Fixing racial discrimination through analytics on online platforms: A neural machine translation approach

Abstract

Few issues have fostered such controversy as racial discrimination, especially against African American. While online crowdfunding is expected to democratize the opportunity of starting new businesses, recent empirical evidence suggests that it is not the case for racial minority groups. This study proposes a novel data-analytic approach leveraging state-of-the-art deep learning techniques to design content for reducing racial bias. Drawing upon economic theories of discrimination, we employ a neural machine translation approach to augment the quality of signals, by which undesired textual content is converted into success-style content which favors toward minorities, while preserving the meaning of content. In doing so, we use adversarial networks to separate content representations from style representations. Then, we train the neural sequence-to-sequence model whose decoder takes both content and style representations to generate the output sentence adapted to the desired success-style. Using a large-scale data including 74,500 crowdfunding projects with over 1.7 million sentences, we present the preliminary result of style transfer to success-style. This study demonstrates how machine learning-based analytics can be integrated into a design science paradigm to address societal challenges.