Human biases in government algorithms

Machine learning is used in data-driven decision making and in governmental application of algorithms. Along with its many benefits, literature shows that machine learning solutions can introduce human bias to their results that they from training data. In this paper the authors study sentiment analysis algorithms to see if they show human bias, by interchanging lists of names associated with gender, race and political orientation in a synthetic neutral template sentence and then observing how the resulting sentiment values change. As a result, the authors find that names as subjects and objects in English language texts alone do not introduce human bias to the sentences' sentiment values. The findings also suggest that pseudonymization might be a better solution to the anonymization of social media postings for data protection compliance purposes than research so far suggested, owing to the general lack of sentiment distortion of the method introduced in this paper.