This study presents a unique paradigm based on contrastive optimum transport that successfully tackles the problems of encouraging diversification and preserving target engagement when creating counter-narratives. First, the comparison pairs between the transported and original features are retrieved using an Optimal Transport Kernel (OTK) module to add hatred target information to the token representations. Second, model degradation is tackled through the use of a self-contrastive learning module. Ultimately, an enhanced decoding strategy that incorporates a target-oriented search mechanism is incorporated to specifically encourage domain relevance and diversity during the inference phase. This tactic adjusts the model’s confidence score by taking target relevance and token similarity into account. The suggested model performs noticeably better than the state-of-the-art approaches as measured by several different measures.

By author

Leave a Reply

Your email address will not be published. Required fields are marked *