ReZG: Retrieval-augmented zero-shot counter narrative generation for hate speech (Neurocomputing)

In order to produce CNs with a high level of specificity for invisible targets, the authors suggest Retrieval-Augmented Zero-shot Generation (ReZG). The proposed multi-dimensional hierarchical retrieval approach extends the retrieval metric from the conventional single dimension to many dimensions appropriate for knowledge that contradicts HS by integrating attitude, semantics, and perplexity. The introduction of an energy-based restricted decoding mechanism allows PLMs to achieve zero-shot CN production by using differentiable knowledge preservation, countering, and fluency constraint functions as control signals for generation rather than in-target CNs. By using the aforementioned methods, ReZG may enhance the specificity of CNs and incorporate external knowledge in a flexible manner. With notable gains of 2.0%+ in relevance and 4.5%+ in countering success rate measures, the results demonstrate that ReZG outperforms strong baselines and possesses higher generalization capabilities.

https://www.sciencedirect.com/science/article/abs/pii/S0925231224019118