Let us say, if you want to change the sentiment of sentence x, whose hidden representations are s and c, respectively. [s for style/sentiment, c for content], then you need to find s* among all the opposite sentences which have the closest c with your sentence.
In other words, the best sentence x = argmax[ cos(c, c) ] among x has an opposite sentiment, where c is the content embedding of x.
Then decode from [s; c], where s = style of x.
Let us say, if you want to change the sentiment of sentence x, whose hidden representations are s and c, respectively. [s for style/sentiment, c for content], then you need to find s* among all the opposite sentences which have the closest c with your sentence.
In other words, the best sentence x = argmax[ cos(c, c) ] among x has an opposite sentiment, where c is the content embedding of x. Then decode from [s; c], where s = style of x.