Specificity-Driven Cascading Approach for Unsupervised Sentiment Modification

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5508–5517, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5508 Specificity-Driven Cascading Approach for Unsupervised Sentiment Modification Pengcheng Yang1, Junyang Lin2, Jingjing Xu2, Jun Xie3, Xu Sun2, Qi Su2 1 Deep Learning Lab, Beijing Institute of Big Data Research, Peking University 2 MOE Key Lab of Computational Linguistics, School of EECS, Peking University 3 Tencent, Beijing, China {yang pc,linjunyang,jingjingxu,xusun,sukia}@pku.edu.cn stiffxie@tencent.com Abstract The task of unsupervised sentiment modifica- tion aims to reverse the sentiment polarity of the input text while preserving its semantic content without any parallel data. Most pre- vious work follows a two-step process. They first separate the content from the original sen- timent, and then directly generate text with the target sentiment only based on the content produced by the first step. However, the sec- ond step bears both the target sentiment addi- tion and content reconstruction, thus resulting in a lack of specific information like proper nouns in the generated text. To remedy this, we propose a specificity-driven cascading ap- proach in this work, which can effectively in- crease the specificity of the generated text and further improve content preservation. The ex- periments show that our approach outperforms competitive baselines by a large margin, which achieves 11% and 38% relative improvements of the overall metric on the Yelp and Amazon datasets, respectively. 1 Introduction The goal of unsupervised sentiment modification is to reverse the sentiment polarity of the input text while preserving its semantic content with- out any parallel data. It not only has a variety of practical applications, e.g., fighting offensive lan- guage on social media (dos Santos et al., 2018), but also serves as an ideal testbed for controllable text generation. However, the input for this task usually contains some specific information like proper nouns. For instance, “Michael” in Table 1 is a human name, which belongs to the specific in- formation. This specific information is important because it is often the subject or object of the sen- tence and the bearer of sentiment. Therefore, it needs to be fully preserved in the process of sen- timent modification. We refer this attribute as the specificity of the output. Input Michael ’s work exceeds my expectations. Shen et al. (2017) Service is terrible. Fu et al. (2018) I was very disappointed. Xu et al. (2018) The girls do horrible work. Table 1: Examples generated by four different previous methods. “Michael” is a proper noun, which belongs to specific information. Most previous work on unsupervised sentiment modification follows a two-step process. They first separate the content from the original senti- ment, either in an implicit (Hu et al., 2017; Shen et al., 2017; Fu et al., 2018; Tsvetkov et al., 2018) or explicit (Xu et al., 2018; Zhang et al., 2018a) way. Then, they directly generate text with the target sentiment only based on the content pro- duced by the first …

Content truncated to 500 words