The category of gender bias is where the best performances were found, with 63 for BERT and 73 for RoBERTa. The general purpose MLMs showed significant bias overall, with BERT scoring 57 and RoBERTa scoring 61. This study aims to compare biases in general purpose and medical MLMs with the StereoSet bias assessment tool. Pre-training on large quantities of text also has the potential to transparently embed the cultural and social biases found in the source text into the MLM system.
![collective bias collective bias](http://ww1.prweb.com/prfiles/2012/04/30/9459986/gI_95082_cb-mm-logo.png)
BERT is a powerful and flexible general-purpose MLM system developed using unlabeled text. NLP systems use language models such as Masked Language Models (MLMs) that are pre-trained on large quantities of text such as Wikipedia create representations of language. Therefore, this paper is the first to demonstrate ingroup bias in Wikipedia – a finding that is of practical as well as theoretical relevance as we outline in the discussion. Finally, a third study ruled out that these effects were driven by translations or the raters’ own nationality.
![collective bias collective bias](https://www.makingtimeformommy.com/wp-content/uploads/2011/11/103_3698.jpg)
This procedure allowed for separate analyses for each conflict, which showed considerable variance in the results pattern with a stronger ingroup bias for (1) more recent conflicts and (2) conflicts in which the proportion of ingroup members among the top editors was larger. These findings were replicated and extended in Study 2, which was limited to the lead sections of articles but included considerably more conflicts and many participants instead of a few trained coders.
![collective bias collective bias](https://deeblake.com/images/portfolio/collective-bias/collective-bias-influence.jpg)
Study 1 featured a content coding of translated Wikipedia articles by trained raters, which showed that articles systematically presented the ingroup in a more favourable way (e.g., Argentina in the Spanish article and the United Kingdom in the English article) and, in reverse, the outgroup as more immoral and more responsible for the conflict. Specifically, we compared articles about the same intergroup conflicts (e.g., the Falklands War) in the corresponding language versions of Wikipedia (e.g., the Spanish and English Wikipedia articles about the Falklands War). By examining socially negotiated and publicly accessible Wikipedia articles about intergroup conflicts, we investigated ingroup bias at a collective level. Individuals tend to present their own group (the ingroup) in a systematically more favourable way (ingroup bias).