Reproduce “Counterspeech the ultimate shield! Multi-Conditioned Counterspeech Generation through Attributed Prefix Learning” on OpenResearch

About this paper

Aswini Kumar, Anil Bandhakavi, Tanmoy Chakraborty

Researchers from IIT Delhi introduced HiPPrO, a two-stage framework for generating counterspeech conditioned on both strategic intent and emotional tone. The framework, utilizing hierarchical prefix learning and preference optimization, outperformed state-of-the-art baselines and commercial LLMs like GPT-3.5 and GPT-4 in producing multi-attribute aligned, effective, and non-toxic responses, achieving an average strategy conformity of 0.929 and emotion conformity of 0.706.

Key insights

• Hierarchical prefix learning effectively integrates and controls multiple conditioning attributes (strategy and emotion) within a single generation process, leading to more nuanced outputs.
• Odds Ratio Preference Optimization (ORPO) can align model outputs with human preferences for constructive and non-toxic language without requiring explicit reward models or reference models.
• Parameter-efficient fine-tuning techniques are effective for adapting large language models to complex, specialized tasks like multi-attribute counterspeech generation with reduced computational overhead.

Summary by alphaXiv.