Wikipedia is a collective intelligence platform that helps contributors to collaborate efficiently for creating and disseminating knowledge and content. A key guiding principle of Wikipedia is to maintain a neutral point of view (NPOV), which can be challenging for new contributors and experienced editors alike. Hence, several previous studies have proposed automated systems to detect biased statements on Wikipedia with mixed results. In this paper, we investigate the potential of cross-domain pre-training to learn bias features from multiple sources, including Wikipedia, news articles, and ideological statements from political figures in an effort to learn richer cross-domain indicators of bias that may be missed by existing methods. Concretely, we study the effectiveness of bias detection via cross-domain pre-training of deep transformer models. We find that the cross-domain bias classifier with continually pre-trained RoBERTa model achieves a precision of 89% with an F1 score of 87%, and can detect subtle forms of bias with higher accuracy than existing methods.
Paper
Full text
Improving Linguistic Bias Detection in Wikipedia using Cross-Domain Adaptive Pre-Training
Semantic Scholar · Linguistics · 2022
Abstract
Wikipedia is a collective intelligence platform that helps contributors to collaborate efficiently for creating and disseminating knowledge and content. A key guiding principle of Wikipedia is to maintain a neutral point of view (NPOV), which can be challenging for new contributors and experienced editors alike. Hence, several previous studies have proposed automated systems to detect biased statements on Wikipedia with mixed results. In this paper, we investigate the potential of cross-domain pre-training to learn bias features from multiple sources, including Wikipedia, news articles, and ideological statements from political figures in an effort to learn richer cross-domain indicators of bias that may be missed by existing methods. Concretely, we study the effectiveness of bias detection via cross-domain pre-training of deep transformer models. We find that the cross-domain bias classifier with continually pre-trained RoBERTa model achieves a precision of 89% with an F1 score of 87%, and can detect subtle forms of bias with higher accuracy than existing methods.