Using Pre-Trained Language Models for Producing Counter Narratives Against Hate Speech: a Comparative Study

Findings of the Association for Computational Linguistics: ACL 2022, pages 3099 - 3114 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Using Pre-Trained Language Models for Producing Counter Narratives Against Hate Speech: a Comparative Study Serra Sinem Tekiro˘glu2, Helena Bonaldi1,2, Margherita Fanton1,2∗, Marco Guerini2 1University of Trento, Italy 2Fondazione Bruno Kessler, Via Sommarive 18, Povo, Trento, Italy tekiroglu@fbk.eu, hbonaldi@fbk.eu, margherita.fanton@ims.uni-stuttgart.de, guerini@fbk.eu Abstract In this work, we present an extensive study on the use of pre-trained language models for the task of automatic Counter Narrative (CN) gen- eration to fight online hate speech in English. We first present a comparative study to deter- mine whether there is a particular Language Model (or class of LMs) and a particular de- coding mechanism that are the most appropri- ate to generate CNs. Findings show that au- toregressive models combined with stochastic decodings are the most promising. We then in- vestigate how an LM performs in generating a CN with regard to an unseen target of hate. We find out that a key element for successful ‘out of target’ experiments is not an overall similar- ity with the training data but the presence of a specific subset of training data, i. e. a target that shares some commonalities with the test target that can be defined a-priori. We finally introduce the idea of a pipeline based on the addition of an automatic post-editing step to refine generated CNs. 1 Introduction Hate Speech (HS) has found fertile ground in Social Media Platforms. Actions undertaken by such platforms to tackle online hatred consist in identifying possible sources of hate and removing them by means of content deletion, account suspension or shadow-banning. However, these actions are often interpreted and denounced as censorship by the affected users and political groups (Myers West, 2018). For this reason, such restrictions can have the opposite effect of exacer- bating the hostility of the haters (Munger, 2017). An alternative strategy, that is looming on the horizon, is based on the use of Counter Narratives. CNs are “all communicative actions aimed at refuting hate speech through thoughtful and cogent reasons, and true and fact-bound arguments" (Schieb and Preuss, 2016). As a de-escalating ∗ Now at the University of Stuttgart, Germany. measure, CNs have been proven to be successful in diminishing hate, while preserving the freedom of speech (Benesch, 2014; Gagliardone et al., 2015). An example of <HS, CN> pair is shown below: HS: Women are basically childlike, they remain this way most of their lives. Soft and emotional. It has devastated our once great patriarchal civilizations. CN: Without softness and emotions there would be just brutality and cruelty. Not all women are soft and emotional and many men have these characteristics. To perpetuate these socially constructed gender profiles maintains norms which oppress anybody. Based on their effectiveness, CNs have started be- ing employed by NGOs to counter online hate. Since for NGO operators it is impossible to man- ually write responses to all instances of hate, a line of NLP research has recently …

Content truncated to 500 words