Evaluating Style Transfer for Text

TL;DR

This paper addresses the lack of standard evaluation methods in text style transfer, proposing improved metrics and best practices validated on a Yelp dataset to enhance reliability and comparability.

Contribution

It introduces new automated evaluation metrics for style transfer, correlates them with human judgment, and provides guidelines for assessing tradeoffs between style transfer aspects.

Findings

Automated metrics correlate better with human judgments.

Models show tradeoffs between style transfer quality and content preservation.

Software tools for evaluation are publicly released.

Abstract

Research in the area of style transfer for text is currently bottlenecked by a lack of standard evaluation practices. This paper aims to alleviate this issue by experimentally identifying best practices with a Yelp sentiment dataset. We specify three 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 propose a set of metrics for automated evaluation and demonstrate that they are more strongly correlated and in agreement with human judgment: direction-corrected Earth Mover's Distance, Word Mover's Distance on style-masked texts, and adversarial classification for the respective aspects. We also show that the three examined models exhibit tradeoffs between aspects of interest, demonstrating the importance of evaluating style transfer models at specific points of their tradeoff plots. We release software with our evaluation metrics to facilitate research.