Unsupervised Sentiment and Style Transfer from Massive Texts

Abstract

Unsupervised style transfer aims to transfer the intrinsic style of text while preserving its content without parallel datasets. Many sophisticated methods using reinforcement learning and neural networks have been developed to address this problem, however, their performance is not very ideal yet. We observe that given massive unpaired texts, there would exist high-quality sentence pairs that have similar style-independent content but different style words. Inspiring by this observation, in this paper, we propose a simple yet effective method without any neural network. Specifically, we consider both embedding similarity and BLEU score to locate similar sentences of different styles for a pseudo-parallel dataset construction. From this pseudo-parallel dataset, we distill the style words and align them into pairs based on statistical signals. We further refine our pseudo-parallel dataset by ignoring the identified style words during similarity calculation. After the style word pairs converged, we put them together as a lookup table to recognize and replace style words for style transfer. Extensive experiments demonstrate that our method is effective in different style transferring settings, such as sentiment and formality, outperforming state-of-the-art methods.