 2011
F. Gediz Aksit
Maastricht University
g.aksit@student.maastrichtuniversity.nl
Abstract. Wikipedia despite having a very small budget has been among the
top ten most visited websites for over half a decade. Being this visible also
generated the problem of ill intended people modifying Wikipedia in a
destructive manner. VandalSense is an experimental tool programmed by F.
Gediz Aksit to automatically identify vandalism on Wikipedia through the use
of machine learning and text mining as well as the use of years of personal
experience. VandalSense is not intended to replace traditional recent changes
patrolling and instead it is intended to be a tool to compliment it.
1. Introduction
Wikipedia, the free encyclopedia has been under the constant attack of vandalism.
Vandalism is generally quickly removed, “but in one particularly well-publicized
incident, false information was introduced into the biography of American political
figure John Seigenthaler and remained undetected for four months.”1
Furthermore Wikipedia is no longer just an online encyclopedia that one can
simply check the previous revision on vandalized pages. WikiReader2, Wikibrowse3
for OLPC (One Laptop Per Child), WikiMiner4 are among the many offline uses of
the encyclopedia with no access to the history page. Furthermore vandalism in offline
editions may remain unchecked for years in these offline devices which in some cases
are the only means of information where internet or even books are unavailable. This
only elevates the importance of the elimination of Vandalism from Wikipedia.
2. Approach
My approach was based on the initial analysis of the problem itself.
Wikipedia receives millions of edits every day and these edits can be classified
initially as edits from logged in users (Accounts) and edits from anonymous users
(IPs) which will be referenced respectively in the rest of this document. German (de)
and Spanish (es) editions of Wikipedia were analyzed using the PAN 2011 corpus for
both these language editions. The two language editions of Wikipedia will be
referenced by their respective two letter shortcut for the rest of this document. The
distinction between the types of contributors is visible in the test and training corpus
as well.
Edit behavior between account and IP type contributors differs considerably.
By very nature accounts are intended to have a more established contribution history
while IP edits are intended to be quick edits often months apart. In the light of this
distinction it is possible to distinguish edits classified as vandalism on the training
corpus based on their account types before the actual training phase.
As visible from the pie charts training corpus for both language editions have
about 30% of the revisions that are vandalism. As also clearly visible of the pie charts
vast majority of the vandalism comes from IP edits and only a minority comes from
logged in users. De wiki training corpus has about 71% (270 out of 379) of the total
IP edits that are vandalism. Same corpus also has about 4% (24 out of 596) of the
total account edits that are vandalism. Es wiki training corpus has about 56% (292 out
of 521) of the total IP edits that are vandalism. Same corpus also has about 3% (15
out of 462) of the total account edits that are vandalism. As a result as visible in the
pie charts the throwaway IP accounts are more prone to vandalism than logged in
users. Because edit behaviors of accounts and IPs are very different and because the
small potential of training from so few edits by accounts edits by accounts were
ignored completely.
Based on the statistics from the training corpus, de wiki is predicted to have
about 3,400 out of 84,114 revisions by accounts while 19,800 out of 27,840 revisions
by IPs are predicted to be vandalism. Es wiki on the other hand is predicted to have
about 500 out of 15,577 revisions by accounts while 10,313 out of 18,417 revisions
by IPs are to be vandalism.
3. Features
The creative nature of vandalism makes its identification more than tricky.
Keywords such as vulgar words and keywords such as Nazi and Jew are common in
vandalism edits however same words – even vulgar words – are welcome on their
respective articles.
Extracting features had proven to be a challenge as each case of vandalism
observed differs from each other significantly. A diff algorithm by M. Hertel5 was
used to compare revisions. The diff result is then stripped of punctuation and then it is
stemmed using snowball libraries.6 Wiki markup was not stripped as the markup itself
is used differently by vandals where vandals typically use the wiki markup in a less
familiar and sloppy way. This project intends to detect three types of vandalism.
• Blanking: This term is used to define edits that result in the mass removal of
content. Edits of this nature typically removes one or more paragraphs. While
there are legitimate reasons for such mass removal information such as the
removal of copyrighted material, such edits are almost always conducted by
accounts rather than IPs.
• Gibberish: This term is used to define edits that result in the mass inclusion of
content. Edits of this nature typically includes one or more paragraphs. While
there may be legitimate reasons for such mass inclusion of information such as
copy paste of material from freely licensed sources, such edits are almost always
conducted by accounts (more so flagged bots) rather than IPs.
• Sneaky: This term is used to define edits that result in the addition of a small
amount of content that is intended to change the meaning of a few sentences or
add a shot well structured messages to avoid detection. Such an edit was used
with Seigenthaler biography controversy mentioned in the introduction section.
Sneaky vandalism revisions typically contain similar keywords which include but
not limited to profanity.
4. Classification
Figure 1: Decision tree structure
User type as well as the three features discussed above were observed and
were used for classification purposes. A decision tree structure was implemented
with an order that is intended to identify the more obvious kind of vandalism first
without involving an unnecessary and time consuming word analysis reducing
resource usage.
