Adobe Inc.
CONVERTING TONE OF DIGITAL CONTENT

Abstract:

Techniques are disclosed for generating an output sentence from an input sentence by replacing an input tone of the input sentence with a target tone. For example, an input sentence is parsed to separate semantic meaning of the input sentence from the tone of the input sentence. The input tone is indicative of one or more characteristics of the input sentence, such as politeness, formality, humor, anger, etc. in the input sentence, and thus, a measure of the input tone is a measure of such characteristics of the input sentence. An output sentence is generated based on the semantic meaning of the input sentence and a target tone, such that the output sentence and the input sentence have similar semantic meaning, and the output sentence has the target tone that is different from the input tone of the input sentence. In an example, a neural network for parsing the input sentence and/or generating the output sentence is trained using non-parallel corpora of training data that includes a plurality of input sentences and corresponding plurality of assigned tones.