Learning to Flip the Sentiment of Reviews from Non-Parallel Corpora

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6311–6316, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6311 Learning to Flip the Sentiment of Reviews from Non-Parallel Corpora Canasai Kruengkrai Yahoo Japan Corporation ckruengk@yahoo-corp.jp Abstract Flipping sentiment while preserving sentence meaning is challenging because parallel sen- tences with the same content but different sen- timent polarities are not always available for model learning. We introduce a method for acquiring imperfectly aligned sentences from non-parallel corpora and propose a model that learns to minimize the sentiment and con- tent losses in a fully end-to-end manner. Our model is simple and offers well-balanced re- sults across two domains: Yelp restaurant and Amazon product reviews.1 1 Introduction Text style transfer is the process of editing a sen- tence to modify specific attributes (e.g., style, sen- timent, and tense) while preserving its attribute- independent content (Shen et al., 2017; Fu et al., 2018; Li et al., 2018). In this work, we are particularly interested in sentiment modification. This task is challenging because parallel data (i.e., aligned sentences with the same content but dif- ferent sentiment polarities) are not always avail- able and they are costly to annotate. Major exist- ing solutions that exploit non-parallel data include training discriminators to guide sentence genera- tion (Hu et al., 2017; Shen et al., 2017; Fu et al., 2018), retrieving similar sentences from another corpus and then editing them (Guu et al., 2018; Li et al., 2018), and applying back-translation (Prab- humoye et al., 2018; Zhang et al., 2018b; Lample et al., 2019). Achieving both accurate sentiment modification and high content preservation can be difficult. Re- cently, Li et al. (2018)’s framework has shown promising results. They delete sentiment-specific phrases (e.g., “works great”) from a given sen- 1The code for reproducibility will be available on the au- thor’s website. (a) the service was great too → the service was n’t too great (b) this one is by far the worst → this one is the best by far Figure 1: The retrieved sentences (right) from the op- posite corpora given the query sentences (left). tence, retrieve new phrases associated with the tar- get sentiment (e.g, “barely used”), and recombine them to generate an output. However, their frame- work relies on separate modules and assumes that clear boundaries exist between the sentiment- specific phrases and the content, which is not al- ways the case. In this work, our objective is to develop a model that can be trained end-to-end from sentence pairs acquired from non-parallel corpora. Our approach is based on the observation that shared tokens between similar sentences in unaligned corpora convey the content we wish to preserve, while non-shared tokens more or less reflect sentiment change. Figure 1 shows the examples of the re- trieved sentences from the opposite corpora given the query sentences. Although these sentence pairs are not perfectly aligned, they provide a use- ful signal of …

Content truncated to 500 words