Bias detection in Wikipedia articles. A study on Polish and English Datasets.

Abstract

Nowadays, an almost unlimited number of information sources and polarized media hamper our ability to distinguish between biased and neutral speech. It creates a need for an automatic tool that could comprehend human language and assess whether presented information is conveyed without any editorial bias. In this work, we introduce models able to detect bias on a sentence level. As other authors before, we continue to utilize Wikipedia revision history to collect examples of biased and unbiased speech. For the first time, this work introduces a corpus of labeled subjective and neutral sentences in the Polish language. Furthermore, we compare the performance of LSTM and BERT models trained on English and Polish sentences. As part of the findings, we present cues for subjectivity that were detected during the analysis. We also present a new dataset WNC-pl a Polish corpus of biased and unbiased sentences.