Balancing the Style-Content Trade-Off in Sentiment Transfer UsingPolarity-Aware Denoising

Anonymous

Abstract: We present a polarity-aware denoising-based sentiment transfer model, which accurately controls the sentiment attributes in generated text, preserving the content to a great extent. Though current models have shown good results, still two major issues exist: (1) target sentences still retain the sentiment of source sentences (2) content preservation in transferred sentences is insufficient. Our proposed polarity-aware enhanced denoising mechanism helps in balancing the style-content trade-off in sentiment-controlled generation. Our proposed method is structured around two key stages in the sentiment transfer process: better representation learning using a shared encoder (pre-trained on general domain) and sentiment-controlled generation using separate decoders. Our extensive experimental results show that our method achieves good results for balancing the sentiment transfer with the content preservation.

Paper Type: long