Explainable Supervised Models for Bias Mitigation in Hate Speech Detection: African American English

Automated hate speech detection systems have great potential in the realm of social media but have seen their success limited in practice due to their unreliability and inexplicability. Two major obstacles they have yet to overcome is their tendency to underperform when faced with non-standard forms of English and a general lack of transparency in their decision-making process. These issues result in users of low-resource languages (those that have limited data available for training) such as African-American English being flagged for hate speech at a higher rate than users of mainstream English. The cause of the performance disparity in these systems has been traced to multiple issues including social biases held by the human annotators employed to label training data, training data class imbalances caused by insufficient instances of low-resource language text and a lack of sensitivity of machine learning (ML) models to contextual nuances between dialects. All these issues are further compounded by the ‘black-box’ nature of the complex deep learning models used in these systems. This research proposes to consolidate seemingly unrelated recently developed methods in machine learning to resolve the issue of bias and lack of transparency in automated hate speech detection. The research will utilize synthetic text generation to produce a theoretically unlimited amount of low-resource language text training data, machine translation to overcome annotation conflicts caused by contextual nuances between dialects and explainable ML (including integrated gradients and instance-level explanation by simplification). We will attempt to show that when repurposed and integrated into a single system these methods can both significantly reduce bias in hate speech detection tasks whilst also providing interpretable explanations of the system’s decision-making process.