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

Natural Language Generation (NLG) has the unique ability to create scalable solutions to address the growing problem of hate speech on the internet. Off-the-shelf NLG techniques, which are mostly sequence-to-sequence neural models, are inefficient at defusing angry interactions because they provide safe, repetitive, and conventional replies regardless of the hate speech or irrelevant comments. To successfully increase the diversity and relevance, the article outlines a three-module pipeline strategy. To encourage diversity, the suggested pipeline first uses a generative model to produce a variety of counterspeech options. Next, it uses a BERT model to filter out the ones that are grammatically incorrect, and finally, it uses a retrieval-based approach to choose the most pertinent counterspeech answer. The effectiveness of our method in producing varied and pertinent counterspeech is shown by extensive experiments conducted on three example datasets.

https://arxiv.org/abs/2106.01625