Generative Sentiment Transfer via Adaptive Masking

Sentiment transfer aims at revising the input text to satisfy a given sentiment polarity while retaining original semantic content. The nucleus of lies in precisely separating information from content information. Existing explicit approaches generally identify and mask tokens simply based on prior linguistic knowledge manually-defined rules, leading low generality undesirable performance. In this paper, we view positions be masked as learnable parameters, further propose novel AM-ST model learn adaptive task-relevant masks attention mechanism. Moreover, sentiment-aware language is proposed fill blanks by incorporating both context capture multi-grained semantics comprehensively. thoroughly evaluated two popular datasets, experimental results demonstrate superiority our proposal.