APPDIA: A Discourse-aware Transformer-based Style Transfer Model for Offensive Social Media Conversations

Proceedings of the 29th International Conference on Computational Linguistics, pages 6063–6074 October 12–17, 2022. 6063 APPDIA: A Discourse-aware Transformer-based Style Transfer Model for Offensive Social Media Conversations Katherine Atwell*, Sabit Hassan*, Malihe Alikhani Computer Science Department, School of Computing and Information University of Pittsburgh, Pittsburgh, PA {kaa139,sah259,malihe}@pitt.edu Abstract Using style-transfer models to reduce offensive- ness of social media comments can help foster a more inclusive environment. However, there are no sizable datasets that contain offensive texts and their inoffensive counterparts, and fine-tuning pretrained models with limited la- beled data can lead to the loss of original mean- ing in the style-transferred text. To address this issue, we provide two major contributions. First, we release the first publicly-available, par- allel corpus of offensive Reddit comments and their style-transferred counterparts annotated by expert sociolinguists. Then, we introduce the first discourse-aware style-transfer mod- els that can effectively reduce offensiveness in Reddit text while preserving the meaning of the original text. These models are the first to examine inferential links between the com- ment and the text it is replying to when trans- ferring the style of offensive Reddit text. We propose two different methods of integrating discourse relations with pretrained transformer models and evaluate them on our dataset of offensive comments from Reddit and their in- offensive counterparts. Improvements over the baseline with respect to both automatic met- rics and human evaluation indicate that our discourse-aware models are better at preserv- ing meaning in style-transferred text when com- pared to the state-of-the-art discourse-agnostic models. 1 Introduction Disclaimer: Due to the nature of this work, figures and examples may contain offensive phrases. The spread of offensive and hateful content on social media can be detrimental to users’ psycho- logical well-being (Waldron, 2012; Gülaçtı, 2010). Anonymity on platforms such as Reddit can fur- ther embolden users to post such hateful content (Ascher, 2019). Further, the sheer volume of con- tent on popular social media platforms can render the human moderation process ineffective (Hassan Figure 1: Example of instances where pretrained language models either fail to remove offensiveness (BART/T5) or drastically alter the intended meaning (DialoGPT) when fine-tuned on our style transfer task et al., 2022) or psychologically damaging for mod- erators (Dosono and Semaan, 2019) and calls for AI systems that can mitigate this problem. AI moderation of social media by simply remov- ing content classified as offensive (Zampieri et al., 2020; Hassan et al., 2021) may reduce diversity in online conversations and deter users from using the platform (Jhaver et al., 2019). Our exploration re- veals that many comments removed by moderators on Reddit contain contributions to the discourse beyond their offensive content. Rather than sim- ply removing these comments from social media platforms, they can be turned into inoffensive state- ments by using alternative words, removing profan- ity, or paraphrasing certain parts, while preserving the overall semantic content. We approach the problem of eliminating offen- siveness from text while preserving original se- mantic content as a supervised style-transfer task, where offensive text is …

Content truncated to 500 words