DEEP LEARNING FOR BIAS DETECTION ON THE ENGLISH WIKIPEDIA

Abstract

On Wikipedia, an online crowdsourced encyclopedia, volunteers enforce the encyclopedia’s editorial policies. Wikipedia’s detailed policy on maintaining a neutral point of view has made the project a popular target for NLP researchers working on bias detection or sentiment analysis. Often, this work focuses on a particular category of bias that Wikipedia identifies; while “weasel words” and “hedges” have both received significant attention, little work has been done on identifying “peacock phrases,” phrases that are overly positive without a verifiable source. In this work, we present a model for identifying peacock phrases that achieves a 0.963 f1 score. We also discuss the general issues inherent in building a dataset from Wikipedia, with this project as a case study. Finally, we demonstrate a way to use Wikipedia’s public infrastructure to host a tool that uses the trained model to give back to the Wikipedia editor community.