Affective Natural Language Generation by Phrase Insertion

We propose a highly precise, production-ready neural model for affective natural language generation. It is designed to add predefined sentiment to neutral utterances without changing the meaning significantly. It works by inferring phrases and their insertion points. In our work we also propose strict correctness criteria and apply them to our inference results achieving human-level precision. The model is not specific to any particular domain like IoT or restaurants review. We use six selected emotion categories, but we also speculate that the model could be applied to other affective categories, like informal style or politeness, without a design change.