Generate, Prune, Select: A Pipeline for Counterspeech Generation against Online Hate Speech

Abstract

Warning: this paper contains content that may be offensive or upsetting.Countermeasures to effectively fight the ever increasing hate speech online without blocking freedom of speech is of great social interest. Natural Language Generation (NLG), is uniquely capable of developing scalable solutions. However, off-the-shelf NLG methods are primarily sequence-to-sequence neural models and they are limited in that they generate commonplace, repetitive and safe responses regardless of the hate speech (e.g., "Please refrain from using such language.") or irrelevant responses, making them ineffective for de-escalating hateful conversations. In this paper, we design a three-module pipeline approach to effectively improve the diversity and relevance. Our proposed pipeline first generates various counterspeech candidates by a generative model to promote diversity, then filters the ungrammatical ones using a BERT model, and finally selects the most relevant counterspeech response using a novel retrievalbased method. Extensive Experiments on three representative datasets demonstrate the efficacy of our approach in generating diverse and relevant counterspeech.

Counterspeakers’ Perspectives: Unveiling Barriers and AI Needs in the Fight against Online Hate

“…Automatic detection [147] and computational analysis of large scale counterspeech [43] have been used to understand its characteristics and to inform effective content. Prior works on automatic generation of counterspeech [91,149] relied on curated or scraped datasets [25,63] and evaluation metrics based on correct countering claims [55] or emotion and politeness [108]. Some methods used limited response intent such as question, denouncing, and humor in dataset [25] and as part of the generation method [53].…”

“…AI has recently emerged as a potential tool or solution [2,34,149] to assist with this increased demand for counterspeech. However, designing an AI system for counterspeech is a unique challenge that requires an understanding of larger context of its impact on those who do it [8].…”

“…On one hand, it is important to reduce the burden on counterspeakers [67], who are sometimes victims of hate speech themselves [20]. On the other hand, naive AI solutions that simply generate counterspeech without human oversight could significantly detract from the authentic, empowering, and emotionally connecting experience that counterspeech provides to users [12,42], despite being a growing area of research in AI [24,108,149].…”

Reinforcement Learning-based Counter-Misinformation Response Generation: A Case Study of COVID-19 Vaccine Misinformation

“…Counter-hate [16,36,92,121] and counter-argument [5,37,40] text generation tasks are also related to our problem setting, where the generated text is aimed to refute the original post spreading hate and any generic argument, respectively. Some proposed models finetune large scale unsupervised language models on the hate-speech or argument text for text generation [75,92].…”

“…Some proposed models finetune large scale unsupervised language models on the hate-speech or argument text for text generation [75,92]. Other models first generate a set of candidate counter-hate/counter-argument replies, and then select one based on the relevance to the original post in a generate-then-retrieve or identify-substitute manner [37,40,121] . Meanwhile, some related counter-hate/counter-argument datasets have also been released [40,71,74].…”