Adversarial Demotion of Bias in Natural Language Generation

Abstract

Natural Language Generation models have been a critical area of research in application-oriented artificial intelligence tasks, such as dialogue systems, machine translation, and question answering. The next crucial step in this direction is to ensure quality of generated text. This work proposes a novel method based on adversarial training to mitigate gender bias in generation systems, and can be extended to remove any unwanted characteristics in the generated text.