Research:Topical coverage of Edit Wars
This project is run by the Research Team as part of the Community health initiative.
This project is focused on the the usage of toxic language [1] including specific behavior, such as wiki-hounding [2] The project focuses mainly on the content of user interactions. Our research adopted a complementary, language agnostic approach, which focuses on edit wars and attempts to differentiate between topic-centered and person-centered conflicts.
Introduction
Background
An edit war occurs when editors who disagree about the content of a page repeatedly override each other's contributions. When edit wars occur across multiple topics, they might be an indicator of a personal attack (instead of topic-centered) is occurring. This behavior might be categorized as wikihounding.
Proposal
Edit wars on Wikipedia have been widely studied. We know they can be dectected in a content-agnostic (without analyzing the text) way. We propose studying the topical span of edit wars and characterizing usual and unusual (potentially toxic) behaviors.
The main tasks to develop such models are:
- Define and implement a robust topic model.
- Define a distance metric for topics (eg: Geography is N steps far from Politics, and M steps far from Sports.)
- Generate a representative dataset of edit wars in Wikipedia.
- Detect pairs or groups of users involved in more than X controversies. X will be defined as part of the study.
- Apply an outlier detection mechanism to find potential cases of harassment.
Methodology
- Define a topic model that allows us to measure topic distance between Wikipedia pages.
- Characterize a user's behavior according to the topics that the user edits and the amount of reverts the user commits.
- Compute the probability of a pair of users co-revising a page and the probability that this co-revision is a revert.
- Based on the aforementioned co-revision probability, identify anomalous behaviors that are potentially related to stalking or wikihounding behavior.
Topic model
Mapping pages to topic
- Mapping Wikipedia pages to Wikiprojects can be done with this query: https://quarry.wmflabs.org/query/23214
Note
- We use wikiprojects as proxies for topic assignments.
- Many Wikipedia pages can be considered relevant to more than wikiproject. For example, the page about Diego_Maradona is relevant to Wikipedia:WikiProject_Biography/Sports_and_games, Wikipedia:WikiProject_Football and Wikipedia:WikiProject_Argentina.
Topic distance
- Wikiprojects can be represented as a graph.
- We use this library: https://github.com/wiki-ai/drafttopic/tree/master/drafttopic/utilities).
- Given that each page can belong to more than one wikiproject, we define the distance between two pages as the minimum shortest path [3] among all pairs of nodes on the Wikiprojects graph.
- Example: Given a page X and Y, with X in Wikiprojects a and b and Y in Wikprojects c, and d. We compute the lenght of shortest path between (a,c), (a,d), (b,c) and (b,d), and return the minimum value among these results. In Python:
def distancePages(Graph,page1,page2): """ Graph: is the wikiprojects graphs return -2 if error (page without wikiproject) return -1 if the two pages are the same else return shortest path """ global pagesToWikiprojects results = [] if page1==page2: return -1 try: pages1Projects = pagesToWikiprojects.get(page1,[]) pages2Projects = pagesToWikiprojects.get(page2,[]) except: return -2 for x,y in product(pages1Projects,pages2Projects): try: results.append(nx.shortest_path_length(Graph,x,y)) except:pass if not results: return -2 else: return min(results)
User behavior
Topical coverage
- Topical stability (us): For each user U, we obtain the distance for the user's next revisions.
- For example, if user U makes three revisions, the first one in the topic 'Sports', the second one in the same topic, and the third one in the topic biology (with distance 4 from Sports), the probability that user U will edit with the topical distance of 0, is 2/3, with the topical distance of 4 is 1/3, and the probability is 0 for the all other distances. This metric gives an idea of user stability in terms of topics.
Reverting behavior
- For each user we compute a portion of reverts (within the dataset) compared with her/his total amount of revisions. Considering that our dataset contains only revisions by users who have made 10 or more revisions.
Note: We are just considering the reverts among these users.
