Create 2019/WikiNetBias: topic polarisation on Wikipedia graph and its effect on users: Submissions:2021/News On Wiki, Phase 2 complete

This submission has been accepted for WikiConference North America 2019.

Wikipedia is one of the most widely used sources of online knowledge and information. The encyclopaedia is the result of human cooperation, thus it is likely to inherit human bias that, on our perspective, arises mainly in two ways:

• Free-to-air bias: article’s content is not neutral but skewed towards a perspective.
• Structural bias:  bias through the hyperlink network of articles. In particular, editors inject bias adding hyperlinks, generally according to their knowledge map that reflects a particular standpoint.

The review process set up by Wikimedia is powerful concerning free-to-air bias since it is directly observable. In contrast, structural bias is hidden and difficult to detect by reviewing a single article because it does not give a comprehensive view of the article in the context of the induced network. 

Generally, a user accesses Wikipedia to broaden her knowledge about a specific topic. Each topic can appear on the encyclopaedia in different shapes: we find topic-self-contained pages or set of articles all referring to the same topic. Depending on the topic representation, we are more or less able to control the possible presence of bias. In the case of topic-self-contained articles, the review process gives guarantees about the neutral perspective a generic user is exposed to. When the topic is represented by many articles, there is not an ongoing process that monitors whether the topic-induced-network is unbiased and thus allows users to get a neutral overview of the field.  Specifically, we believe that it is crucial to monitor possible polarization in the Wikipedia induced hyperlink network associated with specific topics. Roughly speaking, we define a topic-induced-network polarized when a user is likely to be trapped within a knowledge chamber presenting a unique viewpoint about the topic. In particular, the question we want to address is whether users' navigation paths are biased toward one of the possible perspectives. To give an example: we want people accessing pages on US politics to be exposed to both liberal and conservative articles, through a topic-induced-network that is not biased toward a narrow set of opinions.

Recognising one of the strongest points of Wikipedia that of allowing users to easily navigate the encyclopaedia, more specifically a topic, through its hyperlinks structure, for the sake of platform’s reliability, it is of utmost importance to be sure that a user has the same chance of being exposed to knowledge that expresses different viewpoints concerning the given topic. With this presentation we would like to shed light and propose the study of structural bias since, if present, would become source of unreliability that harms and influence a lot the user knowledge building process. In the talk we will present more in detail the idea of structural bias, its source and the intuition behind algorithmic solution to the problem.