https://phys.org/news/2018-07-offensive-language-social-media-unsupervised.html

An architecture for replacing offensive language

Our method is based on the now popular encoder-decoder neural network architecture, which is the state-of-the-art approach for machine translation. In machine translation, the training of encoder-decoder neural network assumes the existence of a "Rosetta Stone" where the same text is written in both the source and target languages. This paired data enables developers to easily determine whether a system translates correctly and therefore train an encoder-decoder system to do well. Unfortunately, unlike machine translation, as far as we know, there exists no dataset of paired data available for the case of offensive to non-offensive sentences. Moreover, the transferred text must use a vocabulary that is common in a particular application domain. Therefore, unsupervised methods that do not use paired data are needed to perform this task.