Towards A Friendly Online Community: An Unsupervised Style Transfer Framework for Profanity Redaction

Proceedings of the 28th International Conference on Computational Linguistics, pages 2107–2114 Barcelona, Spain (Online), December 8-13, 2020 2107 Towards A Friendly Online Community: An Unsupervised Style Transfer Framework for Profanity Redaction Minh Tran†, Yipeng Zhang§, Mohammad Soleymani† †Institute for Creative Technologies, University of Southern California, Los Angeles, CA, USA §University of Rochester, Rochester, NY, USA †{mtran,soleymani}@ict.usc.edu §yzh232@u.rochester.edu Abstract Offensive and abusive language is a pressing problem on social media platforms. In this work, we propose a method for transforming offensive comments, statements containing profanity or offensive language, into non-offensive ones. We design a RETRIEVE, GENERATE and EDIT un- supervised style transfer pipeline to redact the offensive comments in a word-restricted manner while maintaining a high level of fluency and preserving the content of the original text. We extensively evaluate our method’s performance and compare it to previous style transfer models using both automatic metrics and human evaluations. Experimental results show that our method outperforms other models on human evaluations and is the only approach that consistently per- forms well on all automatic evaluation metrics. 1 Introduction Despite the undeniably positive impact social media has on facilitating communication, it is also a medium that can be used for abusive behavior. Many social media platforms do not restrict the lan- guage users use, leading to an overflow of strong language that might not be appropriate for children (Duggan, 2014; Rieder, 2010). Verbal abuse and cyber-bullying is also a common problem on social media. Such phenomena are harmful to the victims, the online community, and in particular adolescents who are more susceptible and vulnerable in such situations (Patchin and Hinduja, 2010; Pieschl et al., 2015). To mitigate such problems, recent studies have focused on developing machine learning models for detecting hate speech (Davidson et al., 2017; Xiang et al., 2012; Djuric et al., 2015; Waseem and Hovy, 2016; Chen et al., 2012; Xiang et al., 2012; Founta et al., 2019). However, little progress has been made regarding the task of transforming hateful sentences into non-hateful ones, a potential next-step after detecting the hateful content. dos Santos et al. (2018) propose an extension of a basic encoder- decoder architecture by including a collaborative classifier. To the best of our knowledge, this is the only approach dealing with abusive language redaction. Unsupervised text style transfer is an important area in text generation that has recently received a lot of attention. Generally speaking, text style transfer is the task of rewriting sentences in a source style to a target style while preserving the original sentences as much as possible. In the context of the paper, we define a corpus to be stylistic if every sample in the corpus shares a common style. Most style transfer approaches are developed and validated on bi-stylistic datasets (Shen et al., 2017; Hu et al., 2017; Li et al., 2018; Prabhumoye et al., 2018; Tian et al., 2018; He et al., 2019; Wu et al., 2019), which require stylistic features on both source and target samples. Some common bi-stylistic datasets for …

Content truncated to 500 words