Sentiment Transfer Using BART Based on Aspect-Based Masking

Abstract

This study deals with fine-grained sentiment transfer, which changes the sentiment polarity of a part of review. In the sentiment transfer, the words to be transformed are mainly limited to opinion terms. Therefore, we propose a learning method for the sentiment transfer using special mask tokens for aspect targets and their opinion terms, aiming at fine-grained control of the aspect-based sentiment transfer in text. In this method, text is masked based on triplets composed of aspect terms, opinion terms, and polarity and an encoder-decoder model is trained to fill masked inputs for sentiment transfer. We fine-tune the pre-trained BART for sentiment transfer in two steps: (1) a masked-language modeling task for only an encoder, and (2) a denoising task for both an encoder and a decoder. In experiments, we show that the proposed method is able to achieve fine-grained transfer, which is evaluated using an aspect-based sentiment analyzer.