Improving Fine-grained Text Sentiment Transfer for Diverse Review Generation

Abstract

In this paper, we focus on the task of generating reviews of diverse sentiment intensity via fine-grained text sentiment transfer. This task aims to revise an input review to satisfy a given sentiment intensity, while preserving the original semantic content. The challenge is to achieve fine-grained control of the sentiment intensity when generating reviews, different from conventional coarse-grained sentiment transfer task that only reverses the sentiment polarity of text. To tackle this problem, we propose a new text sentiment transfer model which encodes the original views by semantic attention and sentiment attention separately, then incorporates a numeric sentiment intensity value for decoding to finely control the sentiment diversity of the output reviews. Moreover, we formalize the optimization problem as to minimize the negative expected reward as well as the Kullback-Leibler divergence between the distributions of the original and generated reviews. Inspired by the existing Seq2SentiSeq approach, we train the model using a cycle reinforcement learning algorithm but introducing a more precise reward for sentiment transformation to include the sentiment difference of the backward generated reviews from the original ones. Experimental results show that our model outperforms existing methods by a large margin in both automatic evaluation and human evaluation.