Counter Narrative Generation for Fighting Online Hate Speech

Studies on online hate speech have mostly focused on the automated detection of harmful messages. Still, tackling hate speech in the standard way of content moderation may be charged with censorship and overblocking. One alternate strategy, that has received little attention so far, is to directly intervene in the discussion with counter narratives (i.e., textual responses meant to withstand hatred and prevent its spreading). While recent advances in NLP can help in automatizing counter narrative generation by using pre-trained language models, challenges such as lack of high-quality data, generic generation, hallucination and multilinguality must be addressed. This dissertation focuses on automatic and effective hate mitigation targeting Islamophobia by data collection and counter narrative generation. Firstly, we tackle the problem of data scarcity. We present CONAN, the first large-scale and expert-based hate countering dataset for English, Italian and French. Then, we present an author-reviewer approach that can create automatically high-quality data while reducing human effort. Secondly, we develop models to generate counter narratives focusing on informative and multilingual responses. We introduce a knowledge-driven pipeline that can produce suitable and informative English counter narratives while avoiding hallucination phenomena. We address multilinguality by presenting approaches to counter narrative generation for Italian, and characterizing the effect of data size and of data quality on model performance. Thirdly, we present an extensive evaluation of automatic counter narrative generation embedded in a platform that NGO operators can use to monitor social media data and counter hate speech. Results show an increased efficiency and effectiveness of operators' activities. We conclude by discussing our contributions and future research directions on building models for hate countering.