Equation 1: Entropy and Entropy gain calculation
All of the three classifiers based on edit patterns require threshold values to
operate. These values are determined through the use of statistical entropy and
entropy gain which is also a key feature of the ID3 algorithm.7 The tree was generated
by hand due to the unstructured nature of the data as well as the minimal availability
of meta data. Features are particularly hard to find as vandalism has far too many
flavors with far too many different patterns. The same patterns can also be observed
with regular edits.
Vandalism detection is achieved by initial count of the regular and vandalism
revisions. These numbers are then compared to the revisions to threshold values for
the latter three classifications. Byte values are looped while seeking a higher entropy
gain. This strategy however has a flaw. Because the majority of the revisions are
regular edits, entropy gain converges towards misidentifying vandalism revisions as
regular edits. To circumvent this entropy gain is capped at a reasonable amount. The
value of .6 was observed to be the more reasonable amount for entropy gain.
• User Type: User type classifier identifies if an edit is by an IP or an account.
Account edits are flagged as regular edits for the reasons discussed in the
approach section.
• Deletion: Deletion classifier calculates the deletion amount based on how many
bytes of information was removed even if the information is replaced by some
other content. The threshold value N is calculated for this classifier.
• Addition: Addition classifier conversely calculates the inclusion amount based
on how many bytes of information was included even if the information replaces
some other existing content. The threshold value M is calculated for this
classifier.
• Word Score: This classifier calculates word score based on the frequency of the
words appearance in regular and vandalism revisions. The effectiveness of word
score classifier is limited as the training set is not a fair representation of the
entire respective languages.
Equation 2: Word score calculation
Although the limited size of the training set makes it difficult to have a
reasonable understanding of the language processed, vandalism identification through
statistical frequency of words that appear in vandalism and regular edits was
attempted. This produced almost random results. It was observed that words common
in vandalism revisions can also be common in regular edits. For instance stop words
are as expected common in both vandalism and regular edits. Instead of stripping stop
words and spending time to manually or algorithmically identify words common on
both types of edits, entropy gain was employed to weight the good and bad word
scores for vandalism identification. A weight value was used to shift the weight
towards good or bad edit score depending on the targeted entropy gain value. For
word score, each words positive score (its frequency in good revisions) and bad score
(its frequency in bad revisions) are individually calculated, weighted and then the
weighted bad score is subtracted from the good score. If the concluding scores
computation is negative in value, that revision is considered to be vandalism. This
approach eliminates problems stemming from stop words as well.
5. Visualization
An ASP.Net web application was developed on top of the algorithms used to
generate the submission to PAN. Intention behind this is to expand the project for
human use to expand the training set based on human submission. The web
application does not have a submission feature implemented yet. Visualization aspect
of the project is essential as the intended use of the entire project is to assist recent
changes patrol by weighting revisions instead of making edits to the wiki directly.
The web application end of the project is also intended to better analyze the
inner workings of the project particularly that of the word score calculations. The web
application is also capable of analyzing the live recent changes processing up to last
500 revisions (restriction due to Wikipedia’s API limit). This web application can be
accessed through http://eva.no-ip.biz/VandalSenseWeb/ and will be maintained as
long as resources allow it. The three fields (upper diff, lower diff, word point) are
actually the ID3 gain values. These fields exist to allow fine tuning. The change field
displays the threshold values including the individual values for words before they get
weighted.
A word cloud representation was implemented to show the frequency of
words in the added revision. The colors represent if the word is positive (green),
negative (red) or neutral/unknown (grey). The color itself isn’t weighted. An
interesting feature of the web application is the inclusion of Google Earth API.8 The
API allows the search of vandalism revisions which may help identify edits from
unrelated IP ranges that are geographically nearby which could in return be used to
identify more vandalism patterns expanding the training set.
6. Conclusions
The approach taken intended to avoid false positives as much as possible.
Only a minority of the edits to Wikipedia is vandalism and false positives would only
contribute to the problem. Both runs on both wikis (de and es) had a recall of 25%
with a precision of about 60% for de wiki and 75% for es wiki. Based on the test set
and training set statistics as discussed in the approach section es wiki receives a
greater percentage of IP edits than of account edits in comparison to de wiki. The
increase in accuracy of Spanish Wikipedia is probably due to training set of es wiki
having a greater number of revisions for IP edits than of the training set of de wiki.
7. References
1
http://www.usatoday.com/news/opinion/editorials/2005-11-29-wikipedia-edit_x.htm, [Online
accessed: 23 June 2011]
2
http://www.thewikireader.com/, [Online accessed: 23 June 2011]
3
http://wiki.laptop.org/go/Wikibrowse, [Online accessed: 23 June 2011]
4
http://meta.wikimedia.org/w/index.php?title=WikiMiner&oldid=503664, [Online accessed:
23 June 2011]
5
http://www.mathertel.de/Diff/Default.aspx, [Online accessed: 23 June 2011]
6
http://snowball.tartarus.org/, [Online accessed: 23 June 2011]
7
http://www.cise.ufl.edu/~ddd/cap6635/Fall-97/Short-papers/2.htm, [Online accessed: 23
June 2011]
8
http://code.google.com/apis/earth/, [Online accessed: 23 June 2011]