Evaluating style modification in text

Abstract

In this thesis, we identify best practices for evaluating style modification, or style transfer, for text. Research of style transfer is bottlenecked by a lack of standard evaluation practices. We define three key aspects of interest (style transfer intensity, content preservation, and naturalness) and show how to obtain more reliable measures of them from human evaluation than in previous work. We also demonstrate stronger correlation between human judgment and a new set of automated metrics: the Wasserstein distance, word mover's distance on texts with style masked out, and adversarial classification for the respective aspects. Lastly, we illustrate aspect tradeoff curves for three state-of-the-art style transfer models to highlight the importance of evaluating style transfer models at specific points on the curves. This can enable direct comparison of the models, facilitating future research in style transfer.