Publication

Abstract : Social media has become an essential facet of modern society and is quickly becoming indispensable. The ubiquity of social media, while having its benefits, also opens up people to a world of toxicity that is often left unregulated. The word toxic has become synonymous with online hate speech, internet trolling, and sometimes outrage culture. To tackle these hate comments, Google and Jigsaw together have developed the Perspective API, which detects toxicity and can be used for moderation. Through this study, we highlight that Perspective API poses a discrepancy in determining the toxicity of sentences that are semantically the same but with different arrangements of words. We prove that changing a sentence from active voice to passive voice and sentence reordering while maintaining the semantic meaning leads to a substantial difference in the toxicity of that sentence. Furthermore, we discuss possible methods to overcome this bias and help toxicity prediction models become capable of predicting more consistent results across different sentences with the same semantics terms.