Wikiprojects graph
- Using the Wikiproject graph generated from this script[4]
- Pages to Wikiprojects: https://quarry.wmflabs.org/query/23214
- The Wikiproject graph (in networkx format) can be downloaded here: [1]
Terminology / conventions
- We use wikiprojects as proxies for topics.
Results
Reverting behavior
- We found a strong correlation between the number of revisions and the number of reverts. This suggests that reverting is part of the task of active users.
- Also, we found a change in the reverting behavior according to the date of user registration (account age/tenure), as expected, older users tend to be less reverted, and do more reverts.
A detailed analysis of reverting behavior can be found here: [2]
Characterization of user topic-focus
- Following our definition of topical stability, we see that 83.95% of 'next' revisions happen on the same page, and 99.25% on the same topic.
- Moreover, 52.88% of users never jump out of the same topic. However, 41.88% of users jump more than 4 steps at least once.
More details can be found here: [3]
Characterization on topical distance in multipage editwars
- In order to have a notion of how frequent reverts and edit wars are across multiple topics, we considered all pairs of users with U, and V, where U has reverted V more than 2 times, and computed the topical distance between all the pairs of pages reverted.
- Next, we computed the mode -- the most frequent value -- for each pair of users and reported the frequency of those values. As expected, most of wars, 71%, focus on one page and 22% are in the same topic but on a different page. The remaining 7% are cross-topic reverts. This reinforces our intuition that cross-topic edit wars are rare.
Distance | % |
---|---|
-1 | 0.7107 |
0 | 0.2297 |
4 | 0.0143 |
5 | 0.0121 |
2 | 0.0103 |
3 | 0.0074 |
6 | 0.0056 |
7 | 0.0034 |
8 | 0.0027 |
1 | 0.0016 |
9 | 0.0012 |
10 | 0.0007 |
11 | 0.0002 |
12 | 0.0001 |
Outliers
- We found that almost 40% of users jump 4 or more steps in the Wikiprojects graph, making difficult to predict the likelihood of two users of co-editing the same page.
Details about this study can be found here: [4]
Conclusions and Main Outputs
- Based on previous work [5], we have implemented and released a model that allows us to measure topical distance between Wikipedia pages.
- We have found that around 99% of 'next revisions' are done within the same topic.
- We have found that just 7% of edit wars are cross-topic.
- However, users involved in this cross-topic edit wars are generally very active users, making difficult to assume that those wars are due person-centered conflicts.
Resources
Raw data
- We considered all users with more than 10 revisions in the selected period (1/Jan/2017 to 16/Nov/2017)
- The revisions from those users can be obtained from this query: https://quarry.wmflabs.org/query/23424
- A list of possible bots was obtained using this approach: Research_talk:Identifying_bot_accounts
Reverts dataset
- The interactions dataset can be downloaded here: [6]
Format: interactions[user1][user2]:[[pageid,timestamp,deltatime,revision_id_reverted,revision_id_reverting], another revert, etc] user1: user reverting user2: user reverted pageid: page_id timestamp: timestamp when the reverted version was created (done by user2) deltatime: delta time from the reverted version to the reverting revision (done by user1). revision_id_reverted: revision_id_reverted (by user2) revision_id_reverting: revision_id_reverting (by user1)
Wikiprojects graph
- Using the Wikiproject graph generated from this script[4]
- Pages to Wikiprojects: https://quarry.wmflabs.org/query/23214
- The Wikiproject graph (in networkx format) can be downloaded here: [7]
Code
- Find all the code used in this study here:
Future work
- Design a probabilistic model for outlier detection, considering the (un)predictability of two users of co-editing the same page.
- Improve mapping system from pages to Wikiprojects.
- The aforementioned query returns around 22% of pages matching with no wikiproject. However, when we manually reviewed, we found cases where the Wikiproject is not correctly (?) assigned as a category in the Talk pages. For example: Classon_Avenue_(IND_Crosstown_Line) belongs to Wikiproject Trains, but that project is not listed as category.
Timeline
Q1, Q2