Abstract
We introduce a new approach to tackle the problem of offensive language inonline social media. Our approach uses unsupervised text style transfer totranslate offensive sentences into non-offensive ones. We propose a new methodfor training encoder-decoders using non-parallel data that combines acollaborative classifier, attention and the cycle consistency loss.Experimental results on data from Twitter and Reddit show that our methodoutperforms a state-of-the-art text style transfer system in two out of threequantitative metrics and produces reliable non-offensive transferred sentences.
Quick Read (beta)
loading the full paper ...