Outcome-Constrained Large Language Models for Countering Hate Speech
Abstract: Counter speech (CS) that challenges or counteracts harmful or discriminatory messages is an effective way to diminish the influence of hate speech (HS). Automatic CS generation methods have been developed to assist efforts in combating online HS. Existing research focuses on generating CS with linguistic attributes, such as being polite, informative, and intent-driven. However, the real impact of CS in online environments is seldom considered. This study aims to develop methods for generating CS constrained by conversation outcomes and evaluate their effectiveness. We experiment with large language models (LLMs) to incorporate into the text generation process two desired conversation outcomes: low conversation incivility and non-hateful hater reentry. Specifically, we experiment with instruction prompts, LLM finetuning, and LLM reinforcement learning. Evaluation results show that our methods effectively steer the generation of conversational systems towards desired outcomes. Our analyses, however, show that there are differences in the quality and style of the generated CS.
Paper Type: Long
Research Area: Computational Social Science and Cultural Analytics
Research Area Keywords: sociolinguistics, task-oriented, text-to-text generation, applications, hate speech
Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis
Languages Studied: English
Submission Number: 1720
Loading