Research talk:Wikipedia Knowledge Integrity Risk Observatory

Content controversiality[edit]

Why is this listed as indicative of a higher risk of misinformation? There has been some research indicating the opposite ("higher political polarization was associated with higher article quality"). Regards, HaeB (talk) 20:48, 22 July 2021 (UTC)[reply]

Thanks @HaeB for your comment (that should have been answered much earlier). It is true that there are studies that showed that "higher political polarization was associated with higher article quality". In fact, other studies in which I participated already indicated that the claim of polarization as a harmful effect in social media is not applicable to Wikipedia because of its design and the characteristics of its community. However, I also think it is important to differentiate between controversiality and polarization, two related but different concepts. The notion of controversiality was given by the claim that "controversial pages are articles that regularly become biased and need to be fixed or that suffer from additions of opinions that do not conform to the Wikipedia neutral point of view". That said, content neutrality, as a category, might capture this phenomenon better than content controversiality, so the (non-definitive) taxonomy is sensitive to change/improvement. As a matter of fact, as a result of a review we are conducting in the area of knowledge integrity, I am finding that knowledge integrity in Wikipedia might be better characterized through content verifiability, quality and neutrality (rather than controversiality). Pablo (WMF) (talk) 08:57, 7 December 2022 (UTC)[reply]

Paid advocacy and conflict of interest editing[edit]

It's rather surprising that this is not covered, considering that 1) the introduction correctly highlights "atroturfing" among the main causes of misinformation problems for web platforms in general, 2) the issue has been discussed as a major concern among the editing community for a long time (see e.g. the hundreds of articles published over the year by the Wikipedia Signpost, recently in form of a regular "Disinformation Report" rubric), 3) it's one of the few content areas where the Wikimedia Foundation felt compelled to take action already, e.g. by changing its terms of use and taking legal action against some players.

Could you talk a bit about what research insights let you conclude that this is not a disinformation risk worth monitoring for Wikipedia?

Regards, HaeB (talk) 02:24, 23 July 2021 (UTC)[reply]

Thanks @HaeB for this comment too. Paid advocacy and conflict of interest editing are indeed threats to knowledge integrity that should be included, ideally within a content neutrality category of the taxonomy. Recent work has proposed methods for detecting biased language is English Wikipedia. However, the risk observatory should rely by design on language-agnostic approaches. That is why I am interested in approaches based on features such as “the percentage of edits made by a user that are less than 10 bytes, since undisclosed paid editors try to become autoconfirmed users before creating a promotional article”. Given all the review work you have done so far, any recommendation in this regard is more than welcome :) Pablo (WMF) (talk) 08:58, 7 December 2022 (UTC)[reply]

HaeB, fine point -- what other tools exist for tracking and responding to that? Pablo, please don't rely exclusively on language-agnostic approaches. That will never be as good as models that include language and meaning; language-agnostic measures are also, looking at the history of the tools arms race for spam, easier to game. Happy to discuss specifics. –SJ talk 18:44, 3 October 2023 (UTC)[reply]

Sj, thanks also for your comment. I agree that language-agnostic approaches will never be as good as models that include linguistic features. For example, new ML models have recently been deployed to predict the probability that a revision will be reverted, with a multilingual approach and a language-agnostic approach. The multilingual one uses mBert (available in 47 languages) and provides better results for IP edits. The language-agnostic approach takes no advantage of the possibilities offered by recent NLP resources and shows lower accuracy with IP edits, but instead provides support to all language editions of Wikipedia. In my opinion, the trade-off between model performance and language coverage is debatable, there is no perfect solution. However, in order to promote knowledge equity, I think it is very positive that not only dozens of languages benefit from these efforts (being also the ones that usually have more and better resources). That said, I'd love to know more about the details of the history of anti-spam tools :) Pablo (WMF) (talk) 08:27, 4 October 2023 (UTC)[reply]

3 Of course our base condition should be models that can be used anywhere! I just wouldn't want "language agnostic" and "multilingual" to be seen as mutually exclusive :) I hope multilingual models can include a combination of language-agnostic features, language-specific but meaning-agnostic features (related to language structure, byte encoding, &c), meaning-specific features, &c. As all languages would benefit from linguistic analysis, we could support all languages for instance by being aware of language clusters and finding the closest match; by having a clear way to build and curate lists of patterns to block + allow. All languages would benefit from a shared blocklist of scam + spam domains, &c. For spam it proved helpful for networks to share data about new attacks, with things like fuzzy pattern-matching to flag astroturfing.

I'd be interested to know why the multilingual revert risk model is only in 50 languages while mBert is in 100. Since mBert itself drew on the top 100 wikipedias by size, would it be helpful to retrain it today with more languages? Which ones are most significant (new language-families, largest population of primary + secondary speakers, &c)? –SJ talk 17:36, 4 October 2023 (UTC)[reply]

Interesting! I also see it very important to have clear ways to build and curate lists of patterns to block+allow, including shared block lists of scam + spam domains (e.g., COIBot benefits from shared data). In a recent study we found imbalances and disagreements between perennial source lists from different language editions (which was proven very effective for the English Wikipedia in a previous study). On the one hand, I am enthusiastic about participatory approaches to address these challenges. On the other hand, I am also aware that not all language communities have the same resources to engage in participatory processes and I am less enthusiastic about extracting patterns from annotation samples in which large communities are overrepresented. For this reason, for list curation, I am starting to see great potential in transitioning from explicit annotations (high quality data but very expensive to obtain) to implicit annotations (e.g., patterns inferred from reverted activity).

Regarding the language coverage of the multilingual revert risk model, @Diego (WMF) is the model owner, so he can provide you with more details. Pablo (WMF) (talk) 12:12, 5 October 2023 (UTC)[reply]